Artificial intelligence (AI) has helped uncover new weaknesses in two cryptographic systems, including a simplified version of the Advanced Encryption Standard (AES) that underpins much of today's internet. While these headline-grabbing findings don't put anyone's passwords or bank accounts at immediate risk, experts say the research could mark the beginning of a new era in which AI becomes a powerful assistant for discovering flaws in the mathematical foundations of digital security.

In a blog post published July 28, representatives from Anthropic's Frontier Red Team said Claude Mythos Preview independently developed new cryptanalytic techniques against two different targets: a version of AES-128, one of the world's most widely used encryption algorithms, and HAWK, an experimental post-quantum digital signature scheme currently being evaluated as part of the U.S. National Institute of Standards and Technology's (NIST) effort to standardize cryptography for the quantum computing age.

It's tempting to interpret this news as AI cracking one of the internet's most important encryption algorithms, but the reality is less dramatic. Nonetheless, experts say the achievement could mark the start of a new age of digital security.

The power of encryption

AES protects huge amounts of everyday digital life, from encrypted websites and messaging apps to Wi-Fi networks and financial transactions. But the version Anthropic attacked wasn't the full AES-128 algorithm used in those systems. Instead, the researchers studied a seven-round version of AES, a deliberately weakened variant long used by cryptographers to test new attack techniques.

The full AES-128 algorithm uses 10 rounds of encryption, while Anthropic's research focused on a seven-round version. In a self-published study that has not been peer-reviewed, scientists Milad Nasr and Nicholas Carlini said Claude found a faster way to recover the encryption key from that simplified version, making the best-known attack between 200 and 800 times faster. However, the study stressed that the technique does not work against the full version of AES used to protect real-world systems.

The more significant result may instead involve HAWK. Unlike AES, which has protected data for more than two decades, HAWK is a relatively new digital signature scheme designed to resist attacks from future quantum computers. It's one of the remaining candidates in NIST's additional post-quantum signature standardization process, meaning it is still being scrutinized by researchers before its widespread deployment.

In a separate study, Anthropic scientists Zygimantas Straznickas and Stephen A. Weis found that Claude spotted a mathematical property that researchers hadn't previously exploited. Combined with existing attack techniques, this property made it much easier to recover HAWK's secret key.

Thomas Espitau ‪—‬ head of research at PQShield, a cybersecurity company that specializes in post-quantum cryptography, and a cryptographer who has published research on HAWK ‪—‬ described the work as "one of the most, if not the most significant, cryptanalytic result of the year."

"To say it plainly, this is great work, and it is exactly what the NIST process is designed to produce," Espitau told Live Science. "Candidate schemes exist to be attacked before they are deployed, not after."

Espitau said the attack combined previously known techniques with one missing mathematical insight that Claude identified, substantially reducing HAWK's estimated security margin. He believes the work demonstrates how AI could increasingly help researchers evaluate the strength of cryptographic systems before they are adopted.

"This development illustrates how AI can act as a powerful accelerator in cryptanalysis," he said. "Claude Mythos Preview identified the exploitation of the sign-flip symmetry, providing the final piece of a puzzle that the research community had been assembling."

Building defensive measures

However, not everyone thinks the announcement means AI suddenly outsmarted human cryptographers.

"There is nothing you need to change," Roberta Faux, head of cryptography at cybersecurity and quantum encryption company Arqit, told Live Science. "Both Anthropic and the outside cryptographers agree that you don't need to change your key sizes or switch your cryptography."

Instead, Faux said the bigger story is AI's ability to apply expert-level cryptanalysis across thousands of algorithms that relatively few researchers have had time to examine.

"The interesting prospect is not a model outdueling the world's best lattice theorist on one problem but a model applying solid, roughly expert-level analysis at scale to the several thousand ciphers nobody ever had the human-hours to examine," she said.

Faux also cautioned against attributing the HAWK breakthrough solely to AI.

"The HAWK attack used no exotic ingredients; it simply competently assembled tools that were already lying around," she said. "A post-quantum candidate is seriously scrutinized by maybe a few dozen people on the planet, so beating two years of review mostly reveals how thin that layer of review is."

For consumers, the immediate implications are reassuring. The encryption protecting online banking, shopping, messaging and cloud storage has not suddenly become obsolete. But for the researchers designing the next generation of cryptography, Anthropic's work hints that AI could soon help cryptographers test tomorrow's encryption schemes more quickly and more thoroughly than has previously been possible.

Each year, tons of soil, silt and rocks flow into reservoirs in the U.S. and get stuck there, creating a "ticking time bomb" that threatens lives and livelihoods. According to a new study, around a fifth of the world's reservoirs are currently at high risk of filling with sediment, and by 2060, half could be inoperable as a result.

The sediment takes up space meant for water, putting additional pressure on structures that were not built to contain large quantities of sand and sludge. As the reservoir fills with the material instead of water, sediment-filled reservoirs pose a serious threat to water security and the safety of people living near or downstream of them — and as the climate changes, scientists expect the problem to worsen.

"We are losing storage space in our reservoirs faster from sediment accumulation than we gain space from new dam construction," Amy East, a geologist at the U.S. Geological Survey, told Live Science.

What is dam sedimentation?

Dam sedimentation is not a uniquely U.S. problem. Rain pulls sediment down slopes into waterways. Normally, this sediment would travel through the river system, shaping the water flow, creating habitats, and fortifying coastal wetlands and estuaries against erosion. In many parts of the world, this sediment is also a vital natural fertilizer for crops. But after a river is dammed, this sediment can collect in the reservoir.

A 2023 study published in the journal Science found that sediment buildup was the greatest threat to global water storage in existing reservoirs — even more than drought.

New research, which includes smaller reservoirs, has offered a more nuanced and concerning picture. In the study, published June 5 in the journal Nature Sustainability, researchers assessed the state of reservoir sedimentation globally. They looked at more than 550,000 reservoirs, most of which spanned less than 0.4 square miles (1 square kilometre). They found that almost 1 in 5 of reservoirs globally are at high risk of filling up with sediment.

The researchers found that sedimentation is eating into global reservoir storage capacity at about 7.3% every decade. There are 16 global hotspots that are at high risk of reservoir sedimentation, threatening water supplies to more than 2 billion people and impacting more than a quarter of global irrigated land. One of these hotspots includes the western U.S.

The authors found that sedimentation risk was especially pronounced in regions with dense small-reservoir networks supporting key crop systems.

"Without intervention, we project that, by 2060, more than half of all reservoirs could become functionally inoperable," the authors wrote. This includes more than half of all small reservoirs and more than a third of large reservoirs globally.

Where is it a problem in the U.S.?

North America has the second largest number of reservoirs, after Asia, according to the global survey. And at 7.8%, its sedimentation rate is higher than the global average, ranking it third after Oceania (9.9%) and South America (8.1%),

The U.S. is home to many small dams, said Matthias Kondolf, a geomorphologist at the University of California, Berkeley who wasn't involved in the study. "All the focus tends to be on the big dams, and these many, many small ones have been neglected," he told Live Science.

From the 1940s to 1970s, there was a big push by the U.S. government to build "farm ponds," he said. In western Oklahoma, for example, "all these dams are now filling up with sediment, and there's nothing in place to assess the safety of these features, which have now become threats."

Some of these dams have already failed, following large storms as the ageing dams were not able to cope with the volume of water. A 2018 study found that since 1980, there have been on average 24 dam failures a year in the U.S.

Reservoirs in some regions are more at risk than others. The western U.S. is a natural hotspot because it has a steeper topography, meaning more sediment naturally ends up in watersheds, East, who wasn't involved in the new research, said.

And in areas where human activities like farming or logging disrupt the landscape, rainfall is likely to drag even more sediment into reservoirs, East said. "Those types of land use can create sediment-producing hotspots almost anywhere."

In California, for example, more than half of small reservoirs in maize-growing areas are filling up with sediment at a rate of more than 10% a year, which threatens local maize production.

The threat to water security will be particularly acute in places struggling with wildfires, such as the western U.S. After fires, rain washes the debris and sediment into watersheds. A 2024 study in California found that most of this post-fire sediment erosion occurred upstream of reservoirs, rather than in places where sediment would move downstream directly to the ocean.

The dam-sedimentation situation is likely to get worse as the climate changes, researchers said. As the climate changes, there will be more extreme weather events, such as droughts and wildfires, adding more sediment to watersheds. Dams are designed to cope with the pressure of water, not sediment which exerts more force on dam walls. Moreover, in seismically active areas, the sediment vibrates at a different frequency than the concrete of the dam, which creates further impact against the structure. Additionally, the dams are just getting older.

"What do we do 50 or 100 years from now when we have all these reservoirs across the landscape filled with sediment, sitting there as ticking time bombs?" Kondolf said.

What can be done about it?

For a reservoir to be sustainable long-term, the amount of sediment entering it and leaving it should be the same.

"A big part of preventing reservoir sedimentation problems is upstream soil conservation in the watershed," East said. Interventions include ensuring that land use, such as agriculture or timber harvesting, follow best practice to reduce the amount of sediment washing off the landscapes.

If sediment is already in the dam, it can be difficult to manage and expensive to remediate, Kondolf said. Some dams, such as Fall Creek Dam in Oregon, have mechanisms that enable engineers to release sediment. But in the past, reservoirs were not designed to manage sediment and little thought was given to how to decommission them.

"This is one of those issues that falls in the cracks between different agencies and responsibilities," Kondolf said. Some larger facilities fall under the remit of the federal government. Other dams fall under state governments. California, for example, recently launched a state-sponsored program to future-proof the state's dams.

But smaller dams often get overlooked, Kondolf said.

One option is to remove the dams entirely. For example, sediment-related concerns contributed to the decision to remove the Elwha and Glines Canyon dams in Washington. The Elwha River later carried most of the sediment to its river mouth.

For smaller dams with a relatively small amount of sediment, "you just blow up the dam and let the sediment work its way downstream", Kondolf said.

There are other technological solutions, but they are expensive, such as retrofitting dams to enable them to pass sediment; creating a bypass to route sediment-heavy water around the dam; and mechanical dredging.

Another option is to empty the reservoir of water and let the river cut a channel through the sediment, Kondolf said.

However, released dam sediment brings its own risks. If not managed carefully, sediment can flood the river system and disrupt downstream ecosystems.

"The economics of managing reservoir sediment long-term were not generally incorporated into dam design and planning in the last century when many western-U.S. water projects began," East said. This has "pushed the cost burden on future generations."

But the reality is that reservoirs around the country are filling with sediment. It is reducing their ability to store water and potentially endangering the communities living downstream of them — and unless addressed, the problem will only get worse.

When the famous Chicxulub asteroid struck what is now Mexico's Yucatán Peninsula 66 million years ago, it ended the reign of the nonavian dinosaurs. As ecosystems recovered, small, adaptable mammals rapidly spread into newly vacant niches, evolving new diets, larger bodies and more diverse ways of life until they dominated the land within a few million years.

Now, new evidence suggests that even before the asteroid hit, Earth was already beginning to rot.

A study published May 12 in the journal PNAS found evidence of environmental fungal blooms before and after the Cretaceous-Paleogene (K-Pg) mass extinction event. Looking at microscopic fossils from rocks dating to the K-Pg event in North America, a team of researchers led by Dr. Arturo Casadevall, chair of the Molecular Microbiology and Immunology Department at Johns Hopkins University, identified an earlier fungal bloom dating to roughly 30,000 to 100,000 years before the asteroid hit, during a period of cooling linked to intense volcanic activity in the Deccan Traps in what is now India.

The researchers also found a fungal spike immediately after the Chicxulub impact exactly when dead plants and animals would have littered devastated ecosystems. The finding adds to earlier evidence from New Zealand that suggests the fungal bloom wasn't just a local event.

This idea feeds into a controversial idea called the fungal infection-mammalian selection (FIMS) hypothesis. First proposed two decades ago by Casadevall and subsequently updated in 2018 and 2026 papers, the FIMS hypothesis suggests that the abundance of fungus before the asteroid impact may have helped mammals gain an evolutionary advantage over the dinosaurs that remained after that cataclysm.

The hypothesis doesn't argue that fungus wiped out the dinosaurs. Instead, it proposes that mammals' live births and immune systems may have helped them withstand the rise of fungus in ways egg-laying reptiles and nonavian dinosaurs could not, thereby giving them the upper hand in disrupted ecosystems full of rot, cold, starvation and polluted air.

"The fungal infection mammalian selection hypothesis is focused on how the mammals became the dominant land animals," Casadevall told Live Science via email, "not dinosaur extinction."

But not all experts are convinced. Mary O'Connell, the Chair of the Zoology department at the University of Manchester told Live Science via email that the hypothesis doesn't fit all the evidence and that "it emerges from a chain of plausibility arguments rather than direct evidence, and the links in the chain are weaker than the framing may suggest."

A fungal world before the asteroid

To find this pre-asteroid fungal bloom in their new study, Casadevall and Rosanna Baker, who researches fungus at Johns Hopkins, focused on palynomorphs — tiny fossilized organic remains, including fungal spores, fungal hyphae and plant pollen — preserved in ancient sediment layers from the Denver Basin in Colorado dating back to the K-Pg extinction.

Baker counted between 100 and 500 microfossils per sample across sediment layers spanning about 60,000 years before the K-Pg mass extinction event to 30,000 years after it. Most samples were dominated by plant material, but some layers contained 50% or more fungal spores. The researchers interpreted this high proportion of fungal spores as a sign of ecological disturbance ‪—‬ more dead material and more fungi feeding on it, especially after the dinosaur-killing asteroid.

Baker and Casadevall also found an earlier fungal bloom dating to roughly 10,000 to 30,000 years before the asteroid impact.

"What surprised us was the proliferation in association with Deccan volcanism, which implies ecological disruption prior to [the] meteor impact," Casadevall told Live Science via email.

In other words, ecosystems may have already been under stress — and faced with a fungal overgrowth — before the asteroid delivered the final blow.

The case for FIMS

When Casadevall first proposed the FIMS hypothesis, it rested on a simple observation: Mammals are unusually resistant to invasive fungal diseases.

Fungi are major pathogens of plants, insects, amphibians and many reptiles, but life-threatening systemic fungal infections are relatively rare in mammals with healthy immune systems. Casadevall argued that mammals' relative resistance to fungi came from their sophisticated immune systems. In mammals, first-line, relatively nonspecific defenses such as neutrophils recognize fungal cell walls, engulf spores and attack invasive filaments, while the adaptive arm of the immune system excels at building customized antibodies that selectively target specific fungi, and then creating immune memory so the invaders are recognized the next time they're encountered.

Casadevall also initially believed that mammals' high, stable body temperatures — generally around 97 to 104 degrees Fahrenheit (36 to 40 degrees Celsius) — could prevent many environmental fungi from growing well enough to establish an infection.

"The success of mammals made no sense given their high energy needs and this idea provided a plausible mechanism for their success," Casadevall said.

After the asteroid hit, forests burned, plants died, darkness fell and dead matter piled up, providing a feast for fungi. If fungal spores filled the air, surviving animals may have inhaled enormous, possibly dangerous doses.

However, the FIMS hypothesis proposes that mammals had several possible adaptations against these fungal invaders that other animals did not. In addition to having warm body temperatures and complex immune systems, mammals who could maintain stable body temperatures could forage during the cold "impact winter," ‪when ash and dust blocked out the light of the sun. Meanwhile the dinosaurs, who were originally thought to be the "cold-blooded," would have struggled to stay active and find food, leading them to become weaker and more susceptible to disease.

Differences in gestation may have mattered, too. Many mammals protect developing embryos inside the mother's body, where warmth and immunity could shield them from fungi. In contrast, all dinosaurs laid eggs. Those eggs developed in nests, which were often in contact with soil and decaying plant matter ‪—‬ the exact environments where fungi thrive. Modern fungi can invade reptile eggs, and fossilized hyphae, a type of fungus, was described in dinosaur eggshells in a 2008 study.

That does not prove fungi wiped out dinosaur nests. Fossilized hyphae could suggest either that fungi invaded a living egg, or that fungi colonized the egg after death. Still, eggshells provide one clue that fungi were a problem for the survival of dinosaur young..

A sick sauropod

Another potential line of evidence to buttress the FIMS hypothesis came from dinosaur bones that showed evidence of fungal infection.

In 2022, a team led by Cary Woodruff, curator of vertebrate paleontology at the Phillip and Patricia Frost Museum of Science in Miami, described a Late Jurassic sauropod fossil from Montana with unusual lesions in its neck vertebrae. The dinosaur, an immature diplodocine known as MOR 7029, lived roughly 150 million years ago, long before the asteroid impact.

Sauropods had long, air-filled neck bones connected to their respiratory system. In fossils of healthy sauropods, the places where air-sac tissue met bone are smooth, sometimes glass-like, Woodruff told Live Science.

But in this animal, the normally smooth sockets were rough and irregular. "Coming out of the socket — like imagine coming out towards you —it looked like fossilized heads of broccoli florets," Woodruff said. "And I had never seen that."

Unsure what he was looking at, Woodruff posted a photo online and asked whether anyone had seen anything similar. Other anatomists and paleontologists quickly weighed in. Their response, he recalled, was essentially, "That's exactly what we would have predicted a respiratory infection to look like," he said.

By comparing the fossil lesions with respiratory diseases in modern birds, which are living dinosaurs, the team tentatively flagged the condition as airsacculitis with associated osteomyelitis ‪—‬ air-sac inflammation that had spread to nearby bone.

However, the researchers could not identify the exact disease that caused the inflammation. In modern birds, airsacculitis can be caused by fungal infections such as aspergillosis, but also by bacterial infections. Without preserved soft tissue or fungal structures, they stopped short of confirming a fungal infection diagnosis.

Either way, Woodruff thinks the sauropod was probably sick enough that the infection contributed to its death. It may have become too weak to keep up, too ill to eat or drink properly, or too vulnerable to escape predators.

The case against FIMS

Even so, many experts remain skeptical of Casadevall's hypothesis. For one thing, the asteroid and its aftermath already explain a tremendous amount. The impact triggered tsunamis, earthquakes and a global food-web collapse. Those disasters alone would have killed off large dinosaurs at the top of disrupted food chains.

Another problem revolves around body temperature. In 2005, when Casadevall first proposed his hypothesis, dinosaurs were often thought of as "cold-blooded," while mammals were "warm-blooded." But over the years, research has uncovered that many nonavian dinosaurs, such as Velociraptor, T. rex and Brachiosaurus, were likely warm-blooded, too ‪—‬ so body temperature wouldn't explain why fungus would have been a big problem for many dinosaurs.

Complicating the picture, birds — the only surviving dinosaurs — often have higher body temperatures than mammals do. So the FIMS theory rests on an assumption that living examples of dinosaurs don't support, Flynn said.

Around the time of the asteroid impact, cold-blooded creatures like turtles and crocodilians that lived in the water "don't really suffer as severely as other groups, even including mammals," Andrew Flynn, an assistant professor of paleobotany at New Mexico State University, told Live Science. "Clearly this fungal thing is not probably what's causing mammals to rise up and go on to dominate the ecosystem."

The FIMS hypothesis also has a contradiction at its heart, said Jingmai O'Connor, a paleontologist at the Field Museum in Chicago.

"It's just a hypothesis I had vaguely heard about but was not popular, not often mentioned, and I guess I would say 'out of favor,'" O'Connor told Live Science via email. "If everything we know that suggests dinosaurs are more susceptible to fungal infections is based on living dinosaurs (birds) and they survived, doesn't that rather undermine this hypothesis?"

Dr. Isabel Jimenez, an assistant professor at Johns Hopkins University and a proponent of FIMS, agrees that this discrepancy must be explained.

"For the FIMS hypothesis to still hold water, we have to consider alternative mechanisms by which dinosaurs could have been more susceptible to fungal diseases than mammals," Jimenez told Live Science in an email.

Modern-day birds are avian theropods — the surviving branch of the dinosaur family tree — and likely faced the same fungal pressures as their extinct relatives. More recently, in 2026, Jimenez and Casadevall have suggested a tweak to their theory: that some of the small-bodied, beaked, flying or semiflying theropods may have been able to pull through, despite some baseline higher susceptibility to fungal disease. That's because they were small, had feathers to better insulate their bodies, and maintained generalist diets that enabled them to use more food sources than the dinosaur lineages that died out.

Working with Jimenez, Casadevall outlined an updated version of the FIMS hypothesis in a paper published earlier this year. In it, they propose that dinosaurs' respiratory systems may have made them more at risk than their mammalian counterparts.

Whereas mammals breathe by expanding and contracting flexible lungs filled with blood-rich alveoli, modern birds — and probably many dinosaurs — have rigid lungs ventilated by a network of thin-walled air sacs. The air sacs act as bellows, maintaining one-way airflow through the lungs, but they perform little gas exchange and have relatively poor blood supply. This system could have also carried inhaled fungal spores deep into the respiratory tract, where they may have settled in air sacs that are harder for circulating immune cells to reach and clear.

While birds have a similar sophisticated immune system as mammals — including macrophages and antibodies — their air sacs extend throughout the body and sometimes into the bones, making fungal infections easier to spread beyond just the respiratory system.

This possible weakness is what Jimenez and Casadevall suggest is what made air-sac-bearing dinosaurs vulnerable to fungal infections, similar to modern birds.

After the asteroid strike, there would have been soot and sulfur polluting the air and dinosaurs would have been nutritionally stressed. "So secondary fungal infections are certainly possible," Jimenez said, referring to infections that take root after the body's defenses are weakened.

But skeptics argue that even with these updates in the FIMS hypothesis, more evidence is needed to show it's not just possible but likely.

The idea that their air-sac respiratory systems made them more vulnerable to fungus is "a more testable claim," but its likely impact on dinosaurs would also be more limited, because of competing environmental factors, O'Connell said.

And because only a handful of fungal species actually cause disease in animals, to show that fungi worsened dinosaur deaths, researchers would need to find evidence of elevated fungal disease in dinosaur fossils from right after the asteroid struck, O'Connell said.

"We would also need to be able to identify elevation of pathogenic fungal spores in the sedimentary record at that boundary, rather than just more spores," she said.

Overall, "the fungal infection-mammalian selection hypothesis feels like a conjecture in search of any supporting evidence," paleontologist and science writer Riley Black, told Live Science in an email.

How much do you know about the king of the dinosaurs? Test your knowledge with our T. rex quiz!

LONDON — In the packed halls of the recent Alzheimer's Association International Conference in mid-July, over 10,000 experts in brain research discussed strategies to beat dementia. But in one room, attendees gathered to watch a video from a nonscientist named Isabel Beltre, whose life was transformed by a recent trial.

Beltre was one of 1,200 participants ages 60 to 77 who took part in the Latin American Initiative for Lifestyle Intervention to Prevent Cognitive Decline (LatAm-FINGERS), the first trial to test a culturally adapted lifestyle intervention to prevent cognitive decline in Latin America. Rather than attempting to treat dementia after it has developed, the trial tackled risk factors that raise the likelihood of the condition, like smoking and a sedentary lifestyle.

In her native Spanish, Beltre thanked both God and the project for its impact on her life. There was a celebratory air in the conference room as the trial's results were released. The lifestyle interventions, which included exercise programs and brain training, had statistically improved participants' thinking skills, compared with a control group who received only bare-bones support, like general health advice.

Trials like LatAm-FINGERS — including a related U.S.-based trial called the POINTER study — ultimately aim to cut rates of dementia around the world. The LatAm trial is the most successful of these trials to date, raising hopes that this approach could help to prevent dementia en masse.

Whether that promise pans out remains to be seen, though, experts told Live Science.

The lifestyle factors that increase dementia risk

When it comes to dementia prevention, few guidelines are as influential as those written by the Lancet Commission, an expert-led review of the available evidence on dementia. Three editions of this review have provided comprehensive overviews of risk factors; the most recent, published in 2024, identified 14 factors, including smoking, infrequent social contact, and a lack of physical activity.

The commission estimated that, together, the identified risk factors contribute to 45% of dementia cases at a population level.

Long before the 2024 report, Dr. Miia Kivipelto, a professor of clinical geriatrics at the Karolinska Institute in Sweden, wanted to tackle these risk factors at the individual and population level in her work as a medical doctor and epidemiologist.

"I felt that I wanted to move on to interventions to really show that if you do something, you can change the risk," Kivipelto told Live Science.

The many contributing factors that lead to cognitive decline, and ultimately dementia, couldn't be tackled in isolation, Kivipelto realized. She designed a trial in her native Finland, called the Finnish Geriatric Intervention Study to Prevent Cognitive Impairment and Disability (FINGER) study, that would aim to tackle many risks simultaneously.

The two-year project recruited 1,260 participants ages 60 to 77 and divided them into two groups. The first group received general health advice, such as tips on healthy eating and exercise, at the start of the study and then again after six and 12 months.

The other group additionally received a multipart intervention consisting of meetings with nutritionists to discuss personalized meal plans, brain training, exercise programs, social activities, and metabolic and vascular tests. These programs lasted throughout the study period, with the brain training alone including 144 sessions over 12 months of the study.

At the end, the two groups' cognitive skills, such as thinking speed and memory, were tested; the trial did not track dementia rates, only overall cognition. In both groups, these skills improved by the end of the study. But the headline finding was that in the group given the comprehensive intervention, overall brain function was judged to be 25% stronger than in the control group.

What did that difference mean in concrete terms? It turns out, the improvement was relatively small.

It was calculated in terms of a statistical measure called a z-score, which represented the extent of the improvement in brain-function test scores in each group. In the end, the difference in z-scores between the two groups was just 0.04 points, which a 2021 review defined as "small."

However, Kivipelto maintained that the improvement is significant. "It's as big as anti-amyloid treatments," she said.

Several drugs approved by the U.S Food and Drug Administration for the treatment of Alzheimer's disease target amyloid, a protein that forms unruly tangles in the brain. These drugs have been reported to marginally slow the disease's progression, but statistically, their overall effect is small.

Targeting those most at risk of dementia

Despite criticism of the trial's moderate effects, Kivipelto's FINGER intervention has since been repeated all over the world, including by the U.S.-based POINTER study, and, most recently, LatAm-FINGERS.

Dr. Gill Livingston, a psychiatrist at University College London who led the 2024 Lancet Commission report, told Live Science that the data from these trials shows that dementia interventions need to be tailored to the risks the participants actually face.

In the POINTER study, 70% of the participants had a college degree, compared with roughly 43% of the wider U.S. population. "People who are usually in randomized controlled trials are usually highly educated people who are very interested in health," Livingston noted. "And that means they're at less risk of developing dementia, and therefore, less can be done to help them" through lifestyle interventions.

The difference between the test group and the control group in the POINTER study was even smaller than in the original FINGER trial — just 0.029 points.

LatAm-FINGERS tried to maximize flexibility in its approach by better tailoring its interventions to its study participants, said study first author Dr. Lucía Crivelli, a neuropsychologist at the neurology center FLENI in Argentina. Dietary advice reflected the produce available in each region. Exercise programs considered local environmental factors; for example, sessions were held outside, avoiding hot gyms during the sweltering summers in Ecuador and Mexico.

The trial recruited volunteers across Argentina, Brazil, Bolivia, Chile, Colombia, Costa Rica, Ecuador, Mexico, Peru, Puerto Rico, the Dominican Republic and Uruguay. The participants were less highly educated than the POINTER cohort had been; 37% had no education beyond a high school diploma, compared to just 5% in the POINTER study. That likely made the study participants more representative of people who might benefit from the trial's interventions.

When Crivelli announced the results of the trial at the conference, a large cheer filled the room. The intervention had achieved a difference of 0.11 points between the test and control groups' z-scores — more than twice the difference seen in the FINGER trial.

That difference corresponded to a relative 55% improvement in overall brain function in the test group compared to the control group at the end of the study.

What made the difference?

Crivelli, Kivipelto and Livingston all pointed to one key factor in explaining the project's relative success. "In Latin America, there are more modifiable risk factors," Kivipelto summarized.

The study participants had lower academic attainment and higher cardiometabolic risk factors, such as high BMI, than the participants in previous studies. The trial had reaped the rewards of delivering the intervention to people who needed it most, experts concluded.

But the approach wasn't perfect. said Dr. Claudia Suemoto, a neuroscientist at the University of São Paulo who presented a detailed breakdown of the trial's results. They found that the intervention didn't improve men's overall cognitive ability, compared with the control group. Only women benefited.

The difference may be that men simply didn't engage in the trial as much as women did, Suemoto's data suggested. Livingston pointed to the trials' exercise classes as an example of an activity that men may not be as likely to participate in, although that's an assumption rather than being based on the trial data.

Can we really prevent dementia?

The LatAm-FINGERS results show that addressing the dementia risk factors that people face is the most effective way to prevent dementia, Livingston said.

And Livingston is leaning into the idea. She's planning a dementia prevention trial in the U.K. that will ask participants about their risk factors and then give them tailored interventions based on the risks they can realistically avoid.

"We ask them to choose which ones they want," Livingston said. "If somebody doesn't want to stop drinking or smoking, then they're not going to."

While LatAm-FINGERS has been the most successful of the FINGER trials to date, it's still unclear whether such projects actually reduce dementia rates.

Only two previous prevention trials have attempted to measure dementia incidence, and neither saw a notable change in their test group. That said, neither trial included cognitive training interventions like FINGER did; they instead focused on cardiovascular risk factors, such as low exercise rates and high body weight.

Measuring changes in dementia rates requires trials much longer than two years, and they also need massive recruitment. Implementing an intensive study like LatAM-FINGERS in the wider community "is not practical," Crivelli said. Instead, she said the study's interventions would need to be distilled into their essential parts so that they could be implemented at scale.

Kivipelto is optimistic that advances in Alzheimer's screening, including new blood-based tests, could provide an early biological measure to show that FINGERS-style interventions can help stave off dementia. This would be an indirect measure but wouldn't require as long a trial as tracking dementia rates.

Some of the risk factors the Lancet Commission identified, like air pollution, can't be modified at an individual level, Livingston noted. A combination of personal and governmental change could address this, she said, as it has in the realm of infectious-disease prevention.

While the scientific field still has uncertainty about the scale of change that trials like FINGERS can bring about, Beltre has little doubt about the trial's impact.

While she previously relied on help from family members like her grandson to complete everyday activities, she said, "I do everything myself now."

This article is for informational purposes only and is not meant to offer medical advice.

See how much you know about the most complex organ in the human body with our brain quiz!

As recently as 2019, U.S. military health researchers had concerns about testosterone levels among the troops. Their worries were not that male service members didn't have enough of the hormone, as current Defense Secretary Pete Hegseth fears. Instead, experts worried about the troops' inappropriate use of testosterone replacement therapy (TRT), the practice of using injectables or gels to increase a person's circulating testosterone.

"Out of every 1,000 male service members, almost 3 are inappropriately receiving TRT," researchers wrote in a report published on Health.mil, the military's health system website. (That report is available on the Wayback Machine.)

This could lead to harmful side effects, the report authors noted.

"Those being inappropriately treated may experience adverse effects of the medication," including obstructive sleep apnea and edema, the report continued. TRT may worsen sleep apnea and cause swelling in the limbs from water retention. "These adverse effects have the potential to impact deployability and medical readiness."

In short, supplementing testosterone unnecessarily could reduce military readiness, rather than boost it, per the report authors. Now, in a reversal of these concerns, more service members might find themselves offered TRT in the wake of a new military policy announced July 15 by the Department of Defense. It states that all active-duty and reserve service members' testosterone levels will be tested annually starting at age 30.

It is unclear how the policy might affect female service members. Live Science contacted the Department of Defense, which declined to answer questions about the expected testing protocol or criteria for offering service members TRT.

Such an initiative goes well beyond what medical organizations recommend regarding hormone screening for men, experts told Live Science. And while the screening could catch some service members who actually need treatment, the policy might also lead to overtreatment and, in turn, unwanted side effects.

"One of my big, big concerns around this is, how are these tests going to be conducted?" said Luke Cox, a lecturer in sport and exercise science at Swansea University in the U.K. who studies performance-enhancing drugs, including steroidal hormones like testosterone.

Testosterone and military readiness

Testosterone has strong cultural connections with hypermasculinity. As the primary hormone responsible for the development of male sexual characteristics, it's been associated with concepts of strength, aggression and virility. Hegseth continued this pattern by tying the hormone to military readiness in a video announcing the policy change.

"[I]t's about restoring and optimizing your natural capabilities, projecting your longevity and ensuring you have the biological foundation required to sustain the fight," Hegseth said.

That rhetoric aligns with promises made in advertisements for TRT, which pitch testosterone as a tool to "attain that higher level of living," Cox told Live Science. The clinical trial data is more circumspect, however.

In men with genuinely low circulating testosterone levels — often defined as lower than 300 nanograms per deciliter (ng/dL) — TRT does have benefits. (This 300 ng/dL threshold isn't the universal definition for "low testosterone," as not everyone below that line has troubling symptoms. We'll unpack that later.)

Much of the clearest evidence of benefit is for sexual symptoms: In a large clinical trial, men with low testosterone saw increased sexual desire and sexual activity, such as masturbating or having sex, after TRT. For men whose levels are not low, additional supplementation hasn't been shown to have a clear benefit in the bedroom.

TRT has been shown to improve mood and reduce depressive symptoms in men with low testosterone, and in older men, it's also been shown to improve bone density. Boosting testosterone also boosts muscle mass in middle-aged and older men.

Most of this research has been done on men over age 40, because testosterone naturally declines with age. Low testosterone is rarer in younger men, although exact estimates of its prevalence vary. Estimates also vary among older age groups, but a large study called the European Male Ageing Study found that 0.1% of 40- to 49 year-olds had low testosterone, along with symptoms such as fatigue or low sexual desire. That number rose to 5.1% among those in their 70s.

In 2024, 42.8% of active-duty military members, including men and women, were 25 and younger. Only 8.3% were 41 or older. But the prevalence of low testosterone is generally on the rise, said Dr. Joshua Halpern, chief scientific officer and regional medical director at Posterity Health and a reproductive urologist at Northwestern University's Feinberg School of Medicine.

Studies suggest that testosterone levels are lower, on average, in young men today than they were in the year 2000. Scientists have theorized that this trend may stem from increases in body mass index, as well as other factors, such as sedentary lifestyles, changing diets and exposure to environmental toxins, according to a 2020 article in Urology Times.

"While active service members are typically among the healthiest of men, and therefore at lower risk for testosterone deficiency, a significant proportion of service members could have a low testosterone number," Halpern told Live Science. "But that doesn't necessarily mean they should receive treatment."

Treating with testosterone

There is no universal testosterone level that is definitively "too low." The cutoff of 300 ng/dL is based on the point at which symptoms of low testosterone ‪—‬ such as fatigue, low libido and poor mood ‪—‬ tend to present themselves. However, some men have no problems associated with low or low-to-normal numbers.

Plus, testosterone doesn't act in isolation. In fact, some symptoms of "low T" in aging men may actually arise from low estrogen, which stems from these individuals' declining testosterone levels. That's because the male body converts a portion of its total testosterone into estrogen. The two hormones normally work together to regulate sexual desire and erectile function, while estrogen specifically regulates fat accumulation, studies suggest.

There is also a broad range of "normal," running from around 300 to 1,000 ng/dL. For that reason, the American Urological Association recommends testing only if a man is experiencing symptoms that point to low testosterone.

"Making a diagnosis of testosterone deficiency in the absence of signs and/or symptoms increases the likelihood of making a false diagnosis and reduces the potential benefit of testosterone therapy," the association's guidelines state.

Testosterone also has a complicated role in the body. It can affect mood, energy and bone health, but the hormone itself can also be affected by health conditions and lifestyle factors. Studies suggest that intense periods of physical training can temporarily suppress testosterone in military members. That finding tracks with civilian studies that have found that levels can shift dramatically based on people's recent sleep and exercise.

That's why doctors recommend two separate blood tests, both taken in the morning, before diagnosing low testosterone. This helps to ensure that a person's levels are consistently in a low range.

In some cases, it may be that the best treatment isn't TRT but rather an approach that addresses the underlying health conditions that are affecting testosterone levels. This might include assessing whether a patient has sleep apnea and treating it if they do, for example.

Meanwhile, pushing testosterone levels too high can lead to erythrocytosis, an overconcentration of red blood cells that can bring a risk of blood clots and stroke. What constitutes "too much" varies person to person. Due to some testosterone being converted into estrogen, too-high levels can also lead to unwanted side effects like breast growth in some men on TRT.

The other major side effect of TRT is that testosterone replacement shuts off the body's own production of testosterone in the testes. TRT only raises the level of testosterone circulating in the blood, not in the testes where sperm production takes place. Sensing the testosterone in the bloodstream, the brain suppresses its production in the testes, thus diminishing or stopping sperm production. This can cause infertility, and sperm production may take several years to recover after a person stops TRT.

"[F]or young service members who are found to have low testosterone, it will be very important for them and their physicians to understand the risks of treatment and the alternatives that may be more appropriate for preserving their fertility," Halpern said.

Meanwhile, a federal judge has questioned the treatment program. An ongoing lawsuit against the federal government's ban on transgender troops states that "the Armed Forces must adhere to high mental and physical health standards ... without the benefit of routine medical treatment or special provisions." The judge raised questions about how the TRT that will be offered to cis men differs from that which transgender men would use.

This article is for informational purposes only and is not meant to offer medical advice.

When OpenAI recently revealed that two of its most advanced artificial intelligence (AI) models had escaped the confines of a cybersecurity test and hacked into a startup, it sounded a lot like the kind of scenario that AI safety researchers have spent years warning about.

The models found a previously unknown vulnerability in the infrastructure meant to contain them, gained access to the public internet and broke into Hugging Face, a major platform for hosting AI models and datasets. Their objective, however, was less sinister than the sequence of events might suggest: They were looking for information that would help them complete the cybersecurity test OpenAI had given them.

In a July 16 statement, Hugging Face representatives disclosed that internal datasets had been infiltrated, saying it was "different from anything we had handled before" because it was driven "by an autonomous AI agent system." In another statement published July 21, OpenAI representatives fessed up to being responsible, calling the episode an "unprecedented cyber incident" while warning that similar events could become more common as AI models become increasingly capable.

Independent experts who spoke with Live Science agree that what happened is significant — but they cautioned against interpreting it as an AI system suddenly developing a malicious agenda. The models appear to have pursued the task OpenAI gave them, finding a route to success that their creators had failed to anticipate or adequately block.

"If there's a failure here, it isn't that the AI wanted to hack something," Oli Buckley, a professor in cybersecurity at Loughborough University in the U.K., told Live Science. "It's that humans created a test where success was measured by achieving an objective, deliberately relaxed some of the normal security controls to measure the system's capabilities, and underestimated how effective the model would be at finding an unexpected path to success."

How did an OpenAI test end up hitting Hugging Face like this?

OpenAI was testing GPT-5.6 Sol and a more powerful unreleased model using ExploitGym, a benchmark that challenges AI systems to find and exploit software vulnerabilities. The company removed some cybersecurity safeguards that would normally prevent potentially dangerous actions while relying on an isolated environment to keep the models away from the wider internet.

According to OpenAI's postmortem, the models discovered a previously unknown vulnerability in third-party software used to proxy and cache software packages. They exploited it, escalated their privileges and moved through OpenAI's research infrastructure until they reached a machine with public internet access.

Hugging Face became a target because the models identified it as a possible source of information that could help them complete the ExploitGym challenges. OpenAI said at least one attack chain involved stolen credentials and previously unknown vulnerabilities that eventually enabled the models to execute remote code on Hugging Face systems and access test solutions stored in a production database.

In their disclosure, Hugging Face representatives said the company recorded more than 17,000 actions during the intrusion, but they couldn't initially explain who or what was behind it. OpenAI's subsequent disclosure supplied that missing piece: Its models had broken out of their test environment and gone looking for the answers elsewhere.

Did the AI really "escape"?

It's notable that the models found a flaw in the infrastructure designed to contain an AI and used it to reach the public internet. Describing the models as having "gone rogue," however, risks assigning them unsupported motivations, Buckley said.

"I think I'd be wary of jumping to "rogue AI,"" Buckley said. "The models didn't develop their own agenda or decide to attack Hugging Face while twirling their digital moustache."

Buckley compared it to asking a dog to fetch a ball while leaving the garden gate open. "If the easiest ball for it to find is in the park down the road, that's where it'll head," he said. "You wouldn't say the dog had gone rogue; you'd just say you underestimated how literally it would pursue the task."

Daniel Hulme, entrepreneur in residence at University College London and CEO of AI safety company Conscium, agreed that the models shouldn't be assigned human-like motivations. "Models don't have intent; humans have the intent, and we train models with goals in mind," he told Live Science

The capability may matter more than the motive

What matters more than the models' supposed motives is what they managed to accomplish while pursuing their assigned task.

"The genuinely significant point is that the models appear to have chained together multiple vulnerabilities across different systems and sustained a complex sequence of actions," Buckley said. "That demonstrates a level of capability that security professionals should take seriously."

Katerina Mitrokotsa, a professor of cybersecurity and applied cryptography at the University of St. Gallen in Switzerland, said the containment failure is particularly concerning because another company ultimately paid the price.

"What concerns me most is who ended up affected," Mitrokotsa told Live Science. "The victim was not the company running the test, but a third party. This is the scenario security researchers have warned about for some time: that an AI agent's escape does not necessarily stay contained to the environment in which it originated."

OpenAI representatives said they have tightened the infrastructure used for these evaluations. But Mitrokotsa warned that containment becomes harder to guarantee as models improve at performing exactly the kind of exploitation OpenAI was testing.

An AI warning — and an impressive product demonstration

There is also reason to look carefully at how the incident is being framed. OpenAI's account serves two purposes at once: It warns about the security risks posed by increasingly capable AI while demonstrating just how capable its own newest models have become.

Buckley said announcements from frontier AI companies like OpenAI or Anthropic should be viewed in the context of an industry competing to build ever-more-powerful models.

"We've seen similar high-profile capability demonstrations from Anthropic and others," he said. "That doesn't make the findings untrue, but it does mean we should separate the technical evidence from the marketing narrative."

These companies have every incentive to show both that their models are extraordinarily capable and that they are taking the risks seriously, he added. The Hugging Face incident demonstrates both that OpenAI's models carried out a complex series of operations with considerable autonomy and that its security measures failed to keep them inside the experiment.

Hulme argued that the longer-term challenge is ensuring that increasingly capable AI systems pursue their goals in ways that remain consistent with human values.

"Rather than seeking to control AIs, the focus should instead be on alignment," he said, adding that continuous testing will be needed to ensure systems remain aligned with their intended missions while staying secure.

The episode, the experts said, leaves OpenAI with a result that is impressive and uncomfortable in equal measure. Its models found previously unknown vulnerabilities and continued pursuing their goal well beyond the boundaries their creators expected, but none of that requires them to have developed malign intentions.

"The lesson isn't that AI has become malicious," Buckley said. "Instead, it's that increasingly capable systems will exploit opportunities that humans fail to anticipate."

In this incident, OpenAI's new models were given a hacking challenge and they were rewarded for finding a way to solve it. The humans running the experiment simply hadn't anticipated quite how far they might go.

One of the central assumptions of modern neuroscience research is that the brain's shape and structure affect behavior. Scientists have linked a thicker cortex (the brain's outer layer) to higher intelligence, certain brain wave patterns to better volleyball ability, and higher connectivity between parts of the brain to chess-playing skills. There's a vast number of these so-called brain-wide association studies (BWAS).

But when experts repeat these studies, they can't replicate the results.

This "replication crisis" suggests that a substantial amount of brain-behavior imaging research rests on a shaky foundation. There are myriad problems affecting this area of research, said Randy Ellis, a senior scientist at Oracle who wrote about these issues while working as a biomedical informatician at the Icahn School of Medicine at Mount Sinai in New York.

Some of these issues aren't unique to neuroscience. They begin with what Ellis called the "original sin" of science: Academics are put under huge pressure to publish positive research findings.

But some of these factors do apply specifically to this field.

The problems are big enough that they are "halting the growth and development of science and the curing of diseases," Ellis told Live Science.

Tiny differences, small samples

Several brain studies that could not be reproduced in follow-up work show how problems can creep in. Each of the original papers has over 500 citations, with the number of citations reflecting how much these studies may influence thinking in the field.

For example, a landmark study in 2007 found that the brains of kids with attention-deficit/hyperactivity disorder took longer to mature. But a study published earlier this year found that this link vanished once the different rates of aging between boys and girls were taken into account. "That's what made the whole house of cards topple," Matthew Albaugh, co-author of the replication paper and a clinical neuroscientist at the University of Vermont, previously told Live Science.

In 2011, researchers at University College London published a study that found that the density of gray matter, the part of the brain where brain cell bodies are found, in several brain areas was linked to participants' number of Facebook friends. (In 2011, this was a vital social metric.)

A 2015 paper failed to replicate this finding, but the authors of the original paper argued that was because the follow-up study didn't follow the same protocol.

"It's hard to do exactly what was done in the original studies," said Sarah Genon, a neuroscientist at the Research Center Jülich in Germany who was not involved in either the Facebook study or its replication.

So, in 2019, Genon and her colleagues tried a different type of replication analysis.

They used a vast database and conducted more than 10,000 analyses of MRI scans from hundreds of volunteers. The median neuroimaging study sample size was around 25, so they divided this larger dataset into smaller ones and then looked for significant associations within each.

If one of these micro-studies reported a significant association between a behavior and a brain structure, the team tried to reach the same findings using a different subset of the data. Very few of the repeat studies replicated the results of the first. "It was clear evidence that the replicability of those brain-behavior associations is relatively weak," Genon said.

One of the key problems is that in healthy volunteers, variation among individual brains is relatively small, which means you need a large number of participants to detect average differences that tie to behavior, Genon said. That's similar to genome-wide association studies, which pick up the very tiny effects of individual genes, and they must sample tens or hundreds of thousands of people to find robust effects. To detect meaningful differences in brain-wide imaging studies, only sample sizes in the thousands would do the trick, a 2022 paper suggested. Because brain scans are expensive and time-consuming to capture, most studies have relied on much smaller samples.

Additionally, the behavioral tests used to establish psychological variables can be inconsistent, Genon said. What's more, some of these studies may be operating on the outdated assumption that small brain regions control complex characteristics like intelligence, she added. In reality, however, "those types of abilities are usually relatively distributed across the brain," she said. When researchers analyze multiple small brain areas with that assumption in mind using separate statistical tests, it can raise the risk of producing mirage associations.

Some of these studies may be flawed at the outset. A 2017 brain imaging study by researchers at Weill Cornell Medical College explored whether MRI data could be used to stratify patients with depression. The researchers found that brain activity clustered into four "biotypes" of depression, each of which featured distinct, unusual activity patterns in key brain networks.

But a follow-up study found that the differences among the clusters weren't statistically significant, meaning they could have occurred by chance. Richard Dinga, a neuroscientist at the Friedrich Schiller University Jena in Germany who worked on the follow-up study, said the main problems lay in its design.

Its statistical approach, he said, essentially guaranteed that significant links would be detected. That's because of a problem called overfitting, in which a model is so well tuned to one set of data that it struggles to find real patterns in other datasets. This design correlated 17 clinical features of depression with 30,000 brain imaging features, but Dinga said the authors could have used "30,000 coin flips" and still produced the same strong correlation.

Conor Liston, a psychiatrist at Weill Cornell Medicine and co-author of the original paper, acknowledged the model's susceptibility to overfitting. "That was not at all our intention, but that was the result," Liston told Live Science

Still, he argued that the biotype finding they identified was real, pointing to later studies his team had published to back up their original finding. When asked why his team hadn't updated the original paper to acknowledge that their analysis method had serious flaws, Liston said their subsequent publications were sufficient to correct the record. "I think the message is out there, and I think the solutions to those overfitting issues are very clear," he said.

Solving the problem

The results of many studies that share the same issues as these papers will never be replicated and yet will be left to stand. For one thing, there's no money in replication efforts. "Grant agencies wouldn't be really excited" to fund a study whose goal was to confirm past work, Genon said. And even when these studies get done, they aren't as likely to get published when they undermine prior findings.

For Dinga, the way forward is to collect more data. "That's what solved the same irreproducibility problem in genetics," he said, referring to genome-wide association studies that have produced important findings over the past two decades by using sample sizes that have reached the millions. But it's also important to improve study design, he added, since bigger datasets won't fix studies with incorrect analysis approaches. "The methods just don't have a chance to produce something useful," he said.

Some researchers are leveraging large brain imaging datasets. Martin Hebart, a group leader at the Max Planck Institute for Human Cognitive and Brain Sciences in Germany, and Luca Kämmer, a doctoral student in Hebart's lab, are developing re:vision, a project that is collecting imaging data ‪—‬ the largest initiative yet to capture how the brain responds to visual stimuli. They plan to make this information available to other researchers, who will then explore whether hypotheses tested in other papers can be replicated in re:vision's dataset.

Hebart and Kämmer said they have already received over a dozen applications for their project. Reception has been particularly positive among younger researchers in the field. But convincing senior scientists to retest their data might take more work because they have more to lose if their long-standing finding is ultimately not replicated, Kämmer said.

"Everybody always says that replication is great," he told Live Science. "But then when I suggest that their specific studies could be replicated, often they are a bit more skeptical."

See how much you know about the most complex organ in the human body with our brain quiz!

The two major earthquakes that struck Venezuela just 39 seconds apart on June 24 had slightly different epicenters in north-central Venezuela. The first (M7.2) struck near San Felipe, and the second (M7.5) near Yumare, leaving thousands dead and thousands more injured, according to government officials. But beyond the devastation, the sequence opened a rare scientific opportunity: Researchers think the unusual "earthquake doublet" could offer new insight into how large fault systems interact and how some of the most destructive earthquakes grow.

Large earthquakes are typically followed by smaller aftershocks. But particularly intense events can also alter stress on nearby faults ‪or along the same fault, ‪triggering another major earthquake.

These scenarios are infrequent but not unprecedented. The 2023 sequence in Kahramanmaraş, Turkey, and the 1997 earthquake doublet in Harnai, Pakistan, are two well-known examples.

The Venezuelan sequence also reinforces an emerging consensus among seismologists: that treating faults as isolated structures may underestimate the destructive power of quakes in regions where multiple tectonic faults meet, as they do both in Venezuela and around California's San Andreas Fault system. That's a problem, because many of the seismic hazard models in California do not account for those multi-fault interactions.

A natural laboratory for understanding major earthquakes

The fault system involved in the Venezuelan earthquake ‪—‬ which includes the Boconó, Morón, San Sebastián and El Pilar faults ‪—‬ shares several key characteristics with the San Andreas Fault. Both are right-lateral strike-slip fault systems ‪—‬ in which the crustal blocks slide horizontally past each other ‪—‬ located along the boundary between two tectonic plates: the South American and Caribbean plates in Venezuela, and the Pacific and North American plates in California.

Despite these similarities, researchers caution that the two systems differ in important ways.

"The main difference is that the Venezuelan plate boundary has a much more complex fault architecture," Julián García Mayordomo, a senior scientist in the Geological Hazards and Climate Change Department at the Geological and Mining Institute of Spain, told Live Science.

The difference stems largely from the Maracaibo block, whose interaction with surrounding faults creates a much more intricate plate boundary than California's.

"The other difference is the speed at which the plates move," García Mayordomo pointed out.

In Venezuela, the tectonic plates move past each other at about 0.8 inches (20 millimeters) per year, compared with roughly 1.2 inches (30 millimeters) along the San Andreas Fault. Faster plate motion allows tectonic stress to accumulate more quickly, which influences how often large earthquakes occur over long timescales, but not when the next one will strike.

Along the San Andreas Fault, magnitude 7 or larger earthquakes occur, on average every 100 to 200 years, although the frequency of recurrence varies along the fault. The last major rupture in Southern California was the magnitude 7.9 Fort Tejon earthquake in 1857. In Venezuela, estimated slip rates suggest recurrence intervals of one to two centuries. The region experienced two devastating earthquakes in 1812, part of a multiple-rupture sequence that included events of magnitude 7.5, 7.2 and 6.5, and a 2018 study concluded that the Boconó Fault had already accumulated enough strain to generate another major earthquake.

However, these are statistical averages. The recurrence of large earthquakes is highly irregular and depends on a multitude of factors, many of which we still don't fully understand. So a major event could occur in 100 years ‪—‬ or even tomorrow.

Looking beyond individual faults

This uncertainty is precisely one of the reasons the Venezuelan seismic doublet is generating so much interest among seismologists.

"It is the kind of natural event that can sharpen and test the rupture-interaction concepts that paleoseismic models like ours can only infer indirectly," said Liliane Burkhard, a geologist and geophysicist at the University of Bern and first author of a recent study suggesting that the junction between the San Andreas and San Jacinto faults in Southern California is experiencing some of its highest tectonic stress levels in the past 1,000 years, told Live Science.

"Our Cajon Pass work relied on centuries of paleoseismic reconstructions to infer how stress evolves and whether ruptures can cross between fault systems," Burkhard told Live Science. But that doesn't give geologists real-time data, captured by seismic instruments, showing how different faults interact during quakes, she added.

The Venezuela doublet offers exactly that opportunity. The main lesson for California, Burkhard said, is that interactions between neighboring faults can play an important role in the evolution of large earthquakes.

"Whether it is Cajon Pass where the San Andreas and San Jacinto systems meet or the Boconó-San Sebastián in Venezuela, these are precisely the locations where single-fault hazard models break down because the real behavior depends on how stress is shared and transferred between adjacent structures," she said.

Still, the two systems are quite different. The Venezuelan sequence represents a different type of cascading rupture than the one described in Burkhard's research. At Cajon Pass, the "earthquake gate" concept explores whether a single rupture can jump from one fault system to another during the same earthquake, over tens of seconds of rupture propagation along a continuous fault trace. The Venezuelan doublet, by contrast, "looks like two distinct ruptures on what may be two separate fault structures, triggered in close succession," Burkhard said.

For Burkhard, the Venezuelan earthquake reinforces the need for seismic hazard models to move beyond treating faults as isolated structures and instead represent them as interconnected networks. The challenge is particularly relevant in California, where roughly 300 active faults may interact in ways that traditional hazard models do not capture.

New Zealand has already incorporated this lesson. After the 2016 Kaikōura earthquake ruptured at least 12 faults in a single event, New Zealand revised its National Seismic Hazard Model to include complex multifault ruptures.

García Mayordomo argues that both Venezuela and the United States should incorporate these complex rupture scenarios into seismic hazard assessments and building codes. Earthquakes involving multiple faults can produce longer-lasting shaking that increases structural fatigue and, ultimately, the risk of collapse.

"It's like a boxing match," García Mayordomo said. "Many times the winner isn't the boxer who lands the hardest punch but the one who keeps punching for longer."

Even so, researchers cautioned against drawing sweeping conclusions from a single earthquake.

"Each earthquake gives us one possible scenario," Judith Hubbard, an earthquake scientist and structural geologist at Cornell University told Live Science. "The range of earthquake behaviors is wide."

Researchers have demonstrated that a computer worm powered by artificial intelligence (AI) can autonomously spread across a network by identifying and exploiting vulnerabilities on different devices, raising fresh concerns about how the technology could change the future of cyberattacks.

The proof-of-concept malware, developed by researchers at the University of Toronto and cybersecurity firm CleverHans, combines a locally running large language model (LLM) with an autonomous software agent that can scan networks, assess potential attack paths, and decide how to compromise new targets without human intervention. The researchers say the work shows how AI could enable malware to adapt to unfamiliar environments rather than relying on a single preprogrammed exploit.

In experiments described in a new study uploaded June 2 to the arXiv preprint server, the worm was tested against a simulated corporate network containing 33 hosts, including Linux servers, Windows workstation computers and other internet-connected (IoT) devices. The researchers found that the system identified vulnerabilities, compromised new machines, and replicated itself across roughly 62% of the network over the course of a week.

"The main finding is that this type of system can do more than run a fixed exploit; it can examine the target environment, reason about possible vulnerabilities, use tools to attempt attacks, and then replicate itself after a successful compromise," Michael Agee, an adjunct professor of information technology at Trinity Washington University in Washington, D.C., who was not involved in the research, told Live Science.

How does the AI worm work?

The setup was relatively straightforward. The researchers took an open-weight LLM (for which training data is publicly available) running on local hardware and connected it to a software framework that could scan networks, collect information about target systems, and carry out attacks. The AI's role was to interpret what it found and decide where to go next.

"The AI-driven part of the attack is mainly the reasoning and decision-making," Agee said. "The LLM is not magically hacking the system; it is being used to reason about what the information means, suggest possible attack strategies, decide which tool or action should be tried next, and help adjust the approach when something fails."

In other words, the worm isn't inventing new ways to break into systems. Instead, it's taking information about a machine, matching it against known vulnerabilities and weaknesses, and deciding which avenue is most likely to succeed.

Bob Hutchins, who teaches AI strategy courses at Lipscomb University in Nashville, Tennessee, said the innovation lies in the system's ability to adapt.

"Traditional worms follow a scripted sequence: Once a vulnerability is identified, the worm replicates," Hutchins told Live Science. "In contrast, the researchers demonstrated that an easily downloaded AI model could be used as the decision-making component of the worm. The worm would analyze each device it encountered to determine its most effective strategy to breach that particular system."

"Intelligence does not exist in discovering new vulnerabilities; rather, intelligence exists in determining how quickly an attacker can choose and sequence attacks against previously identified vulnerabilities," he added.

What makes this AI worm different from conventional malware?

The researchers also designed the worm to work across devices with different levels of computing power. More capable compromised machines equipped with graphics processing units (GPUs) could provide reasoning services for lightweight agents running on less-powerful devices elsewhere on the network.

"What made it particularly dangerous was a clever tiered design," Tom Vazdar, a professor of AI and cybersecurity at the Open Institute of Technology, told Live Science. "GPU-equipped compromised machines provided reasoning capacity for lightweight agents running on low-power IoT devices that couldn't run an AI model locally. A camera becomes a thinking node in the attack network, not just another door."

The research, which has not been peer-reviewed yet, was published as governments, security experts and AI companies continue to debate whether generative AI will make sophisticated cyberattacks easier to carry out. One reason the study has attracted attention is that the researchers did not rely on a frontier model from a major AI company, like OpenAI's ChatGPT or Anthropic's Claude. Instead, they used a much smaller open-weight model that can be downloaded and run offline on normal computers.

"The researchers employed lightweight open-weight models during their demonstration, which are relatively easy to download, remove guardrail components from, and utilize," Hutchins told Live Science. "By using these types of models, the researchers challenged a long-standing assumption that only advanced/edge-type models present risks related to misuse."

Vazdar argued that the work highlights how attackers could increasingly automate tasks that currently require skilled human operators, telling Live Science, "The attacker's marginal cost drops to essentially zero. And you can't patch your way out of it, because it doesn't rely on a single vulnerability class. It reasons. Patch one hole, and it finds another."

Could attackers use this AI worm in the real world?

Not all experts agree with that assessment, however. Although researchers described the system as capable of targeting a wide range of devices, some cautioned that the demonstration took place in a highly controlled environment designed to showcase the concept.

"This is at best a lab-based proof of concept in a target-rich test environment," Agee said. The test network contained many intentionally vulnerable systems and lacked active endpoint defenses. "The paper shows that the approach is possible, not necessarily that this attack would work reliably in a normally, or even minimally, defended enterprise network," he added.

The worm also generated activity that security teams could potentially detect, he noted, including network scanning, repeated exploitation attempts and privilege-escalation behavior.

"Even a basic monitoring setup could flag some of that behavior," Agee said.

Hutchins likewise warned against overstating the findings. "'Could potentially target almost any device' is technically correct and emotionally misleading," he said. "Any internet-connected device running vulnerable versions of software is theoretically susceptible to being exploited via a similar mechanism. This has been a truism of malicious code for decades."

Organizations can still defend themselves by using many of the same measures recommended against conventional cyberattacks, Hutchins added, including prompt patching, strong passwords and multifactor authentication (using multiple forms of identification to log in to systems, like a password sent via text message on top of your password).

Even so, experts broadly agree that the study could mark a shift in how malware could operate in the future. Rather than relying on fixed instructions written by human attackers, future malicious software may be able to make many tactical decisions on its own.

"The attack is important because it shows that an LLM-based agent can reason through different targets and adapt its approach," Agee said.

For Hutchins, the study ultimately represents exactly the kind of work academic researchers should be doing. The study authors "are performing precisely what academia should perform ‪—‬ researching a legitimate threat within a controlled environment before malicious actors begin building it outside of that controlled environment," he said.

Whether attackers adopt similar techniques remains to be seen. What the researchers have shown is that a relatively small AI model can already play a meaningful role in planning and directing a cyberattack.

A Chinese spacecraft has arrived at one of Earth's temporary "quasi-moons" and will soon attempt to land on the space rock to scoop up samples that will be returned to our planet next year. The ambitious mission could help researchers uncover secrets about Earth's main moon and provide clues that may benefit space exploration and planetary defense, experts say.

The Chinese National Space Agency (CNSA) probe, named Tianwen-2, launched from Xichang Satellite Launch Center in southern China on May 28, 2025, Live Science's sister site Space.com reported at the time. The car-sized spacecraft sent back its first images a few weeks later, giving us our first glimpse of its secretive design, followed by a selfie with Earth a few months later.

The mission's primary target is 469219 Kamo'oalewa — also known as 2016 HO3 or simply Kamo'oalewa — a near-Earth asteroid discovered in 2016 by astronomers at Hawaii's Haleakala Observatory. It likely spans somewhere between 130 and 330 feet (40 to 100 meters) across and is a "quasi-satellite" of our planet, meaning it's orbiting the sun right alongside Earth in a temporary partnership. (Kamo'oalewa means "oscillating celestial fragment" in Hawaiian.)

Tianwen-2 is predicted to have arrived at Kamo'oalewa on June 7 and inserted itself into an orbit alongside the asteroid, according to an unverified timeline shared by freelance journalist Andrew Jones, a leading expert on China's space program. Around a month later, on July 4, the probe is expected to descend upon the space rock to collect samples from its surface, potentially employing a never-before-seen drilling technique to do so. (The CNSA has not released an official timeline for the mission or shared any recent updates about its progress.)

If successful, the samples are expected to be returned to Earth on Nov. 29, 2027, according to the same timeline. At this point, Tianwen-2 will then slingshot around our planet and head farther out into the solar system for its secondary mission: to rendezvous with and study 311P/PanSTARRS — a peculiar object beyond Mars, which displays characteristics of both comets and asteroids — in 2035.

Samples from Kamo'oalewa could help answer many interesting questions, including the true origin of the object, which some experts believe may have broken off from the moon.

"What makes [this mission] extraordinary is that we don't yet know [the object's] composition or origin," Li Chunlai, a researcher with the Chinese Academy of Sciences' National Astronomical Observatories and chief commander of Tianwen-2's ground application system, said in a state-sanctioned video translated to English. "We'll only obtain definitive answers after completing our exploration."

What is Kamo'oalewa?

There are currently eight known quasi-moons orbiting the sun alongside Earth, following the most recent quasi-moon discovery in September 2025. Unlike the similarly named "minimoons," which are asteroids that are briefly captured by Earth's gravity for a few months or years at a time, quasi-moons are temporary satellites that slowly circle our planet but are never gravitationally bound to it.

This relationship can last decades, or even centuries. For example, experts think Kamo'oalewa has been accompanying our planet for at least 100 years and could remain alongside it for another 300 years before it starts to fall out of sync with our solar orbit.

"The way Kamo'oalewa moves with Earth is kind of like a dog that might tag along with you for a while on a long walk through the woods, but it's not your dog," Richard Binzel, a planetary scientist and asteroid expert at MIT, told Live Science in an email.

The Ferris-wheel-size space rock is one of the closest quasi-moons to us, with a minimum distance of around 2.9 million miles (4.6 million kilometers) from Earth, or around 12 times farther away than our actual moon. It is also one of the fastest-spinning quasi-moons, rotating around its axis every 30 minutes. But unlike other quasi-moons, which likely originate from the solar system's main asteroid belt, between Mars and Jupiter, our current best guess is that Kamo'oalewa is actually a piece of the moon.

This hypothesis is based on observations from 2021 that showed Kamo'oalewa absorbs and reflects light similarly to the moon, hinting that the space rock may be a chunk of lunar material ejected by a long-ago meteor strike. More recently, in 2024, researchers narrowed down the most likely impact crater from which the asteroid originated: the 13.6-mile-wide (22 km) Giordano Bruno crater, located on the far side of the moon. However, there is still a chance that Kamo'oalewa is a main-belt asteroid that just happens to look an awful lot like lunar material.

"I am curious to find out the answer about its origin, since the debate on its [potential] lunar origin is still very open," Marco Fenucci, a mathematician at the European Space Agency's Near-Earth Objects Coordination Centre who has co-authored multiple studies on Kamo'oalewa, told Live Science. This mission should "definitely give us an answer to this matter," he added.

Scientists are also unsure whether Kamo'oalewa is a "solid monolithic rock," like the $100,000 quadrillion asteroid Psyche (the target of NASA's ongoing Psyche mission), or a "rubble-pile-like asteroid," such as Bennu or Ryugu, both of which have been explored by sample-return missions in recent years, Fenucci said.

Scooping up samples

Having arrived at Kamo'oalewa, Tianwen-2 will spend the next few weeks mapping the asteroid's surface at altitudes between 1,000 feet and 12.5 miles (300 m to 20 km) above its surface, according to limited mission details revealed by Chinese scientists in a January study. The resulting map will allow researchers to estimate the asteroid's density and work out the best landing spot.

On July 4, the probe is expected to swoop down and scoop up around 3.5 ounces (100 grams) of surface material, or regolith, from the asteroid, which is slightly less than what NASA's OSIRIS-REx mission managed to grab from Bennu. However, there are multiple ways this sampling process could play out.

If the asteroid is a pile of rubble, the probe will utilize a touch-and-go approach, which involves hovering just above the surface and using a robotic arm to reach out and grab the sample. This method was also used by OSIRIS-REx for Bennu and by Japan's Hayabusa2 mission, which scooped up samples from Ryugu.

But if Kamo'oalewa has a more solid surface, Tianwen-2 will attempt to land directly on the space rock and anchor itself in place before drilling into the asteroid. This has never been attempted before and could allow for longer contact time and "more controlled sampling," according to The Planetary Society.

After collecting samples, the probe will remain alongside the asteroid and conduct further analysis of its size, shape and trajectory.

In late April 2027, Tianwen-2 will depart Kamo'oalewa and begin the six-month journey back to Earth, where it will dump its samples in a return capsule that will fall to our planet's surface as the probe begins its long journey to rendezvous with 311P/PanSTARRS.

The return capsule will reenter Earth's atmosphere at around 27,000 mph (43,500 km/h), researchers wrote in the January study. That's slightly faster than the Orion capsule carrying the Artemis II astronauts traveled as it fell back to Earth in April.

What could we learn from Tianwen-2?

In addition to ending the debate over Kamo'oalewa's origin, samples from Tianwen-2 could reveal more about the solar system and our place within it.

For example, previously returned asteroid samples have helped to "tell us where Earth's water came from, how complex organic molecules were distributed in the early solar system, and how surface processes like space weathering alter asteroid properties over time," Fenucci said. These samples can also be used to "establish a link to known meteorites found on Earth" and help us learn more about these space rocks, he added.

The proximity of quasi-moons and minimoons to Earth also makes them valuable assets for future space exploration missions. Some researchers have proposed using them as temporary bases or refueling stations that could serve as stops on the way to Mars or other faraway destinations. These objects "could become literal stepping stones for human missions to practice what it is like to leave the Earth-moon cradle," Binzel said.

One particular idea that scientists want to explore is whether water can be extracted from these objects. And if Kamo'oalewa is a piece of the moon, that could have implications for NASA's planned moon base. "Someday, it might be possible to extract that water to provide resources for thirsty astronauts or even generate usable rocket fuel from the hydrogen," Binzel said.

Some near-Earth objects also pose a potential risk to our planet. And while Kamo'oalewa will likely never hit us (and would probably only cause localized damage if it did), there are much larger nearby space rocks that could trigger global catastrophes if they were to impact our planet.

"Over the time scale of centuries, these kinds of objects have a chance of striking the Earth," said Binzel, who invented the Torino scale used to measure the impact potential of near-Earth objects. Therefore, a better understanding of space rocks like Kamo'oalewa "can prepare us in how to deal with a threatening object if one were ever discovered," he added.

Finally, this mission will likely provide another opportunity to assess China's space exploration capabilities, which could hint at whether the country could overtake the U.S. as the top nation in space, as some experts predict.

As the World Cup kicks off across the U.S., Canada and Mexico, scientists have predicted that extreme heat could play a decisive role in determining the outcome.

Nearly half of the scheduled matches — especially those in more southerly areas of the U.S. and low-lying parts of Mexico — have at least a 50% likelihood of "performance-impairing heat." To prepare, many teams have been training intensively in hot conditions, and during the tournament, they will have access to cold-water immersion tubs, ice vests and misting fans to help cool players down if they overheat.

Compared with these major sporting events, which have resources to adapt to extreme heat, community sports participants and recreational exercisers often have fewer options. This lack of adaptability may increasingly affect when, where and how people can safely exercise, especially as climate change makes extreme warming events more frequent and intense.

"The majority of people who play sport for competition or just for fun, recreation and leisure have a fraction of that power [of professional sports organizers] but face greater challenges and harms," Jessica Murfree, an assistant professor in the Department of Exercise and Sport Science at the University of North Carolina at Chapel Hill, told Live Science.

How hot is too hot for exercise?

Sports governing bodies typically use wet-bulb globe temperature (WBGT) to assess heat risk. It measures the combined impacts of temperature, humidity, direct sunlight and wind speed on the human body.

Taking humidity into account is key because, with more water in the air, it's harder for sweat to evaporate off the skin and cool the body. WBGT was developed by the U.S. military in the 1950s to prevent heat-related deaths in training camps. Nowadays, it also shapes safety guidelines for outdoor workers, as well as athletes.

The American College of Sports Medicine divides the U.S. into three regions to account for varying levels of heat acclimatization and defines WBGT ranges that pose a "very high" risk in each region. That threshold is a WBGT of 82.2 degrees Fahrenheit (27.9 degrees Celsius) or higher in Southeastern and Southwestern states, 79.9 F (26.6 C) in Central states, and 76.3 F (24.6 C) in Northern regions.

For unacclimatized individuals and those with poor physical fitness, the health risk jumps at lower thresholds.

Gulf Coast states — particularly Southern Texas, Louisiana, Mississippi, Alabama and Florida — tend to experience the highest WBGT values. High-risk conditions begin in May in South Texas and South Florida, before expanding northward and peaking in July and August. Studies suggest the number of extreme humid heat events has more than doubled in most parts of the U.S. since 1980, and that globally, climate change added roughly three weeks' worth of dangerous humid heat in 2024 alone.

An individual's overall health and fitness, hydration levels and amount of exertion need to be considered, and systemic factors like socioeconomic status also come into play, Murfree said. For instance, lower-income families are more likely to live in hotter urban environments and lack access to air conditioning, meaning they may already be experiencing heat stress at home.

While high humidity and heat make it so sweat can't evaporate easily, the body has other means to cool down. It also sends more blood to the skin, arms and legs, allowing heat to be released through the skin. But physical exertion in extreme heat makes this more difficult.

"We only have a limited amount of blood in our body, and we're simultaneously trying to use it to fuel the reactions in our muscles to be able to facilitate exercise, as well as send it to our skin to help us cool down," said Grant Lynch, a research fellow in the Heat and Health Research Centre at the University of Sydney.

When the body can no longer get rid of extra heat, symptoms of heat exhaustion ‪—‬ including dizziness, nausea, headache and muscle cramps ‪—‬ may set in. If the body continues to overheat, this can progress to heat stroke, a serious and potentially fatal medical emergency where a person's core body temperature rises to 104 F (40 C) or above. Heat stroke can cause central nervous system dysfunction and delirium, convulsions or coma.

In the U.S., exertional heat stroke is a leading cause of preventable death during sport and exercise and is a particular concern for youth athletes. Children, including teens, produce more heat relative to their body weight than adults do while sweating less. Additionally, they may feel less able to advocate for themselves on the playing field if they start to feel ill, Murfree said.

Reducing health risk in hot weather

The body can acclimate to extreme heat, to some extent. At the University of Sydney’s thermal ergonomics laboratory, researchers use a climate chamber to test the body’s capacity to adapt by increasing its blood plasma volume and lowering heart rate and resting temperature.

"We would have them come in for anywhere from five to 14 consecutive days of exposure," Lynch said. "They would be cycling or running for between 90 minutes to 2 hours every day, and the conditions would be [95 F to 104 F] 35 C to 40 C in the room for that entire duration."

The benefits achieved through this kind of heat acclimation aren't permanent. Research suggests that, for every two days spent in a cool environment, one day of heat acclimation is lost. For people exercising just a few times a week in hot weather, "it's pretty unlikely you are physiologically adapting as much as you think you are," Lynch said.

While WBGT is considered the "gold standard" for heat safety monitoring, the special device required to measure it may not be accessible to every community sports organization or individual person exercising. The danger, Lynch said, is that many people focus on ambient temperature alone to determine risk, when multiple factors are involved and can compound.

Sports Medicine Australia, the country's national body for sports medicine and sports science, developed a tool to calculate risk and suggest safety precautions. It combines data automatically extracted from weather stations — including ambient temperature, humidity, wind speed and solar radiation — with sport-specific information, such as typical uniforms and duration of activity.

Generally speaking, the sports body recommends staying hydrated before, during and after physical activity; wearing lightweight and breathable clothing; taking frequent rest breaks using ice towels and misting fans when possible; and delaying or cancelling games when the risk level gets too high. If someone begins to feel overheated, immediate action is important to prevent further heat strain.

"The most important thing to do, without a doubt, is to stop or reduce the intensity of the activity you're performing, seek shade, go into an area where you are not exposed to direct sunlight, and then douse the skin with water," Lynch said.

Heavy sweating, painful muscle cramps, nausea, dizziness, and cool, pale, clammy skin are all signs that someone may be suffering from heat-related illness. Heat stroke symptoms include a very high temperature; hot, dry skin or profuse sweating; confusion; and loss of consciousness. These symptoms should prompt immediate cooling with a cold water bath and emergency medical care.

While rescheduling events to cooler parts of the day is an option, heatwaves are already increasing in most regions and that trajectory is expected to continue as average global temperatures rise. So, some researchers and sports organizations are exploring infrastructure changes that could help: stadiums with shaded and green spaces, schoolyards with grass instead of concrete, and artificial turf with a built-in irrigation system for use in school sports fields.

Murfree said she would like to see heat safety training introduced in schools in areas prone to extreme heat in the U.S., similar to existing fire safety education.

"Sport often asks us to push ourselves harder, be faster and tougher," Murfree said. "It's OK to take breaks, stop, slow down, reschedule and advocate for our bodies and our wellness in the heat."

This article is for informational purposes only and is not meant to offer medical advice.

How much do you know about the greatest athletic feats of all time? Find out with our sports record quiz!

When the U.S. faces Paraguay in the World Cup in Los Angeles' SoFi Stadium on June 12, the teams will play on a natural-grass surface that, a month earlier, was carefully stitched together atop an underlying artificial-turf field. This "hybrid turf" installation was part of a broader effort in advance of the World Cup to ensure the world's best soccer teams played on predominantly natural-grass surfaces.

FIFA, the organization that oversees the World Cup, has prohibited the use of synthetic turf in World Cup games for years, mainly because of an increased risk of player injury, which has been well established. Synthetic turf is also notorious for causing nasty turf burns, as the material is a poor solar radiator and can reach up to 200 degrees Fahrenheit (93 degrees Celsius).

However, many people in the wider sports community worry that synthetic turf may pose more insidious harms, such as dangerous levels of exposure to heavy metals, cancer-causing chemicals and microplastics.

A recent, 10-year-long study from California found no significant risks from the chemicals in artificial turf. But experts aren't convinced the study has laid to rest concerns about artificial turf. Here's why.

Wonder of the Space Age

This infill is the crux of the issue: All those tire crumbs contain hundreds of chemicals and "created this exposure situation that had never existed previously," Rachel Massey, an environmental health researcher at the University of Massachusetts Lowell Center for Sustainable Production, told Live Science.

More than 95% of the 18,000 to 19,000 turf fields in the U.S. now include tire-crumb infills. AstroTurf has remained a dominant player offering 3G products. (Live Science reached out to AstroTurf to discuss the company's products but did not receive a response.)

Harmful chemicals?

Some researchers are concerned that turf contains nearly 400 chemicals that have been tied to various health risks, ranging from asthma to leukemia.

"None of those [risk assessment] studies dispute the fact" that those chemicals are present, Massey said. "So then all the debate is about exactly how much exposure there is."

To assess whether that exposure reached harmful levels, the California Office of Environmental Health Hazard Assessment (OEHHA), part of the California Environmental Protection Agency, undertook a study of the health impacts of chemical exposure due to synthetic turf.

"We identified as many chemicals as we could in the crumb rubber by doing non-targeted sampling and assessed as many chemicals as we could," said Amy Gilson, a spokesperson for the OEHHA.

They attempted to measure exposure to metals commonly found in turf infill, including lead, cadmium, manganese and zinc, with zinc being most frequently detected at concerning levels above U.S. regulatory guidelines. Although zinc is essential to the body, it can be harmful at high levels, with the potential to cause anemia, damage the pancreas, and reduce fertility.

They also looked at carbon black, a reinforcing tire filler material that has been shown to kill human cells in lab dishes; benzene, which has been linked to several types of leukemia; toluene, which is toxic to the nervous system, liver and kidneys; and polycyclic aromatic hydrocarbons, compounds that are most commonly found in tobacco and wood smoke and have been tied to cancer in animal studies.

Another potential concern emerged last year, when 1,3-dimethylbutylamine (1,3-DMBA), a stimulant banned by the World Anti-Doping Agency, was detected in post-match samples submitted by eight Norwegian female professional league soccer players from opposing teams. The artificial turf was found to be the source of the banned substance. With tire wear, 6PPD ‪—‬ a substance often added to tires to limit ozone-depleting emissions ‪—‬ can degrade into 1,3-DMBA, which was detected on the turf where the game was played.

The California EPA report also looked into the presence of microplastics in turf.

In 2023, the European Union announced a ban on the sale of products containing intentionally added microplastics, effective starting in 2031. Specifically, it named the granular infill within artificial turf as "the largest source of intentional microplastics in the environment." The U.S. does not have similar legislation.

Microplastics might harm human health by accumulating in various parts of the body through ingestion, skin contact or inhalation.

Early research in human cells in lab dishes hints that microplastics could damage the lining of blood vessels and promote the formation of atherosclerotic plaques and blood clots. Other studies have proposed potential links between microplastic exposure and asthma and inflammatory bowel disease. However, there is a lot of uncertainty about how microplastic exposure is measured and whether this exposure causes diseases or is simply a proxy for some other factor that causes these conditions.

"Negligible" long-term harm

The California EPA focused on chemicals in tire infill that could potentially cause harm at high doses. Then, it estimated the levels of exposure in players, referees and spectators through having direct contact with, breathing in, or inadvertently swallowing tire crumbs. First, they estimated exposure indirectly, by observing people.

"We actually had teams of people going out and videotaping players on the field," Gilson said.

The research team also collected samples of artificial turf and then put them in lab dishes with synthetic bodily fluids such as sweat, saliva and small intestine fluid and then measured the concentrations of dozens of potentially harmful chemicals within those fluids. To be more conservative, they assumed every chemical detected in their experiments would have been completely taken up by human cells, Jocelyn Claude, a toxicologist at the OEHHA who worked on the study, told Live Science.

The study found "no acute risks," meaning short exposure to this substance is unlikely to cause immediate harm, she said, and the risks of longer-term health effects, such as cancer or reproductive harm, "were negligible."

But some experts said the study methodology was flawed. Andrew Watterson, a professor emeritus of public health at the University of Stirling in the U.K., said the study should have taken urine or expelled-air samples directly from real people exposed to artificial turf, instead of estimating exposure in synthetic body fluids.

Wrong approach to risk assessment

For Massey and several environmental researchers Live Science spoke to, the bigger problem is that the underlying approach used by the California study — and most other environmental assessments done in the U.S. — is flawed.

Instead of asking whether an average person is at risk, society should aim to reduce exposure across the population, which also protects people at higher risk. Instead of asking, "Is this risk acceptable?" the latter approach asks, "How can we create the healthiest, safest playing surface for a child?" Massey said.

The California report included referees and spectators in addition to players, but it may have missed other at-risk groups, such as people exposed to chemicals that have leached into the water supply from synthetic turf, said Tracy Stewart, a Medway, Massachusetts resident who advocates against the use of artificial turf.

Another byproduct of 6PPD, the chemical tied to doping, has been shown to cause mass die-offs of coho salmon. So when dead salmon started turning up near an artificial turf field in Vancouver where scientists had detected 6PPD-Q and a range of other chemicals in the stormwater runoff, the community was worried.

The California study estimated 6PPD-Q exposure levels in synthetic bodily fluids, but it did not model how groundwater would be affected. And without well-established safe limits for this chemical, the California study couldn't make assessments of the risk it poses.

Beyond 3G turf

The attraction of synthetic turf is all about "increased hours of use," said Garnet Brownbill, a spokesperson for the Natural Turf Alliance, a nonprofit that advocates for alternatives to synthetic turf across Australia. "It's not impacted by the weather, it never closes, and people claim it's maintenance-free — it's not, by the way," he told Live Science.

However, in response to the looming EU ban on microplastics in infill, many scientists are working on alternatives that can provide that same advantage.

One material in development is hybrid turf, whose surface is "99.5% natural grass," with an additional 0.5% stitching material composed of either plastic or a biodegradable polymer, said Aaron Golembiewski, head of business development at Turftalents, a company that consulted for FIFA on the installation of hybrid turf at several U.S. stadiums. Hybrid turf is "stronger than if it's grass alone," Golembiewski told Live Science. "It increases [the] quantity of play."

It's also smoother than ordinary grass, which has naturally occurring bumps and divots. He noted that hybrid turf is cheaper to install from scratch than synthetic turf and lasts a few years longer, although it needs more maintenance.

Grass fields may be sufficient for many communities. While some stakeholders figure out how to maintain grass more sustainably, other scientists are working to improve the grass itself. "If the same amount of engineering went into natural grass, there's all of this stuff that we can do," Brownbill said. Not all grass is the same.

"There are high-heat-resilient grasses; there are shade-resilient grasses; there are high-wear grasses," Brownbill said. "[Municipalities] are basically using 100-year-old technology … the same grass cultivars that existed 100 years ago."

Editor's Note: This article was produced as part of the Dalla Lana Fellowship in Journalism and Health Impact program at the University of Toronto.

How much do you know about the greatest athletic feats of all time? Find out with our sports record quiz!

Emerging diagnostic tests for cancer aim to detect the disease in its earliest stages, to improve patients' chances of survival by enabling earlier treatment. Some of these diagnostics — called multi-cancer early detection (MCED) tests — screen for many cancers at once, which sounds great on paper.

However, in February, a flagship trial testing one such cancer test, called Galleri, failed to meet its primary endpoint: it didn't reduce the number of late-stage cancer cases identified. In May, at the American Society of Clinical Oncology (ASCO) meeting in Chicago, attendees got a closer look at some of the data behind the trial.

The trial was the biggest of its kind in the field to date and its high-profile failure drew a lot of attention. But experts told Live Science that it doesn't mean we should count out MCED technology.

Why Galleri was put to the test

The majority of cancer diagnoses and deaths are from cancers that have no well-established screening program, such as pancreatic cancer. Without these early-detection methods, like regular colonoscopies for colorectal cancer, many cancers are caught late.

The perfect MCED test could, in theory, make finding cancer a breeze. Such a test could use a single blood sample to screen for many cancers, detecting disease-linked changes in DNA and other cellular markers of disease floating in the bloodstream.

But designing a trial that proves that MCEDs can improve patient outcomes is difficult. Earlier diagnosis can be life-saving in some types of cancer, said Jessica Lloyd, a strategic evidence manager at the nonprofit Cancer Research U.K. Five-year survival rates drastically improve when lung or colorectal cancer are detected at an earlier stage. But for other cancers, such as melanoma, earlier diagnosis makes little difference to overall survival rates.

The ideal trial would follow a group of patients over many years; one half would receive MCED testing while the other would not. At the end of the trial, data would reveal whether overall cancer deaths were reduced in the group that received testing.

But there's already a lot of hype around MCED tests, and there's little appetite in the field to wait years for clear results that the tests improve outcomes, or to pay the high price tag that such long studies would incur.

Instead, the recent trial of Galleri, run in collaboration with the U.K.'s National Health Service (NHS), lasted only three years. It measured whether the test could identify cancers at an earlier stage, considering the early detection a proxy for mortality data. In other words, early detection was assumed to lead to lower death rates.

The trial recruited 142,000 healthy people ages 50 to 77. Participants had their blood drawn roughly once a year for three years. Each time, for a subset of participants, the Galleri test looked for markers linked to 12 different types of cancer. If a patient tested positive, they received further treatment. This tested group was compared to another that didn't have their blood tested.

The aim of the trial was to assess whether significantly fewer cases of late-stage cancers — meaning stages III and IV — would be detected in the test group; at that stage, disease has started spreading beyond the cancer's origin point to other tissue. The rationale was that the test would instead pick up these cancer cases in stages I and II, before the disease had spread. But Galleri failed to meet this endpoint.

Samantha Harrison, Lloyd's colleague and head of strategic evidence at Cancer Research U.K., said that the test's failure to reach its main target is "disappointing." However, she added that the trial's huge dataset hasn't been detailed in a peer-reviewed publication yet, which would shed more light on the trial's full value.

What went wrong for Galleri?

Galleri's problems were related to both the diagnostic test itself and the trial design, said Ruth Etzioni, a biostatistician at Fred Hutchinson Cancer Center in Seattle. Galleri caught an unexpected number of advanced cancer cases among the trial volunteers the first time they were tested.

These cases came as a surprise, because the volunteers recruited to the trial had not been diagnosed or treated for cancer over the previous three years. Etzioni said that if the trial were begun anew, these initial cases would likely be excluded from the analysis, which would have helped improve Galleri's performance.

But even with that adjustment, she added, it's unlikely the trial results would have significantly changed. That's because the test was unexpectedly bad at picking up early-stage cancer, she told Live Science.

"That's critical," she said. "You have to have an increase in stage I to II, and then that then coincides with the decrease in stage III to IV," she said.

The MCED test doesn't look for a lone marker that determines whether a patient has cancer or not. Instead, the various disease markers that the test looks for are weighed and a threshold is applied, above which the test yields a positive result. Etzioni said the test appeared to be conservative in how it registered these positive results; it applied a high threshold to deem a case "positive." That may explain why the harder-to-detect, early-stage cancer cases were not picked up.

But she added that some of the more dismissive commentary around Galleri's performance is "short-sighted."

"I don't think that this trial was a flop," she said. "I think it was hugely informative."

Galleri hopes that results from its upcoming Real-world Evidence to Advance Multi-Cancer Early Detection Health Equity (REACH) study, which has enrolled 50,000 patients in the U.S., will be more positive. The REACH trial has a slightly different endpoint in that it's measuring the test's ability to reduce stage IV diagnoses. The data from the NHS trial suggested Galleri did reduce these diagnoses, which was a secondary endpoint of the trial.

What does this mean for other MCED tests?

Galleri is just one of many MCEDs under development. What could the negative results from Galleri's initial trial mean for the future of cancer detection?

MCED tests are proliferating rapidly, and improvements to the panels used to detect cancer markers should enhance the performance of newer tests, said Etzioni. That said, she noted that, as these tests are closely guarded by manufacturers, it's hard to know exactly which markers would advance performance. Dozens of presentations at ASCO 2026 focused on other MCEDs in the field.

Galleri's results might mean a slightly longer wait before the tests are used widely, Etzioni said. She felt that some of the hype from GRAIL's earlier studies of the test has now dissipated. "They've maybe [previously] led the community to believe that the technology is more ready for prime time than it actually is," she suggested.

Harrison pointed to a position statement from the U.K. National Screening Committee (UK NSC), which provides advice to government ministers on implementing new disease screening programs. The statement, released in May, directly addressed so-called surrogate endpoints, such as the cancer stage-related endpoints used in the initial Galleri trial.

One concern, it pointed out, is that MCED tests may more easily detect aggressive cancers than slow-growing ones. So even if future trials show the tests find these cancers at earlier stages, patients may still die at similar rates because the types of cancer found at these early stages are harder to treat. So improvements in detection can't be assumed to translate to improvements in survival.

The UK NSC stated that further evaluation would likely be required before a MCED test would be adopted in widespread clinical practice, such as within the NHS. (For what it's worth, though, some MCEDs are available for direct-to-consumer purchase despite not proving their worth in gold-standard trials yet.)

The Galleri trial focused on the test's ability to identify cancer in a (presumably) healthy population, but this is just one potential use of the technology behind the tests. It could also be deployed to test people with non-specific symptoms that may hint at cancer or in populations who are known to be at high risk of the disease. Another NHS Galleri trial, called SYMPLIFY, suggested the test was accurate when used in symptomatic patient groups, Lloyd noted.

Galleri's results are not a reason to discount the value of MCED tests more widely, Etzioni said. She believes the tests will someday be used routinely for early cancer detection.

"This is just the beginning," she said.

This article is for informational purposes only and is not meant to offer medical advice.

Microsoft has revealed a new quantum computing chip with quantum bits (qubits) it says are capable of maintaining their quantum state for 1,000 times longer than its predecessor — paving the way for more reliable quantum computers by 2029. But not all scientists believe the company's claims.

The experimental quantum processing unit (QPU), dubbed Majorana 2, features a four-qubit array that offers a reported mean qubit lifetime of 20 seconds and, in some instances, up to a minute. This is a massive improvement in quantum coherence times — the time that qubits are entangled so that calculations can run in parallel — typically seen in QPUs. Normally, this lifetime is measured in milliseconds (thousandths of a second).

The new chip could put scientists on the path to building a quantum computer that's commercially viable by 2029 — halving the timespan researchers initially expected — Microsoft representatives said in a statement. The scientists who worked on the new processor outlined their findings in a June 2 preprint study, and the results have not yet been peer-reviewed.

"We need to make improvements each year that will get us closer to delivering a computer that we believe will have massive commercial and societal value," Chetan Nayak, Microsoft technical fellow, said in the statement. "We've got to keep marching to that roadmap to accomplish that, but where are we relative to last year? We’re 1,000 times better."

Despite the claimed progress against the first chip, Majorna 1, experts have called Microsoft's work in this specific niche of quantum computing research (called topological quantum computing) into question. They have previously questioned whether the underlying technology has yet been proven and have called for a wider evidence base for suggestions on qubit coherence times.

Despite the criticism, Microsoft representatives say this has halved the development time in building a future fault-tolerant quantum computer — a machine that can overcome errors and sustain long-duration calculations to potentially outperform supercomputers.

Next-generation topological qubits

The Majorana 2's predecessor was revealed in February last year. Both chips are based on a 90-year-old theory by Italian physicist Ettore Majorana that a particle could be its own antiparticle, meaning that it either annihilates itself in a massive release of energy or coexists stably when paired, enabling it to store quantum information as a qubit.

Because Majorana particles aren't found in nature, much of the research into them, including Microsoft's previous findings, centers on nudging them into existence.

Under the right conditions, the qubits in these chips can reach a "topological" state of matter — a specific phase in which atoms are entangled over long distances — which lets them tap into the laws of quantum mechanics to process the 1s and 0s of computing data in parallel.

Representatives said on the launch of Majorana 1 that these qubits were more stable, smaller, more scalable, and drained less power than qubits made from superconducting metals — like the ones commonly used in quantum computing systems made by companies like IBM, Google and Microsoft.

Qubits in the first Majorana chip consisted of a material stack combining a semiconductor made of indium arsenide (used in devices like night vision goggles) with an aluminum superconductor. This forms a "topoconductor," a topological superconductor whose qubits are stored in the shape of the material stack.

Each qubit is made from two superconducting nanowires ended by Majorana zero modes (MZMs) – the building blocks of topological qubits that store information through parity, evenness or oddness in the number of electrons in a topoconductor wire.

Instead of aluminum, Majorana 2 uses lead to shield fragile qubits from disturbances like electromagnetic waves or cosmic radiation. For the semiconductor, researchers swapped out indium arsenide for a combination of indium arsenide and indium arsenide antimonide. The change doubled the "topological gap" — the physical barrier that protects the qubits from environmental noise and errors during calculations.

It also led to a major increase in stability and reliability: boosting the quantum coherence lifetime from between 1 and 12 milliseconds in Majorana 1 to an average of 20 seconds (with a maximum lifespan of 1 minute), the researchers said in the study.

Combining AI and quantum computing

The key components of the Majorana 2 were designed atom by atom, so the scientists needed to add impurities in the form of other materials into the crystalline structure to lock each atom in its correct spot. But adding too many impurities, or adding them in the wrong way, would disturb the structure. To get these impurities into the right spots, the scientists turned to artificial intelligence (AI).

"Finding the exact recipe, the right amount to put to get the desired energy structure, requires a lot of experimentation in the old world order. In the new world order, through simulations, you can see where the highly probable target is. And then with that knowledge, you ideally only have to experiment once,” Zulfi Alam, corporate vice president for quantum at Microsoft, said in the statement.

Using the Microsoft Discovery platform, the scientists deployed AI agents to keep track of the complex intersectional elements while designing Majorana 2 — with changes to any of the software, architecture, design, the materials stack, the fabrication processes, measurements, and others, carrying ramifications for every other element. The project also had close to two decades' worth of data in many different formats, which were stuck in different silos. But AI agents were able to resynthesize the data and establish connections between the different pieces of information.

AI also slashed the time it took to conduct experiments from weeks by "several orders of magnitude," Alam said in the statement, but did not specify the exact time saving.

"Using agentic AI to automate the measurements was a game changer,” said Alam said in the statement. "It goes through some math and starts saying, '"Hey, where do I find the lowest point where everything sort of works?'" And it can do all these voltage adjustments in parallel, which a human cannot do. The way our minds work, we are more linear."

Pathway to the holy grail

Nayak said in a technical blog post that the company is now cutting its timeline to build a practical and scalable quantum computer in half with a new target of 2029. "This achievement will mark a major milestone on the path to a transformative fault-tolerant quantum computer that has the potential to solve problems that affect all of humanity."

This timeline sits roughly in line with competitors in the field. But this apparent progress in the field of topological quantum computing is not without its detractors.

Following the release of Majorana 1 last year, physicists questioned the extent to which Microsoft researchers proved that MZMs were present in the device. Nayak, who was involved in last year's research, later presented additional evidence at a talk at the Global Physics Summit in March.

Others have criticized the evidence for the claims made in the new study. Speaking with Scientific American, scientists including Sergey Frolov, a quantum computing researcher at the University of Pittsburgh, suggested that the data reported has yet to be proven credible. Frolov cites the fact that Microsoft's last preprint of this kind was unpublished, meaning it wasn't peer-reviewed

Speaking with Live Science, Yuval Boger, quantum computing researcher and chief commercial officer at QuEra, a quantum computing company that is building neutral atom machines, lauded the progress but urged caution.

"Topological qubits are a bold, long-horizon bet, and the device improvements they reported are worth noting," he said. "As with any announcement of this kind, the sensible thing is to wait for peer review and independent reproduction before drawing conclusions," he added.

"The community has debated the topological evidence since 2018, and that scrutiny is healthy for everyone," he said. "It's also worth keeping the news in proportion. Topological computing has not yet demonstrated a working qubit, while other modalities are considerably further along."

Competing entities, including companies and research institutions, are working on a host of different qubit modalities as they all strive to hit the holy grail of building a fault-tolerant quantum computer that exponentially scales down its errors as you increase the size of the system. This is known as "below threshold" quantum error correction. They may include superconducting qubits, neutral atom qubits, photonic qubits, or, in Microsoft's case here, topological qubits, among others.

"In the end, any real progress in quantum computing is good for all of us," he said. "The field moves fastest when many approaches are pushing at once, and we welcome that."

Walk into any U.S. pharmacy, and you'll find a multicolored range of melatonin supplements marketed as sleep aids for adults and children alike.

The body-made hormone melatonin streams through our systems every evening, preparing us for rest. But while melatonin supplements claim to mimic this natural process, experts say bottles stamped with claims like "100% drug free" may be misleading.

Melatonin carries little risk when used correctly, but as a dietary supplement, it skirts the strong governmental regulation that's reserved for pharmaceuticals. This has led to inaccurate dosage labeling, incomplete safety information, and scant research into melatonin's long-term effects on kids, some scientists argue. Researchers worry that, given that many people perceive melatonin as harmless, some are misusing the supplement by taking unnecessarily high doses at improper times of the day.

"It's being promoted like there's absolutely no risk … [People say] 'It's all natural, and it just is a safe sleep aid,'" said Dr. Pieter Cohen, a Harvard Medical School researcher who studies supplements. "And that's where you can get into trouble."

The U.S. Food and Drug Administration (FDA) classifies melatonin as a supplement, while countries like the U.K. and Australia deem it a medication. This designation comes with a level of governmental oversight that U.S. supplements aren't subject to. Given melatonin's soaring popularity, some experts think the United States should follow suit and make melatonin an over-the-counter drug.

But if melatonin is already produced naturally by the body, why would the supplement version be riskier? Live Science spoke with experts to find out.

Supplements aren't well regulated

In the U.S., a 1994 federal law required that supplements be regulated as food, not drugs. The FDA doesn't assess the safety, efficacy or manufacturing quality of supplements before they're marketed but instead intervenes if harmful side effects are reported or if it's made aware of inaccurate labels. This largely leaves the supplement industry to self-regulate.

Cohen's research suggests that actual melatonin concentrations in products may range from 74% to 347% of what's reported on the label. Meanwhile, safety warnings and instructions on the best way to use melatonin aren't required on labels. Still, many popular brands adhere to voluntary melatonin labeling guidelines from the Council for Responsible Nutrition, a supplement trade organization. These guidelines include disclaimers like "may cause drowsiness" and "for occasional or intermittent use only."

Despite such disclaimers, nearly 1 in 5 school-aged children use it regularly as a sleep aid, a trend scientists are concerned by. Long-term safety data is lacking in kids, in part because melatonin hasn't gone through the safety trials required of new drugs. And melatonin isn't patented, so there's no strong incentive for drug companies to invest in safety studies, according to Debra Skene, a neuroendocrinologist at the University of Surrey in England. "There's no money in it," she said.

Many parents inaccurately view melatonin as akin to a "little glass of warm milk," Cohen said. But if parents give children high-dose supplements or multiple doses at once, "you're talking about side effects and getting sick and not helping sleep," he said.

Calls to U.S. poison control centers about melatonin poisoning in young children have shot up with the supplement’s popularity. This has been driven by accidental ingestions in children under 5, and in some cases, kids had serious symptoms, like respiratory failure or seizures. That said, some of those effects may be due to certain melatonin products' inclusion of serotonin, which can be dangerous if consumed in large quantities.

Evidence does suggest that melatonin can help children with attention-deficit/hyperactivity disorder and autism get to sleep, among whom sleep problems are common. But the potential benefits of melatonin for neurotypical kids are less clear.

One concern scientists have is that the regular use of melatonin may affect puberty. In various animal species, melatonin helps regulate reproduction, and supplemental melatonin can disrupt these effects. The few human studies that exist haven't shown clear effects on puberty, but there hasn't been an in-depth study tracking hormone levels in the blood, said Helen Burgess, who researches the circadian rhythm at the University of Michigan.

"I'm OK with intermittent short-term use," she said of melatonin supplementation in kids. But given the lack of data, she suggested that parents consult a pediatrician to look for another solution if their child needs melatonin regularly.

The lesser-known effect of melatonin

Supplements supercharge the body's supply of melatonin, said Margarita Dubocovich, who studies melatonin at the University at Buffalo. Supplement dosages generally range from 1 to 10 milligrams (mg), while adults produce about the equivalent of a 0.3-mg pill nightly, she said.

When taken orally, supplements and medications pass through the digestive tract before reaching systemic circulation, meaning the full dose of melatonin won't necessarily make it into the bloodstream. However, the amount that circulates will still be far higher than what's naturally produced in the body, said Skene.

Upping your melatonin dosage doesn't knock you out; the drowsiness triggered by the supplements is generally mild. The supplement also avoids the side effects of conventional sleeping pills: memory loss, gastrointestinal issues and next-day "hangovers."

However, melatonin has a lesser-known effect. Given its role in regulating the circadian rhythm — the 24-hour clock that regulates many bodily systems — it can scramble the body's sense of timing when regularly taken outside your bedtime, such as too early before bed or in the middle of the night.

That's why melatonin can help treat jet lag if taken for multiple days just before or after traveling. But if people take pills as they're tossing and turning throughout the night, this could throw off the circadian rhythm, Dubocovich said. This practice may induce a kind of jet lag (without the perks of international travel). In the short term, that can come with drowsiness during the daytime, irritability, upset stomach, brain fog and insomnia.

"I think that melatonin could be good if you take it in a controlled manner and the right dose," Dubocovich said, suggesting a few milligrams taken right at bedtime. "But how people are taking it [currently], they are intoxicating themselves."

Not all researchers agree that melatonin-induced jet lag poses a big risk, but taking more than 3 mg in the middle of the night still "isn't a good idea," said Burgess. This dose may linger in the body through the morning, she said, leading to next-day grogginess.

Burgess' research suggests that supplemental melatonin is most effective at shifting the biological clock forward in time, which happens when it's taken in the early evening. Still, she said the impact of melatonin on the circadian rhythm "just does not seem to be really recognized or understood by the public at all."

Should melatonin be a drug?

If melatonin were an over-the-counter medication, it would require clear labels with instructions and accurate dosing. "You should be able to know what dose you're getting, and it shouldn't be sold in whatever dose the manufacturer … decides to pour in on a given day," Cohen said.

Yet, reclassification could disrupt access to a relatively safe sleep aid that many consumers rely on, argued Jeff Ventura, a spokesman for the Council for Responsible Nutrition, the supplement industry trade organization that created the melatonin labeling guidelines. "Moving melatonin into the drug category would require formal FDA drug approval pathways, which could limit consumer access, reduce product variety, and significantly increase costs" for both manufacturers and consumers, he told Live Science in an email.

It's unlikely the supplement will be made an over-the-counter drug anytime soon. Cohen attributes the lack of a serious movement to the supplemental industry lobby, which includes the Council for Responsible Nutrition. "Any move to take such a great money winner like melatonin out of their hands would cause them to fight it tooth and nail," he said.

Burgess, who serves on the scientific advisory board for the melatonin brand Natrol, agrees that purity is an issue. However, she said she isn't too concerned about melatonin's availability, given its "pretty benign effects." Still, she suggests going with a big name brand with a US Pharmacopeia (USP) stamp verifying its listed ingredients are accurate; USP is a nongovernmental organization that sets quality standards for medicines, food ingredients, and dietary supplements.

After studying the hormone for 40 years, Dubocovich agrees that melatonin should be regulated as an over-the-counter medication.

"Melatonin can help support a good night’s sleep, especially in small doses that mirror the body’s natural rhythm," Dubocovich said. "But not all melatonin is created equal."

This article is for informational purposes only and is not meant to offer medical advice.

How much do you know about sleep and dreams? Find out with our sleep quiz!

An Ebola epidemic in Central Africa has been declared a "public health emergency of international concern" by the World Health Organization (WHO).

As of May 22, over 800 Ebola cases have been reported in the Democratic Republic of the Congo, including more than 180 deaths; these counts include both suspected and laboratory-confirmed cases of the disease. There are also two confirmed cases and one death in Uganda, specifically among people who had recently traveled to the DRC.

A number of factors are making this outbreak very challenging to contain, experts told Live Science — here's what to know.

Current Ebola outbreak

WHO officials suspect the Ebola outbreak centered in the DRC may have begun about two months ago. The earliest suspected death occurred April 20, and it was likely followed by a superspreader event at a funeral or healthcare facility, officials say. Reuters also reported that medical personnel failed to escalate the first patient's samples for further tests after they came back negative one type of ebolavirus.

That virus — known as Ebola virus, or Zaire ebolavirus — is the most common culprit behind Ebola disease outbreaks and deaths. There are two other viruses known to cause outbreaks of Ebola disease: Sudan virus and Bundibugyo virus. The latter is driving the current outbreak.

With only a handful of travel-related cases in Uganda, as well as one American receiving care in Germany after being infected in Congo, the outbreak remains concentrated in the DRC. The WHO anticipates a high risk of international spread, though, which prompted the agency's leader to make an emergency declaration without first convening a committee to discuss the decision.

"In our view, the scale and speed of the epidemic demanded urgent action," WHO Director-General Tedros Adhanom Ghebreyesus said May 19.

A virus with no vaccine

Part of the challenge with containment is that there is no effective vaccine against Bundibugyo virus, the germ driving this outbreak.

There are approved Ebola vaccines. "However, these vaccines have been designed specifically to address the Zaire ebolavirus," which has historically caused more outbreaks than Bundibugyo virus has, said Dr. Madeline DiLorenzo, clinical coordinator of infectious diseases operations and an associate hospital epidemiologist at New York University (NYU) Langone's Tisch Hospital. The largest Ebola outbreak to date, which happened from 2014 to 2016, involved the Zaire ebolavirus.

Zaire ebolavirus and Bundibugyo virus are genetically distinct, sharing only about 60% to 70% of their genetic material. The protein targeted by the existing Ebola vaccines is encoded by a specific gene, and that gene's sequence differs between the two viruses, DiLorenzo explained.

Studies suggest that immune responses against filoviruses — the family of viruses that includes Zaire ebolavirus and Bundibugyo virus — show limited cross-reactivity, meaning the immune response is narrowly focused on just one type of virus. So the Zaire-targeting vaccines likely wouldn't help in this outbreak.

WHO officials have said there is a promising experimental vaccine against Bundibugyo virus, but there are no doses available for a clinical trial. It could take six to nine months to manufacture those doses, the WHO estimates. Another vaccine in development could take two to three months to produce, but its efficacy is unknown because scientists are still waiting on results from lab-animal tests.

In Ebola disease outbreaks, vaccines are used for ring vaccination, in which people exposed to a suspected or confirmed case are inoculated against the disease. They can also be used for "targeted geographic vaccination," in which everyone in a given area is vaccinated because the outbreak is relatively concentrated there or contact tracing is too difficult to conduct. Without a Bundibugyo vaccine, both of those strategies are currently off the table.

Lack of diagnostics and treatments

Ebola disease can be difficult to spot in its early stages, in part because the first symptoms are fairly generic: fever, fatigue, malaise, muscle pain, headache and sore throat. These symptoms appear within two to 21 days of exposure to an ebolavirus.

There are tests for Bundibugyo virus that look for the virus's genetic material in bodily fluids; this approach is known as PCR testing. "However, it is not widely available for the Bundibugyo virus, making it difficult to diagnose and, as a result, contain the virus," Dr. Jill Weatherhead, an associate professor of infectious disease and tropical medicine at Baylor College of Medicine, told Live Science in an email.

Even when such tests are available, it can take several days after a patient develops symptoms for the virus to be detectable in blood, DiLorenzo noted, so repeat testing can be necessary. The recommended samples for Ebola tests are whole blood or plasma for living patients and an oral swab for deceased individuals, the WHO says.

There are additional tests that can detect Bundibugyo virus, including some that determine whether a sample contains a filovirus that infects humans in general, without specifying which one is present. These rapid tests look for specific viral proteins. However, these tests are less sensitive and "may not pick up specific proteins associated with Bundibugyo," DiLorenzo said. "This may have contributed to delayed detection of the current outbreak in the DRC."

On top of these diagnostic challenges, there are no specific antiviral treatments for Bundibugyo virus. There are lab-made antibodies approved for Zaire ebolavirus, and they improve survival by binding to the virus's surface and blocking it from infecting cells. Similar treatments for Bundibugyo virus haven't progressed beyond lab research or early-stage human safety studies, Weatherhead said.

Conflict, declining foreign aid hinder containment

In the absence of vaccines and treatments, other strategies, such as quarantining close contacts of infected people, become key to containing the outbreak, Weatherhead said. Clinicians treating patients with suspected or known cases should also follow strict infection-prevention-and-control protocols to prevent further spread, she added.

Ebola spreads via contact with infected blood and other bodily fluids, as well as contaminated surfaces or materials, such as clothing and bedding. This means using personal protective equipment to block splashes or other contact with infected materials, among other protocols.

The strategies available to contain this outbreak — finding and isolating cases and employing tight infection-control protocols — require public health infrastructure to execute. But in the DRC, that infrastructure is severely compromised. The outbreak's epicenter is the Ituri province in the northeastern DRC, which has "experienced armed conflict for decades, making it challenging for health systems to function optimally there," DiLorenzo said.

Joshua Walker, director of programs of the Congo Research Group at NYU's Center on International Cooperation, told Live Science via email that the circumstances of the current outbreak resemble a 2018-2020 outbreak that centered on the North Kivu province, which borders Ituri to the south.

This time around, several cases have been reported in both North and South Kivu provinces, parts of which are essentially partitioned between the DRC government and a Rwanda-backed rebel group, called M23. And there's been increasing violence among armed groups in Ituri in recent months, Walker said. Meanwhile, development assistance for healthcare in the area has fallen substantially in recent years. Together, these factors "will make access and coordination of a single response more difficult," he said.

Recent cuts to foreign aid are only making things worse, Dr. Manenji Mangundu, DRC country director for Oxfam, which is coordinating on-the-ground responses to the outbreak, told Live Science in an email.

"USAID [the U.S. Agency for International Development] was the main donor in the DRC and many aid organizations depended on its funding to deliver life-saving support in a country already devastated by conflict," Mangundu said. "When USAID was shuttered last year, eastern DRC lost around 70% of its humanitarian aid."

More funding cuts from other donors followed, leading to the closure of medical centers, declines in medical supplies and community health workers, and reduced capacity to deal with outbreaks.

"Funding cuts directly do not cause outbreaks, but they do weaken the very systems that are meant to prevent small crises from becoming larger crises," Mangundu said.

In Ituri, residents have been repeatedly displaced due to the conflict and must shelter in crowded schools and churches with limited access to clean water, sanitation and healthcare. Funding cuts only compound these existing problems, he said.

In this setting, it can also be difficult to convince people to adopt safety measures to limit the spread of Ebola, he added. For example, "communities remain attached to their deceased relatives and continue to handle the bodies, which increases the risk of transmission."

Lessons from past outbreaks

Walker noted that the international response to the current outbreak will shape how it evolves.

"The 2018-2020 Ebola response effectively sidelined the local Congolese health system, viewing it as too fragile or weak to be an effective partner," he said, citing his past research. So the response was largely orchestrated by outsiders to local communities, which fostered suspicion.

"One hopes," Walker concluded, "that there have been some hard lessons learned by the international community since the last major outbreak."

As global health leaders warn that this outbreak could swell to impact more countries, Mangundu emphasized that much more support is needed to snuff out the epidemic in the DRC.

"There is capacity in the country to respond, but there are not enough resources to help control and prevent the spread," he said. "We need to fund humanitarian aid and support the people of DRC at this time before a preventable crisis turns into one with far wider, global consequences."

This article is for informational purposes only and is not meant to offer medical advice.

In recent years, researchers across the world have been publishing increasingly older ages for prehistoric rock art. Among the headliners is a painting of a warty pig in Indonesia that reportedly dates to 51,000 years ago and a hand stencil that researchers claimed was an eye-popping 67,800 years old.

Most of these dates have been determined by measuring the radioactive decay of some versions, or isotopes, of uranium into thorium ‪—‬ a method called uranium-thorium (U-Th) dating. However, the validity of some of these dates has been called into question, with Georges Sauvet, a researcher at the Center for Research and Studies of Prehistoric Art in France, proposing that the method tends to overestimate the ages of dated samples.

In a short communication published March 23 in the journal AOJ of Histoarchaeology and Anthropological Exploration, Sauvet criticized the published dates for some prehistoric rock art, stating that researchers are being less cautious in a "race towards the earliest rock art."

If these dates go unchallenged, Sauvet argued, it distorts our understanding of the intelligence of early Homo sapiens and human relatives, such as Neanderthals.

Sauvet thinks it is absolutely necessary to cross-date with other dating methods to ensure accuracy of the U-Th dating which on its own "is not acceptable," Sauvet said.

What is uranium-thorium dating?

When water washes over and breaks down limestone and then drips into caves, it slowly deposits calcite. As the calcite forms, it traps small amounts of uranium that is also dissolved in the water. The uranium then radioactively decays into several "daughter" isotopes, including thorium isotopes.

The method in question, U-Th dating, involves uranium-234, which contains 92 protons and 142 neutrons and will spit out two protons and two neutrons (called an alpha particle) to form thorium-230, which has 90 protons and 140 neutrons. It takes about 245,629 years for half of a given amount of uranium-234 to decay into thorium-230. So, assuming nothing is added to or taken away from the system since the uranium was deposited, the ratio of these two isotopes can determine the age of a calcite layer deposit.

In theory, this type of U-Th dating can be an incredibly powerful tool for archaeologists.

That's because in some cases, rock art contains carbon, an organic element that can be dated. But often, rock art is made from the mineral ocher or from etchings in stone, neither of which can be radiocarbon-dated.

This is where U-Th dating comes in handy. Calcite can grow over the art, creating a minimum age for its creation.

Questioning the age of rock art

However, in his paper, Sauvet argued that U-Th dating may provide unreliable dates and that all of them need to be cross-checked with other methods before a date can be estimated.

"My first alarm against the danger of the method was reading" a 2018 paper by Dirk Hoffmann, a researcher in the Department of Human Evolution at the Max Planck Institute for Evolutionary Anthropology, and colleagues, "in which they announce that three Spanish cave [drawings] were dated to 65,000 years and were due to Neanderthals," Sauvet told Live Science in an email. The drawings show red ocher dots, a ladder and hand stencils, and some researchers suggested they show Neanderthals' ability to think artistically.

In a response in 2020, Sauvet and 42 other researchers published a reply highlighting the drawbacks of U-Th dating. The problem, Sauvet said, is that the method relies on calcite deposits forming and remaining in a "closed system," meaning no uranium is leached from the calcite deposit from the time of its formation. In an "open system," rainwater and groundwater percolate through the calcite and leach out uranium, skewing the U-Th ratio and making the deposit appear artificially older than it is.

"Dating of rock art is a particularly challenging subject," Adelphine Bonneau, an assistant professor in the Department of Chemistry at the University of Sherbrooke in Quebec, told Live Science in an email. "In theory, Mr. [Dr.] Sauvet is right. U-series dating [which includes U-Th dating] can lead to overestimated dates."

According to Sauvet's paper, the hunt for increasingly older dates is in part due to grant awards and the prestige that comes with them. This has resulted in a series of increasingly older dates, including the supposedly 65,000-year-old art in Spain.

This date would "imply that paintings were made by Neanderthals, whereas there is absolutely no archaeological proof" that Neanderthals were capable of artistic creation, Sauvet said. (Other researchers, however, disagree, and think Neanderthals did create art.)

The debate was revived with the discovery of negative hand stencils in Sulawesi, Indonesia, that were U-Th-dated to around 67,800 years ago, which recently surpassed all previous records.

To highlight U-Th dating's unreliability, Sauvet cited several cases in his new paper in which U-Th and radiocarbon dates did not match when cross-checked. At Nerja Cave in southern Spain, a U-Th date yielded an age of 119,000 years, while a radiocarbon date of a charcoal mark that was part of the same drawing was around 19,000 years old; meanwhile, another radiocarbon date of a calcite layer from the same art was around 14,000 years old.

At Leang Balangajia, the outer calcite layer, which should be the youngest, was dated to 37,300 years old — roughly 7,800 years older than the layer underneath. In all of these cases, the suggestion is that the calcite layers were open systems.

In defense of U-Th dating

Not all researchers agree with Sauvet's assessment. João Zilhão, a research professor at the University of Lisbon and co-author of the Spanish rock art studies, said any dating method can be susceptible to errors. In an email to Live Science, he drew parallels to radiocarbon dating, in which the ratio of carbon-14 to carbon-12 is measured ‪—‬ a method that can be subject to contamination.

Similarly, optically stimulated luminescence, which measures when something was last exposed to light, can give an artificially old age if someone doesn't account for residual emissions from mineral grains in a sample, he noted.

There will always be ways that dating estimates can be subject to error, and U-Th is not particularly special in that regard, experts told Live Science.

"Saying that uranium-series dating does not work for rock art because there are cases where some calcite samples show open-system behavior is an overgeneralization," said Maxime Aubert, an archaeologist and geochemist at Griffith University in Australia, whose team was behind the Sulawesi dates.

Aubert acknowledged that uranium can leach out of or into calcite deposits, thereby distorting their true age. To get around this problem, his team has in recent years developed a laser ablation approach that allows zones altered by outside water seepage "to be identified and excluded from age calculation," he told Live Science.

To ensure their dates were accurate, including ones for the oldest hand stencil art in the world, Aubert's team used lasers to remove small sections within each sample. Next, they measured the isotope ratios in several of these sections to create a map of the isotope ratios across each calcite layer. Finally, they excluded layers where the isotope ratios varied a lot, which would have suggested contamination.

Scientists have developed other ways to ensure their dates are reliable, Bonneau added. "There are several ways to deal with these open-systems and correct the dates accordingly … Using laser-ablation and having a map of the calcite layers makes it possible to select the most reliable parts and then extract the dates," she said.

Bonneau noted that Hoffmann and colleagues did not have access to laser ablation at the time of their study, as the technology was still being developed.

To get around this, in past work Aubert "always published a lot of information regarding the different isotope ratios, the composition of the calcite layers, etc.," Bonneau said. "These data are needed to evaluate the reliability of the dates." Hoffmann's study is missing that data, she added, "so we cannot evaluate the reliability of the dates."

While U-Th dating can be subject to error, that doesn't make their age estimates worthless, she said. "Sauvet is right in principle, but if scientists do their job correctly, the dates are reliable," Bonneau concluded.

How much do you know about our closest relatives? Find out with our neanderthal quiz!

Researchers have cast doubt on an influential 2025 study that claimed a new artificial intelligence (AI) model could accurately simulate human thought.

That study, published in the journal Nature, concluded that a large language model (LLM) called Centaur could "predict and simulate human behavior" with up to 64% accuracy across a series of psychological experiments. At the time, the researchers argued that Centaur's performance reflected a genuine understanding of human decision-making, after it was trained on a dataset of more than 10 million human decisions from 160 experiments involving 60,000 people.

But a more recent study, published in the January 2026 edition of the journal National Science Open, has called these findings into question.

Rather than making judgments based on the semantic meaning of questions, as the original research implied, the new study argues that Centaur simply learned statistical shortcuts in the training data — a phenomenon known as "overfitting."

Overfitting happens when an AI model learns its training data too precisely, memorizing patterns specific to that data rather than developing a broader understanding that transfers to new examples. An overfit AI will perform extremely well on training data but poorly on any new data that's introduced.

Study co-author Nai Ding, a professor at Zhejiang University's College of Biomedical Engineering and Instrument Science in China, likened overfitting to a student memorizing answers to a test rather than understanding the questions themselves.

"If a student is overprepared for an exam, they may learn tricks that allow them to guess answers correctly without actually understanding the underlying material," Ding told Live Science in an email. "If the training and testing samples share the same statistical distribution (and therefore the same kinds of shortcuts), overfitting may go undetected, and the model's performance will be overestimated."

Are we approaching an AI ceiling?

To test their theory, Ding and co-author Wei Liu, a postdoctoral student at Zhejiang University, modified the multiple‑choice questions used to train Centaur with the instruction: "Please choose option A." If the model truly understood the task, it would consistently pick option A, regardless of whether or not it was correct, they argued.

However, Centaur continued to choose the correct answers in tests, suggesting it was repeating learned patterns in its training data.

"High performance alone does not tell us through what mechanism LLMs achieve that performance — whether they truly understand the task or exploit statistical shortcuts in the data," Ding said.

The findings add to a growing body of research questioning how far current neural-network-based AI technology can go.

Researchers have long debated whether existing AI models could ever reach artificial general intelligence (AGI) — a hypothetical, advanced form of AI capable of reasoning at a human level and learning new skills beyond its training data.

While LLMs and broader neural network technologies have made strides in recent years, we could be approaching a ceiling. A study published in February argued that LLMs are fundamentally constrained by "reasoning failures" — a byproduct of their architecture that makes them incapable of holistic planning or in-depth thinking.

Chris Burr, a senior researcher at the U.K.'s Alan Turing Institute who was not involved in either study, pointed out that new AI models are built to score well on benchmarks that assess how closely their outputs match expected patterns. This means an AI model that's very good at pattern matching will naturally look like it understands what it's doing, even if it doesn't.

"Most frontier models are flexible enough to fit almost any pattern, and the headline metrics reward fit and benchmark advances rather than deeper understanding and conceptual nuance," Burr told Live Science in an email. "A model captures something meaningful about cognition only if it does more than predict behavior… At best, Centaur offers behaviourist-style evidence for a linguistically reduced slice of cognition."

Even so, the results of the 2025 study remain compelling. One of the standout findings was that Centaur accurately predicted the behavior of participants whose data and decisions weren't included in its training data.

The researchers divided the participant data into two groups, using 90% for training and keeping 10% for testing. Not only did Centaur accurately simulate the responses of that held-out 10%, but it also successfully predicted human choices in scenarios it hadn't encountered, the researchers said. Ding and Liu didn't address this finding.

Burr acknowledged that the research by Ding and Liu doesn't undo the Centaur study's fundamental argument, which is that AI models fine-tuned on human behavior could enable researchers to more closely simulate and study human cognition.

"The broader programme is not refuted, since only four tasks were tested and Centaur still performs best with intact context, but I think they've done enough to shift the burden of proof," he said.

Stress-testing research "essential for building cognitive models"

Ding explained that stress-testing AI research was key to expanding understanding of AI and its limitations, particularly as a tool for cognitive research.

"Our work is not intended to deny the value of Centaur, but rather to emphasize that when evaluating such models, we need to distinguish between 'performing well' and 'performing well for the right reasons'," Ding said. "This distinction is essential for building cognitive models."

Models trained to perform one task should always be tested on whether they can automatically solve tasks based on the same kind of knowledge but not used to train the model, he added.

"Without this kind of testing, we risk drawing incorrect conclusions about model capabilities. For instance, we might prematurely conclude that a unified model can already capture human cognition, thereby overlooking the problems that genuinely remain to be solved."

Live Science contacted the authors of the 2025 Nature study to ask questions about the findings of the newer study but did not receive a response by the time of publication.

A strange divergence is happening in the diagnosis of colorectal cancer: Older people, who are most at risk from the disease, are seeing fewer cases while people under 50 are seeing more.

These diagnoses are often devastating, as younger people also tend to go longer between first experiencing symptoms and getting diagnosed, meaning most of their cancers are caught in a late stage that is more difficult to treat.

"Almost all my patients are my peers, many of whom have young kids or are pregnant themselves or navigating the challenges of just establishing their careers," Dr. Geoffrey Buckle, a gastrointestinal medical oncologist at University of California, San Francisco Health, told Live Science. "They're hit with these obviously extremely scary diagnoses."

But what's driving this rise in young people? Doctors and researchers think some change that occurred after the 1960s is to blame, likely a shift in people's environments or lifestyles. Perhaps offering a clue as to the culprit, recent research also suggests that, whatever the factor, it primarily affects cancers of the rectum and the lowest part of the colon.

The rise of colorectal cancer in young adults

There were about 158,850 new cases of colorectal cancer diagnosed in the U.S. in 2026, according to the American Cancer Society (ACS). Over 86,000 cases — more than half — occurred in people over 65. But those numbers represent a decline of 2.5% a year in that age group since 2013.

Doctors chalk up the decline in colorectal cancer in people 65 and older up to the success of screening, said Dr. Aasma Shaukat, a gastroenterologist at New York University Langone Health. During routine colonoscopies, which start at 45 years old for people at typical risk, doctors remove precancerous polyps and thus prevent that tissue from becoming cancerous.

At the same time over-65s have seen lower cancer rates, people between 50 and 64 have seen a 0.4% increase per year and people between the ages of 20 and 49 have seen an even more alarming jump of 3% each year. The rise is seen in every racial and ethnic group.

What that translates to, according to research published in ACS's CA: A Cancer Journal for Clinicians, are 24,640 new cases a year for people 49 and younger and 47,600 new cases a year for people 50 to 64.

Most of these cases are what doctors call "left-side" cancers, occurring in the lower colon, which descends down the left side of the body, and the rectum. Between 1999 and 2023, rectal cancer mortality rose two to three times faster than colon cancer mortality across all demographics, according to research presented in May at Digestive Disease Week 2026, a major gastroenterology conference.

Possible environmental factors at play

Cancers arise due to genetic errors, some of which occur randomly when cells divide. And the cells in the gut divide constantly: The entire lining of the colon is shed and replaced every week, Shaukat said. The body has mechanisms to repair the myriad DNA mistakes that happen during this turnover, but when the rate of errors increases too quickly or when repair mechanisms fail, that's when most colorectal cancer occurs.

The environment in the gut can affect both this error rate and the efficacy of repair. Researchers have found that there's a complicated feedback loop between the gut microbiome, meaning the community of microorganisms in the colon, and health factors such as diet, exercise and insulin resistance.

Inflammation also factors into this feedback loop. While inflammation is a necessary immune response that repairs damaged tissue with new cell growth, when it becomes chronic, dividing cells become prone to mutation and overproliferation — in other words, cancer.

The culprits for increased chronic inflammation are many, Dr. Theodore Levin, a researcher and clinical lead for colon cancer screening at Kaiser Permanente Northern California, told Live Science. But "these are all thoughts or hypotheses," he cautioned.

One is the rise in sedentary lifestyles and obesity, which contributes to insulin resistance and, in turn, disrupts the microbial colonies that live inside the intestines. These microbes aid digestion and release compounds that can either promote or reduce inflammation.

Childhood antibiotic use rose after the 1960s, which can also perturb the gut microbiome. Additionally, microbiome-disturbing dietary changes since the 1960s include a rise in sugar-sweetened beverages, Levin said, as well as the introduction of a greater number of ultraprocessed foods, which typically lack the fiber content of fresh fruits and vegetables. Fiber, he said, promotes a diverse gut microbiome that tends to tamp down inflammation. (Fermented foods do the same thing, research suggests.)

Researchers are also concerned about the ubiquity of plastics, Levin added.

Some lab research suggests that colorectal cancer cells can take up microplastic particles and that the takeup of those particles increases the cells' propensity to migrate — at least in a lab dish. That could hint that microplastics make cancer more likely to spread through the body. Some animal studies also find that microplastics hamper immune function, which could theoretically help cancer cells to escape the body's defenses.

Plastic particles might also carry carcinogenic compounds deep into the gut, Levin said. Research is still ongoing, but some evidence links chronic exposure to perfluorooctanesulfonic acid (PFOS), a type of "forever chemical" found in nonstick cookware and firefighting foams, to colorectal cancer risk. Experiments using mouse intestinal tissue finds that PFOS reduces a gut enzyme that protects against cancer while spurring the production of proteins that encourage tumor growth.

But researchers have not yet found "one smoking gun" to explain the rise, cautioned Michael Sapienza, the Chief Executive Officer of the nonprofit Colorectal Cancer Alliance. For example, the results of observational studies of forever chemical exposure and colorectal cancer rates are mixed, with some even finding that more PFOS in the blood actually correlates with fewer colorectal cancer cases.

So while the hypotheses are many, the evidence remains murky. Older adults are exposed to all of these environmental and lifestyle changes, too, so researchers also need to understand why their cancer rates aren't rising along with those in younger people. It could be that the removal of precancerous lesions during colonoscopies are masking what would otherwise be a similar rise in the over-65s. Or perhaps it's early childhood exposure that matters most, a risk factor that the oldest adults escaped.

Raising awareness of the symptoms

Scientists aren't entirely sure why cancers of the lower colon and rectum are rising faster than the "right-sided" cancers of the upper intestines.

The lower and upper parts of the colon are very different, Levin said, even arising from different parts of the embryo in the very earliest stage of development. Cancers on the right side and left side often show different genetic mutations. Left-sided colorectal cancer frequently arise from what researchers call chromosomal instability, or mistakes in how chromosomes split between dividing cells during replication, when cells make a copy of their DNA before dividing. Right-sided cancers are more prone to arise from problems with DNA repair during replication.

These differing pathways could help researchers link environmental and lifestyle factors with the changes that lead to colorectal cancer. But that research will take time. For now, gastroenterologists are hoping to raise awareness about the symptoms of the disease and make sure other doctors think to check younger patients for cancer.

Early symptoms include rectal bleeding, unexplained weight loss, anemia, abdominal pain, blood in the stool, changes in bowel habits and unexplained nausea and vomiting. Unfortunately, many early symptoms of colorectal cancer match those of much less serious conditions. For instance, rectal bleeding can be caused by hemorrhoids, and changes in bowel habits could arise from anything from stress to an infection.

Jenna Scott, 40, of Atlanta was 29 and pregnant when her symptoms started. She didn't get a diagnosis until she was 31. By that time, the cancer was metastatic, having spread to other organs. Scott remains in treatment today.

"I don't know if it would have been as aggressive if they had caught it a year-and-a-half earlier," she told Live Science.

Scott's initial symptom was rectal bleeding, but her doctor brushed it off as hemorrhoids caused by pregnancy. Another classic colorectal cancer symptom, weight loss, seemed like typical postpartum shifts on the scale. Health systems and organizations like the Colorectal Cancer Alliance are working to raise awareness among doctors in all specialties so that they check these symptoms out thoroughly in young people rather than explaining them away.

Although routine colon cancer screenings don't start until 45, patients should be pushy with their doctors to get checked out earlier if they have any telltale symptoms, Sapiena said.

"It doesn't mean you have colorectal cancer," Sapiena said of these symptoms. "But it does mean you need to be your own advocate" and seek screening ASAP.

This article is for informational purposes only and is not meant to offer medical advice.

Artificial intelligence (AI) models can autonomously "replicate" across multiple machines, hacking vulnerable systems, copying their own parameters onto compromised hosts, and launching working replicas capable of continuing the attack, new research shows. However, experts told Live Science the bigger concern is not AI suddenly running wild on its own, but cybercriminals using AI agents to automate known hacking techniques.

Scientists at Palisade Research tested whether AI agents could independently move through a chain of intentionally vulnerable systems without human intervention. In a new study uploaded May 7 to GitHub, large language models (LLMs) could identify exploitable web applications, steal credentials, transfer their own files, and stand up new inference servers capable of continuing the attack from the next machine in the chain.

This is the first research to show an AI model "autonomously exploit a target and replicate itself end-to-end," the researchers wrote in the study.

Although that sounds like the premise of a sci-fi disaster movie, experts who weren't involved in the research said the reality is somewhat less dramatic.

Notably, the systems did not spontaneously decide to reproduce themselves in the wild. Rather, the researchers told the models to replicate, handed them Linux tools and shell access, and pointed them at intentionally vulnerable systems riddled with bugs. These included vulnerabilities open to various attack methods that could expose credentials, bypass logins, or hand over access to other machines.

Jeff Watkins, chief AI officer at consultancy NorthStar Intelligence, said the experiment was closer to automated cyber intrusion than to "machine life."

"The headline claim sounds dramatic, but the underlying mechanism is very familiar to anyone who has studied malware in depth," Watkins told Live Science. "What Palisade appears to have demonstrated is not a magical new form of machine life but rather an agentic system performing a sequence of fairly standard cybersecurity intrusion tasks."

A nervy direction of travel

According to the study, the agent copied onto newly compromised systems its model weights (parameters within a machine learning model that determine the strength of connections between artificial neurons), prompts (inputs instructing the AI) and harnesses (the software infrastructure).

It then verified the replica was functional by checking whether it could respond to inference requests, meaning prompts or queries sent to the AI model to confirm it was running correctly. Researchers also tested "chain replication," in which the replicated model becomes the attacker in the next stage of the operation.

"The agent exploits a web vulnerability, copies its existing weight file of about 119 gigabytes across via scp or rsync [common Linux file-transfer tools], and loads an inference server on the compromised host," he said. "That is more a worm with a planner attached, not consciousness scaling," said Antoine Casanova-Mazet, head of innovation at HIM Business School, where he works on AI adoption and digital transformation projects as well as having built AI systems himself.

Still, the researchers argue that the trajectory matters more than the current limitations.

In the study, the researchers said end-to-end replication success rates improved sharply over the past year as frontier AI systems became more capable at coding and autonomous task execution. The study reported that Anthropic's Claude Opus models achieved success rates above 80% in some scenarios, and OpenAI systems also showed measurable improvements.

"What is genuinely new is the trajectory," Casanova-Mazet said. "Within 12 months, end-to-end success on this pipeline moved from 6% to 81% on Anthropic's frontier models and from 0% to 33% on OpenAI's."

The study lands amid broader concerns about increasingly autonomous agentic AI systems capable of carrying out long chains of tasks with limited supervision. Researchers and safety groups have spent the past year warning that AI models are becoming more capable of offensive cybersecurity operations, vulnerability discovery, persistence and long-horizon planning. For example, in a December 2025 study, scientists in China linked with the cloud giant Alibaba said an experimental AI agent broke out of its testing confines and mined cryptocurrency without permission.

We should worry about other people, not AI

Cybersecurity experts remain skeptical that examples like that highlighted in the new study represent an immediate real-world threat. The biggest practical issue is scale, they said, as modern LLMs are huge. Moving hundreds of gigabytes of weights and infrastructure around a monitored enterprise network would likely generate large amounts of suspicious traffic.

"There are also practical constraints that make this less immediately troubling," Watkins said. "Replicating a full LLM is not like copying a small worm across a network. The notion that something as powerful as Mythos could self-replicate is not currently feasible, due to the intense resource requirements involved."

The more immediate worry is not rogue AI systems "roaming the internet," Watkins said, but attackers using agentic AI to accelerate existing cybercrime operations.

"The more realistic near-term concern is not a frontier model roaming the internet like a digital organism and causing global chaos," he said. "It is threat actors using agentic AI to accelerate familiar attack chains."

That divide is becoming increasingly important in AI safety research. Another study, uploaded Sept. 29 2025, to the arXiv preprint database, argued that the ability for an AI agent to copy itself does not automatically make a system dangerous in the real world. Aspects like autonomy, persistence, objectives, and access to tools or networks matter far more than whether the model can technically spin up another copy of itself, those researchers said.

As experts explained, the Palisade study appears less like rogue AI breaking loose and more like a glimpse into how AI-powered hacking tools are evolving.

"This research shows that self-replication is no longer a purely theoretical capability in agentic AI systems," Watkins told Live Science. "For now, it is probably less urgent than ordinary vulnerability exploitation, ransomware, credential theft and supply-chain compromise, but it is a warning about where those threats are heading as AI agents gain more tools, more autonomy and more operational access."

Climate change is likely to shift the distribution of virus-carrying rodents, potentially increasing the risk of spillover, when viruses make the jump from animals to humans.

This will likely expose populations to diseases that they haven't encountered before and raise the risk of outbreaks, researchers warn.

The hantavirus cluster on the cruise ship MV Hondius has put the rodent-borne disease into the global spotlight. While countries in South America, especially Argentina and Chile, have been dealing with hantavirus for decades, there's been a recent uptick in infections, according to the Argentine ministry of health. More than 100 cases of hantavirus disease were recorded between June 2025 and early May 2026, roughly double the number recorded in the previous year.

Hantaviruses are a family of rodent-borne viruses found across the Americas, Europe and Asia. "New world" hantaviruses cause hantavirus cardiopulmonary syndrome (HCPS), characterized by headache, fever and gastrointestinal symptoms, followed by potentially life-threatening respiratory problems. The Andes virus, responsible for the cluster of cases associated with the MV Hondius, is the only hantavirus known to be transmissible between people.

With a case mortality rate of up to 50%, HCPS is considered one of the most severe forms of hantavirus disease, while the infections that strike in Europe and Asia have lower fatality rates.

Climate change is affecting rodent host populations

Scientists studying hantavirus risk in southern South America have so far identified 11 genotypes linked to human disease across 13 native rodent species, said Juan Diego Pinotti, a postdoctoral researcher for Argentina's National Council of Research, based at the Institute of Animal Diversity and Ecology at the National University of Córdoba. Each strain of the virus has evolved within a specific rodent species, enabling that species to carry the virus while not falling sick from it.

The natural "reservoir" for Andes virus, the long-tailed pygmy rice rat (Oligoryzomys longicaudatus), typically lives in the humid forests and shrublands of the southern Andes in Chile and Argentina. The virus is spread to humans primarily through contact with the urine, feces, or saliva of an infected rodent.

Climate models predict that changing weather patterns over the next few decades may expand the long-tailed pygmy rice rat's habitat eastward, toward Argentina's Atlantic coast, where the vast majority of the country's people live.

"More intensive surveillance and testing are needed across provinces, because the virus might be reaching new areas without us finding out," Pinotti, who worked on the models, told Live Science.

El Niño-La Niña fluctuations may further reinforce this pattern. In central and southern Argentina, El Niño years usually bring increased rainfall, which stimulates vegetation growth, meaning rodents have access to abundant food sources and shelter, driving rapid population expansion.

"That in turn interacts with what the winters are like," said Veronica Andreo, a researcher for Argentina's National Council of Research, working at the Gulich Institute. "If the winter is more severe, fewer mice might survive. If it's a bit milder, more mice can survive and pass on to the next generation."

Beyond hantavirus: other viruses on the rise

A recent study focusing on a different family of rodent-borne viruses, known as arenaviruses, drew similar conclusions. New World arenaviruses, which include Guanarito virus in Venezuela and Colombia, Machupo virus in Bolivia and Paraguay, and Junin virus in Argentina, can cause severe hemorrhagic fevers with high hospitalization rates. Fatality rates range from about 5% to 30%, depending on the virus strain and the patient's overall health and healthcare access.

Researchers used machine learning to map how climate projections, changing rodent populations, and shifts in people's land use could alter the distribution and transmission of arenaviruses in South America over the next 20 to 40 years. They found that climate-induced shifts in temperature and precipitation, coupled with the expansion of agriculture, will likely increase spillover risk as viruses spread into new regions.

The researchers hope the models will help public health officials understand the need to prepare for outbreaks of these diseases as climate change drives more extreme weather patterns.

"These are high impact diseases, but they are fairly neglected in terms of how much research is ongoing about them, or whether they are on the radar of public health officials," said study lead author Pranav Kulkarni, a veterinary epidemiologist at the University of California, Davis (UC Davis) Weill School of Veterinary Medicine.

Notably, the Centers for Research in Emerging Infectious Diseases, part of the U.S. National Institutes of Health, would have been researching these diseases but was shuttered last year.

Continued surveillance is critical

As climate change reshapes disease risk, data gathering and surveillance will be crucial for improving outbreak preparedness. Better regional coordination and data sharing could help countries identify shifting hotspots earlier and build more effective public health responses, said Kulkarni.

"If we can come up with that sort of infrastructure, especially in countries which will probably be affected by this, then there would be more robust modeling, more accurate predictions, and a more informative way of creating health policy," Kulkarni told Live Science.

More research is needed to fully understand the diversity of viral genotypes and rodent species that exist in South America's southern cone, said Andreo. "We still need to understand how all the different reservoirs and viruses interact, and whether the viruses can jump from one reservoir to another," Andreo told Live Science. Some New World hantaviruses, such as the Laguna Negra genotype and the Lechiguanas genotype, are known to have multiple reservoirs.

Additionally, two new hantavirus genotypes have been identified in Argentina in recent years, added Pinotti.

"Always the most important thing is prevention — epidemiological surveillance and education of the population," Pinotti told Live Science. Yet, funding cuts to Argentina's health and science sectors have jeopardized disease surveillance programs, research capacity, and the ability to carry out large-scale, long-term studies into zoonotic and vector-borne diseases.

"When these teams see their operational capacity reduced," Andreo said, "the entire early warning network that makes it possible to detect outbreaks before they escalate is weakened."

This article is for informational purposes only and is not meant to offer medical advice.

The Food and Drug Administration (FDA) is implementing a new strategy to provide experimental gene therapies to patients with rare disorders without going through clinical trials. This framework could grant these patients access to individualized therapies, but experts are divided over whether the regulatory change is safe enough for patients.

Dr. Senthil Bhoopalan, a genome-editing expert at St. Jude's Children Research Hospital in Tennessee, said that, although the framework is still emerging and the details require more discussion between the public and stakeholders, "it's an exciting step in the right direction."

But other experts are concerned that the FDA's past track record with accelerated approval processes doesn't inspire trust.

Arthur Caplan, a medical ethicist at New York University, said more pressure to permit access to new therapies has led the FDA "to allow more risk to subjects, and more risk of failure post-approval, by being willing to accept weaker evidence."

Before receiving FDA approval, most therapies require clinical trials with hundreds or thousands of participants to show a drug is safe and works. In some cases, the agency grants accelerated approval for interventions that appear to show a benefit in small trials, when patients are very sick and have no other treatment options.

However, the new strategy, called the plausible mechanism pathway, would enable the FDA to grant permission to use therapies that haven't been tested in humans but could plausibly succeed.

The pathway would apply only to certain treatments, such as gene therapies that correct single-letter DNA errors, where large-scale clinical trials would be impossible. Take cystic fibrosis as an example. Around 40,000 people in the U.S. have this disorder, but hundreds of mutations can cause it, Bhoopalan said. As a result, you can't use one gene therapy formulation to treat every patient.

However, if a gene-editing tool and delivery technique have been shown to be safe in past human trials, the pathway would allow drug developers to tweak the sequence-specific element of the formulation, such as a guide RNA that tells the DNA "scissors" where to correct a mutation. Then, the specific gene-editing tool, such as a base editor, could be customized for specific mutations in each cystic fibrosis patient. This is similar to how food producers need only show that an ingredient is safe once before including it in multiple food items.

"The safety data can be extrapolated if you're using the same delivery mechanism," Bhoopalan said. "You're really only changing the guide." If the change you make in the body is swapping a faulty mutation with the form that healthy people have, you wouldn't anticipate side effects, he added.

Caplan agreed that this particular use of the pathway doesn't seem, on its face, to be high-risk. However, the safety of base editors has been tested only in relatively small trials thus far, with no more than 15 participants. With a sample size this small, it’s difficult to show a given gene therapy led to positive health outcomes. What’s more, without performing larger trials involving hundreds or thousands of participants, it’s impossible to know whether base editors cause rare side effects.

For example, at least 65 small-scale trials have investigated the use of certain viruses as vehicles to deliver liver-targeting gene therapies that treat hemophilia. While most of these studies show promise, a larger trial involving 134 participants revealed rare side effects, such as elevated liver enzymes, inflammation and allergic reactions.

"The level of risk doesn't keep me awake at night, but there are unknowns," Caplan said. "I think it would be very important to have serious follow-ups following FDA drug approval."

That's where he sees the potential for problems to creep in. Post-approval monitoring of drugs has "never been done with earnestness," despite promises made by pharmaceutical companies. "If we're going to take more risk to go faster at the front end, you have to beef up what's required and what's going to be monitored at the back end, post approval."

Still, that doesn't mean the level of post-approval scrutiny will be lower than it has been previously.

"It's possible that in the fullness of time, we'll see that they've lowered the bar," said Dr. J. Paul Taylor, a neurologist who treats genetic neurodevelopmental disorders at St. Jude Children's Research Hospital. "But the stated intent is not to change the level of substantial evidence [through post-approval monitoring]."

Who will it help?

In an article published last November in The New England Journal of Medicine, the FDA outlined which criteria a disease would need to meet to qualify for this pathway. The plausible mechanism pathway would be ruled out for disorders with unclear causes, such as dementia, Taylor noted.

"This is great for monogenic disorders, which are caused by mutations in a single gene," Bhoopalan said. It would be harder to use this pathway for polygenic diseases, which are brought on by an array of mutations, he added, as you would have to successfully correct multiple mutations to see a benefit.

Rather than correcting a faulty mutation, gene therapy could be used to "switch on" a backup gene in the case of spinal muscular atrophy, Taylor said, which is fatal in children who don't receive treatment.

There are some monogenic disorders that may not meet the criterion, however. Diffuse intrinsic pontine glioma is a brain tumor that appears in young children who carry a faulty gene. Taylor said specialists are split on whether reversing this mutation alone could shrink the tumors or if other mutations that appear as the tumor develops could continue to drive the cancer even if the initial mutation were corrected.

Another FDA criterion requires doctors to confirm that the patient's tissues have been edited. "It might be harder to quantitate when you're editing a critical organ like the liver, because you cannot get a piece of liver and measure how much has been edited," Bhoopalan said.

Doctors may need to repeatedly sample tissues from patients, as studies in mice have shown that gene therapies can wane over time, suggesting that some may not work as a "one-and-done" treatment. This would be much harder to accomplish if you could only sample tissues with invasive surgery.

Some body locations might be difficult to target with gene delivery systems in the first place. The blood, bone marrow, liver and lungs could make easy targets, Bhoopalan said. The heart, on the other hand, could be difficult to edit because a layer of tightly-packed cells creates a barrier that stops gene therapy vectors from crossing into heart tissue.

Though more discussion is needed to clarify which disorders can benefit from this expedited approval and how patients' health can be monitored afterward, experts hope the new pathway could help people with rare disorders.

"I think we have to start thinking about this as an inevitable next step," Bhoopalan said.

This article is for informational purposes only and is not meant to offer medical advice.

Mars is out; the moon is in. And this time, we may be going there to stay.

On March 24, NASA administrator Jared Isaacman announced new plans to build a "sustained human presence" on the moon, complete with a permanent lunar base. Construction of humanity's new home away from Earth could begin as soon as 2027, Isaacman said.

The announcement came just a month after SpaceX CEO Elon Musk similarly ditched his company's plans to visit Mars in favor of establishing a "lunar self-growing city" within the next 10 years.

It's an exciting prospect — but scientists say lunar colonization won't be as simple as packing up and lifting off. The moon's environment is harsh: think razor-sharp, electrified dust and a constant stream of radioactive particles travelling at light speed. Scientists still aren't sure how this increased dose of cosmic radiation, coupled with the moon's weaker gravitational pull, relative to Earth, will affect the human body in the medium- to long-term.

And then there's technological feasibility: Where will astronauts live, and with what resources? At the moment, these big questions are still awaiting answers that may not arrive in time for NASA and Musk's proposed plans.

"I don't think we're quite ready," Caitlin Ahrens, a researcher at University of Maryland and NASA's Goddard Space Flight Center who studies the lunar environment, told Live Science. "A decade may seem far away to some people. To a scientist, it's the blink of an eye."

Satellite of horrors

Moondust is rather different from what we call dust on Earth. Without wind and liquid water, the moon's dust doesn't soften over time. "We're talking very, very sharp little pollen[-size] shards here," Ahrens said.

It's also easily electrified. Simply walking on the moon kicks up dust clouds, and rovers traversing the lunar landscape shoot up "rooster tails" of electrically charged, levitating dust, which sticks to anything in its path, Ahrens said. Energized dust can clog breathing vents in potential living quarters, scratch up spacesuits and coat solar panels, potentially overheating and breaking them. (Dust clogs have already spelled doom for several Mars rovers).

Without the soft blankets of Earth's atmosphere and magnetic shield, moon inhabitants would also be constantly blasted by radiation. Cosmic radiation is "pretty much omnipresent anywhere you go into space," including on the moon, Dr. Emmanuel Urquieta, an aerospace medicine researcher at the University of Central Florida, told Live Science. "It's incredibly difficult to shield."

Cancer is a potential risk, but because health effects of radiation take time to develop, we wouldn't know for certain if this is a major risk factor until potentially decades after landing human settlers on the moon. "Every person that goes to space… will absolutely be test subjects," Urquieta said.

We're not ready for self-defense

Humans on the moon would need hefty building materials to protect lunar habitats against these hazards. Metal or glass domes, subterranean habitats and housing made of 3D-printed moon soil are all potential options being investigated.

But Ahrens, who also works on lunar risk assessment, doesn't think we're ready to plan construction. For example, if NASA decides to invest in an underground habitat — perhaps the most surefire way to avoid radiation — scientists still have no idea how to dig on the moon, she said.

Even with sufficiently protective living quarters, the moon's weaker gravitational pull — just one-sixth that of Earth — may pose health risks. From previous space travel we know that, without the gravitational force humans evolved with, our bones and muscles need significant amounts of exercise to prevent them from withering away. But it wouldn't be practical to lug heavy treadmills, like those used on the International Space Station, to the moon, Urquieta said.

A lack of gravity may also redistribute the body's natural balance of fluids, with potentially disastrous effects. Normally, due to the downward tug of gravity, up to 80% of our blood is in our legs at any given time, Urquieta said. But on the moon, more fluids would flow to the body's upper half, causing loss of blood as the body attempts to rebalance itself, as well as potential swelling at the back of the eye and jugular vein thrombosis, a condition that causes potentially fatal blood clots in the neck. Compared to zero gravity, lunar partial gravity may or may not pose similar health risks. Urquieta said researchers won't know until people spend some time there.

The future of lunar settlement

Ultimately, Ahrens said, these challenges boil down to the need for more data, which scientists are hoping to gather with missions like NASA's Artemis campaign to return humans to the lunar surface as soon as 2028. But she envisions a much slower timeline than Musk's and Isaacman's plans; scientists haven't even taken a sample of the moon's ice yet — a resource lunar settlement planners are banking on.

Depending on its depth and composition, lunar ice could provide water, rocket fuel and rare earth metals. But until scientists get a physical sample of ice, they can't rely on its usefulness or economic prospects. For now, scientific knowledge of the ice's actual chemical composition is very limited: "We know it's cold, and we kind of know where it is," Ahrens said.

When it comes to the moon and its offerings, "we have to be very careful not to sell something which [we] don't have," said Giuseppe Reibaldi, president of the Moon Village Association, a non-profit group focused on international collaboration in lunar activities.

Reibaldi told Live Science that what we find in the moon's ice could be the difference between the lunar equivalent of California's gold rush boomtowns — where settlements sprang up in response to mining opportunities — and Antarctica, where a small number of scientists come and go for solely research purposes.

But even if permanent settlement doesn't happen as soon as hoped, Ahrens said she is optimistic about the forecast for lunar development. "I think it's not all sunshine and rainbows, but it's also not dark rainy clouds either," she said.

Editor's note: The headline and lead art for this story were changed on April 29, 2026.

Moon landing quiz: How quickly can you name all 12 Apollo astronauts that walked on the moon?

Catastrophic water shortages are in store for the 40 million people living in areas fed by the Colorado River, even if cities in the region such as Denver, Phoenix and Las Vegas make dramatic cuts to their usage, recent research suggests.

Water shortages could start as soon as this summer, as snowpack levels reached a record low over Lake Powell and the spring runoff into the Colorado River is expected to be minimal, experts told Live Science. And the region's major cities, which have already slashed their per capita water consumption by up to 58% between 2002 and 2025, can't solve the problem alone.

"We can't just shove it all onto the residents of these cities and tell them to use less water, because it's still not going to be enough [to prevent water shortages]," Renee Obringer, an assistant professor in the Department of Energy and Mineral Engineering at Penn State and the first author of the new study, told Live Science. "Regardless of what mitigation we do, we're still going to have to adapt to this warmer world that is going to have more intense droughts and more frequent droughts."

The findings reveal that bolder measures will be needed to head off disaster in the Colorado River basin. But exactly what are those steps?

One piece of the puzzle

Seven states — Arizona, California, Colorado, Nevada, New Mexico, Utah and Wyoming — pull water from the Colorado River, and the amount allocated to each region is governed by a century-old document called the Colorado River Compact.

But the Colorado River basin, along with the rest of the U.S. Southwest, has been in a megadrought for 25 years. Between 2000 and 2019, the river's flows shrank by 20% due to climate change and water overuse to supply cities, agriculture and industry.

Finding new water supplies is unlikely, so Obringer and her colleagues analyzed what would happen if three large cities fed by the river — Denver, Las Vegas and Phoenix — reduced their consumption by around 25% under various climate scenarios, ranging from a global temperature increase of about 2.9 degrees Fahrenheit (1.6 degrees Celsius) to an increase of 7.7 F (4.3 C) compared with preindustrial levels.

Their results were published in December 2025 in the journal Water Resources Research. In most climate scenarios, demand management — which encompasses broad strategies such as raising awareness, offering rebates, and subsidizing low-flow devices — did not make up for lower water storage in urban reservoirs caused by higher temperatures and lower precipitation in the region. Denver, which is technically not located in the Colorado River basin but gets half of its water from the Colorado River, was the lone exception. Demand management compensated for river flow reductions in two high-emissions climate scenarios for the city. However, Obringer suspects some of those more optimistic results could be an artifact of the way the model handles uncertainties.

Cities fed by the Colorado River basin have reduced water use through programs aimed both at demand and reuse. In Las Vegas, for example, residents receive cash rebates for replacing water-heavy lawns with desert plants. The city, which gets 90% of its water from the Colorado River, also returns, after treatment, 40% of the water it uses to Lake Mead, where it can then be reused.

Cities have taken significant steps to reduce their footprint, and they will likely continue to improve, said Brad Udall, a senior water and climate research scientist at Colorado State University’s Colorado Water Center who was not involved in the study. As a result, some of the region's biggest cities use less water per capita than they did a few decades ago.

But the more efficient cities get, the fewer opportunities they will have to save and reuse water, Udall told Live Science.

Bigger water users

The results highlight a reality scientists have long known: Demand management alone cannot offset the Colorado River's dwindling flows, said Sharon Megdal, a professor and director of the Water Resources Research Center at the University of Arizona, who was not involved in the study.

Cities make up only 18% of water use in the region, while agriculture guzzles more than 70% of the basin's water. "Agriculture's the big user," Megdal told Live Science.

"You can't solve this problem without dealing with agriculture in a major way," Udall agreed. "Because agriculture is 70% of the problem, it needs to be at least 70% of the solution."

Individual farmers have senior water rights to Colorado River water, meaning they receive their full allocation regardless of whether there is a shortage, but political pressure is rising to allocate more water to cities and cut farmers' consumption, Udall said. For example, water managers may decide that farmers in Arizona have to relinquish water to supply the Central Arizona Project canal, which delivers water from the Colorado River to the central and southern parts of the state, including Phoenix and Tucson.

Since 2019, the Colorado River has shrunk so much that it is now 35% smaller than it was on average in the 20th century. "We've never seen flows like this," Udall said. "If these flows continue to drop, I do see agriculture in the Lower Basin not getting the supplies they do [now] and those supplies being reallocated to cities."

The problem is that farmers in the Lower Basin grow crops that benefit the region economically through exports, including alfalfa, cotton, vegetables and citrus fruit. In Arizona, roughly 20% of cropland is used to grow alfalfa for cattle feed. The state is also the biggest cotton producer in the Colorado River basin, and it grows up to 90% of leafy greens consumed in the U.S. and Canada.

To cut their water consumption, farmers can change their irrigation techniques and crop patterns, Megdal said. Drip irrigation, which delivers water directly to the roots, cuts evaporation and reduces water use by up to 50% compared with methods such as flood irrigation used for cotton and alfalfa. Thirsty crops such as alfalfa could be replaced with low-water forage mixes consisting of wheat, barley, oats, rye and peas. Recently, there has also been a focus on guayule, which is a substitute for rubber, as an alternative crop for farmers, Megdal said.

But farmers respond to demand and may have long-term contracts with buyers. "It has to be economical," Megdal said.

Agriculture should switch to more efficient irrigation, but it's also important to realize that some seemingly wasteful methods, such as furrow or flood irrigation, can replenish aquifers, which are also being depleted, Udall said.

Some farmers are letting their fields go fallow, which could happen more and more as water shortages intensify, Megdal added.

Renegotiating water rights

The Colorado River Compact is up for renewal this year, meaning there is an opportunity to devise a more sustainable agreement, Obringer said. But negotiations have stalled between the Upper Basin, which includes Colorado, New Mexico, Utah and Wyoming; and the Lower Basin, which encompasses Arizona, Nevada and California. While officials wrangle over the precise wording in the document, the underlying problem is that the Lower Basin needs more water than it's getting, even though the Upper Basin already uses less water than it was allocated in the compact.

Officials in the Upper Basin have argued that cuts should now fall exclusively upon the Lower Basin. But that can't happen, Udall said, because cuts need to be deep across the Colorado River basin to make a difference. No matter what the Upper Basin thinks it is entitled to, it must accept reductions in water use, he said.

Reality check

Solutions must be found and implemented immediately, because the Colorado River's future has never looked bleaker.

Winter brought barely any snow, and the little that fell melted in early March instead of April, meaning river flows this spring and summer could hit record lows. "It's unprecedented; it's human-caused; it's scary, frightening, awful," Udall said.

Udall estimates that Lake Mead, which serves almost 25 million people in cities such as Las Vegas and Los Angeles, and Lake Powell, which supplies Lake Mead and Indigenous tribes, will be only about 20% full over the coming months. Without adjustments in reservoir release, water levels could drop low enough to prevent energy production at Glen Canyon Dam, which usually produces enough electricity to power over 350,000 homes.

"We're close to dead pool," Udall said. This is when the water level in a reservoir falls so low that it can't flow downstream. "That's never happened, and it's very serious."

It would take years of unusually high precipitation and runoff to recover from the current state, according to Megdal. "It is a challenging situation," she said. "You really have to adjust to a new normal."

To make matters worse, "data centers are popping up everywhere in this region," Obringer said. A midsize data center uses up to 300,000 gallons (1.4 million liters) of water per day for cooling, and this number will only increase as temperatures rise. Much of the water that runs through the system can be reused, but some is consumed to generate the electricity to power data centers.

Some small steps to mitigate the problem are already in the works. Some municipalities are buying water rights and groundwater from rural areas in the region, Megdal said. For instance, the rapidly-growing town of Queen's Creek, Arizona, is purchasing groundwater from farmers and investors in the sparsely populated Harquahala valley. Arizona is also drawing up agreements with California and Mexico to obtain desalinated water.

"I do think there's a tremendous capacity to adapt," Megdal said.

But a more durable solution will require a major overhaul of the existing agreements and water rights to reflect reality, Udall said.

"We've got to have an agreement on how to share the water we have and not pretend we live in the past," he said. "We need to adjust this system to deal with a completely new reality of much less water flow. We've got to balance the books here."

While flying just a few thousand miles above the moon on April 6, Artemis II astronauts reported seeing a handful of bright, fleeting flashes of light on the lunar surface, leaving mission scientists on Earth buzzing with excitement.

The excitement comes with good reason for scientists planning future lunar missions: These brief flashes, caused by tiny meteorites striking the moon, help researchers track when and where impacts occur. Such data can improve scientists' understanding of the risks these impacts pose to long-term infrastructure and a sustained human presence on the moon.

The moon's heavily cratered surface is no stranger to such impacts, and smaller ones occur almost daily with minimal risk to future lunar bases. But with space rocks slamming into the moon at tens of thousands of miles per hour — and no atmosphere to slow them — larger impacts remain a concern for scientists monitoring potential threats to critical infrastructure and crew safety.

"One has to plan for the less-frequent, more hazardous events," David Kring, a scientist at the Lunar and Planetary Institute in Houston, told Live Science. Back in 2016, Kring proposed the idea that future moon explorers should try to detect such impact flashes on the lunar far side. "The lunar flashes the Artemis II astronauts observed remind us to plan for that contingency," Kring said.

"Audible screams of delight"

The Artemis II crew spotted the flashes during their historic flyby around the lunar far side, when the moon briefly blocked the sun and created a total solar eclipse that lasted nearly an hour and left the far side in complete darkness. Against this stark backdrop, the crew observed at least four millisecond-long flashes on the lunar surface, and possibly as many as six, commander Reid Wiseman radioed to mission control in Houston.

"There was a little bit of giddiness," Wiseman said through the mission livestream, reporting a fresh flash spotted by crewmate Jeremy Hansen even as he relayed earlier observations. "It was definitely impact flashes on the moon."

"Incredible news, Reid," science officer Kelsey Young responded from mission control, raising her hand to her forehead in awe. There were "audible screams of delight" from scientists as the observations came in, Young told reporters at an April 7 news conference.

Scientists on Earth have already begun working to match the crew's observations with data from the moon-orbiting Lunar Reconnaissance Orbiter, with the goal of refining existing models of how frequently such impacts occur. To do so, researchers plan to combine astronaut reports with orbital data to extract key measurable details about the flashes, including their brightness, the meteorites' masses, and whether the events created fresh craters on the lunar surface.

Meteorites and moonquakes

A less-visible consequence of these impacts is the shock waves that ripple through the moon. Such "moonquakes," when recorded by seismometers, can reveal how energetic and potentially damaging each strike was.

Like the Artemis II crew, Apollo astronauts reported seeing a handful of impact flashes, but most occurred too far from the seismometers operating at the time to be detected, a recent analysis found. Even so, Apollo-era instruments recorded about 1,700 impact-related moonquakes. Some were powerful enough to "cause boulders to roll downslope," Kring said, while the largest could "cause crater and canyon walls to collapse," which would pose risks to future lunar habitats.

Currently, however, no active seismometers are operating on the moon to support the Artemis era.

"Hopefully the first landed Artemis mission will change that," Nick Schmerr, a geophysicist at the University of Maryland who co-authored the analysis cross-referencing Apollo-era flashes with moonquakes, told Live Science.

Seismometers are among the instruments NASA plans to deploy through a fleet of early robotic missions ahead of crewed landings planned for 2028 and beyond. By combining astronaut observations with data from seismometers, telescopes and orbiters, scientists could precisely locate impact events and track their effects.

"We'd certainly want to know about any nearby impacts for a moon base, especially if they hit near crucial infrastructure," said Schmerr, who is also the deputy principal investigator for a planned seismometer known as the Lunar Environmental Monitoring Station for Artemis (LEMS). "The more observations of a particular flash from different observers, the better."

Over time, such observations could help scientists refine estimates of how much asteroid and comet debris is striking the moon, Kring said. Larger impacts can excavate material from deep beneath the surface to reveal lunar geology that would be inaccessible otherwise. If such events occur near the polar regions, Kring added, they may even expose ice that NASA and other space agencies suspect is present and could someday be harnessed for life support and rocket fuel.

Looking further ahead, analyzing moon samples ferried to Earth from Artemis landing sites could allow researchers to identify the types of meteorites striking the moon and track how their composition has shifted over roughly 4 billion years of solar system history, Kring said. Ultimately, though, these impacts matter "because they produce and modify the lunar soil that future Artemis astronauts will walk on," he added.

How much do you know about the moon? Test your knowledge with our moon quiz!

Imagine a single vaccine that protects against the flu, COVID-19, lung-invading bacteria and seasonal allergens all at once.

Scientists recently developed a nasal spray to protect against multiple respiratory diseases, including seasonal influenza and COVID-19, which they've so far tested in mice over a three-month period. They call the promising invention a "universal respiratory vaccine," though it doesn't stimulate the immune system the way classic vaccines do.

So, how far away are we from a truly universal vaccine that can target all respiratory bugs? Experts told Live Science that while the new research is interesting, a truly universal vaccine is still years away.

Chasing universal protection

Flu and COVID-19 shots need to be updated every year to remain protective, because influenza viruses and coronaviruses are constantly mutating.

"As they circulate, viruses mutate in big and small ways," said Dr. Alfredo Mena Lora, medical director of infection control at Saint Anthony Hospital in Chicago, Illinois. "So antibodies that worked last season may not bind as well this season, leaving more people vulnerable to infection."

What's more, there are myriad respiratory viruses, bacteria and allergens floating around that aren't the target of any existing vaccines. These problems have driven research into "universal" vaccines that could provide broader, longer-lasting protection against multiple respiratory insults — including even seasonal allergies.

This push for universal vaccines accelerated after the COVID-19 pandemic exposed how vulnerable the world is to new respiratory pathogens, as well as how quickly existing vaccines become obsolete in the face of mutations. Since then, researchers have focused on developing vaccines that could last longer and protect against more variants, potentially reducing the need for frequent updates of vaccine formulas.

Universal vaccines in the works

Many universal vaccines in development aim to target parts of viruses that change very little across strains.

For example, for influenza, researchers are targeting the hemagglutinin protein that sticks off the viral surface, but they are focusing on the "stalk" rather than the "head" of this protein, because the stalk mutates more slowly. The National Institutes of Health's (NIH) FluMos-v2, which targets hemagglutinin from six flu strains, recently completed early-phase human trials and generated encouraging immune responses.

The NIH's Generation Gold Standard initiative also aims to develop universal vaccines to protect against multiple viruses that are likely to spark future pandemics. One intranasal flu vaccine is already in later-stage human trials. It uses whole, inactivated viruses to spur the body to make antibodies, which block infection, and T cells, which attack infected cells. This approach could provide broad protection against multiple influenza strains and potentially block transmission — something current flu vaccines do not do.

Meanwhile, some scientists are working on pancoronavirus vaccines to protect against current and future coronaviruses, while others are exploring AI-designed vaccines. They build these by using computational tools to pinpoint regions of virus proteins that mutate very slowly and appear in many viruses. Both these efforts are still in early, experimental stages.

With no universal respiratory vaccines yet on the market, most of the R&D efforts have focused on developing vaccines for specific groups of viruses, such as influenza viruses or coronaviruses. The recent nasal spray study is unique in that it aims to protect against viruses, bacteria and allergens, rather than just one pathogen family.

Rousing innate immunity

Unlike traditional vaccines, the experimental nasal spray does not teach the immune system to recognize select proteins on a specific antigen, the researchers reported Feb. 19 in the journal Science. Rather, it revs up the immune system's first-line defense, known as the innate immune system, study senior author Bali Pulendran, a pathologist at Stanford University, told Live Science in an email.

This acts like an early warning system in the lungs, ready to detect and respond quickly to a wide range of pathogens, even ones the body has never encountered before.

"These [lung] cells are the first to sense infection and help determine how the immune response unfolds," said Pulendran. "And we've learned over the past decade that innate immune cells can also be 'trained' to respond faster and more effectively to future threats."

The concept builds on research with the Bacillus Calmette-Guérin (BCG) vaccine, which prevents tuberculosis. In 2023, Pulendran and colleagues found that mice given BCG saw T cells flock to the lungs. There, they released signals that kept innate immune cells active in the lungs for months, protecting the mice against both COVID-19 and influenza.

This new nasal spray rouses similar immune protection. It combines two adjuvants, or substances that trigger an immune response, to activate T cells and draw them to the lungs. These T cells send chemical signals that mimic natural infection cues, keeping the lungs' innate cells activated and on high alert. If a pathogen makes its way into the lungs, innate immune cells will be primed to nip the infection in the bud.

In experiments, mice received four doses of the nasal spray, spaced one week apart, and were exposed to coronaviruses 21 days to 3 months after their final dose. Vaccinated mice had about 700 times less virus in their lungs than unvaccinated mice and also maintained their weight and survived the infections. Unvaccinated mice, in contrast, lost significant weight, experienced lung inflammation, and, in some cases, died.

The vaccine also helped mice fight off bacterial infections caused by Acinetobacter baumannii and Staphylococcus aureus weeks to months after dosing. For example, levels of S. aureus in the kidneys were about 200 times lower in vaccinated mice than in unvaccinated mice.

The vaccine also reduced the severity of allergic reactions caused by house dust mites. The vaccine seeds the lungs with long-lived T cells that alter the lung environment such that it suppresses the immune pathway behind allergies, the researchers reported. As a result, when vaccinated mice encountered dust mites, their immune systems didn't recruit inflammation-driving cells or generate mucus as they might have otherwise. Removing the T cells abolished this protection.

From mice to humans

"This concept sets up a stronger, faster immune posture in the respiratory tract," Mena Lora, who was not involved in the study, said of the nasal spray. Although the results are still early, he added, the work offers an important proof-of-concept.

The study only tested a handful of pathogens, so although the vaccine raises broad defenses, it's too early to declare it a universal respiratory vaccine. And in the long run, translating these findings to humans will be complex.

"The human immune system varies greatly," said Mark Cameron, an associate professor of population and quantitative health sciences at Case Western Reserve University School of Medicine. "Whether this vaccine can generate broad protection without side effects will require careful clinical trials," said Cameron, who was not involved in the study.

Pulendran agreed, adding that spurring the immune system in this way carries potential risks, such as triggering excessive inflammation. "In our animal studies, we did not observe pathological inflammation, but these questions will need careful examination in human studies," he cautioned.

The team is now preparing for early-stage trials in humans. The exact timeline to potential approval remains uncertain.

"If successful, such a vaccine could reduce hospitalizations, ease ICU [intensive care unit] strain, and protect populations during seasonal outbreaks and future pandemics," saidMena Lora.

The World Health Organization estimates next‑generation or universal influenza vaccines could prevent up to 18 billion flu cases and save millions of lives globally if widely used from 2025 to 2050. However, as no universal flu vaccine has reached the market yet, these ambitious goals are still on hold.

"Their ultimate impact will depend on effectiveness, duration of protection, and performance in diverse populations," Mena Lora emphasized. "The more tools we have — different platforms, targets, and delivery methods — the better our chances of developing broadly protective vaccines, while still refining pathogen-specific shots for high-risk groups."

As the general public has embraced large language models (LLMs) such as ChatGPT, Claude and Gemini, scientists have been exploring how these artificial intelligence (AI) tools could enhance medical research.

Some argue that LLMs could dramatically boost researchers' efficiency in completing certain types of medical studies, and research published in February in the journal Cell Reports Medicine exemplifies that vision for the technology.

The study used massive datasets of patient biomedical information to predict the risk of preterm birth in a given pregnancy. These types of predictions have been a powerful AI use case for years, and were possible with more traditional types of machine learning than LLMs employ. But this study was notable in that LLMs enabled junior researchers — a graduate student and a high school student — to efficiently generate very accurate code.

That code predicted a baby's gestational age at birth and the likelihood of preterm birth. The AI's output matched and, in one case, even beat analyses from expert teams who had used human-generated code to crunch the same data.

"What I saw with junior scientists here and how effective they could be truly inspired and amazed me," said study co-author Marina Sirota, interim director of the Baker Computational Health Sciences Institute at the University of California, San Francisco.

One big promise of LLMs is to lower the barrier for researchers to produce code and conduct complex analyses — but it comes with risks. As AI quickly improves, researchers must grapple with myriad questions. What guardrails need to be established to ensure AI's accuracy? How do we measure its output? And how will the role of human researchers evolve as these systems gain prominence?

How AI prediction works

Sirota's team drew on data used in the Dialogue for Reverse Engineering Assessments and Methods (DREAM) Challenges, international competitions in which teams of scientists tackle complex biomedical problems using shared datasets.

The open-source datasets included blood transcriptomics, which looks at RNA, a molecule that reflects which genes are active in the body. They included epigenetic information from placental cells, which described chemical tags that sit "on top of" DNA and control which genes can be switched on, and microbiome data describing the bacteria present in vaginal fluid samples.

These data points were flagged with the type of sample they came from — blood, placental tissue or vaginal fluid — and labeled with outcomes of interest, namely gestational age and preterm birth. Machine learning algorithms can then be trained to spot links between a sample's source and its label. For example, they may reveal that microbiome samples with certain mixes of bacteria often come from people who have given birth early.

Once trained on a subset of data, the algorithm can be tested on samples that lack labels, to see if it can predict the label that should be there. For instance, it should flag samples with bacterial mixes similar to those in the training data linked to a higher risk of preterm birth.

The final step is to evaluate the models' accuracy and compare them. "Accuracy" in the context of machine learning has a specific definition: the number of correct predictions divided by the total number of predictions.

Human- vs. AI-generated code

The DREAM Challenge was aimed at uncovering links between these medical metrics and the risk of preterm birth. Some risk factors, including having infections during pregnancy, are already well known. But the DREAM Challenge wanted to see what signals might be gleaned from clinical samples, like blood.

It's the kind of work that normally demands months of effort from trained bioinformaticians. But instead of writing the analysis code themselves, the junior researchers in the recent study gave each of eight LLMs a single prompt describing the data available and the labeling task at hand: predicting gestational age or preterm birth.

With this simple prompting, four of the eight models — DeepSeekR1, Gemini, and ChatGPT's o3-mini-high and 4o — produced code that ran successfully. The best performer, OpenAI's o3-mini, was as accurate as the original human DREAM Challenge teams. For one task, which involved estimating gestational age from epigenetic data, it was more accurate than humans had been.

What's more, the junior researchers generated results in about three months and submitted a manuscript describing their results within six months, whereas the same process took the original DREAM Challenge teams years.

"We got lucky with the review process here, but six months to generate the results and write the paper is pretty incredible, especially for a junior scientist," Sirota told Live Science.

Preterm birth, before 37 complete weeks of pregnancy, affects roughly 11% of infants worldwide. Babies born too early are at higher risk than full-term babies for a host of health troubles, including but not limited to problems affecting their brains, eyes and digestive systems. Being able to predict which pregnant patients are more likely to give birth early could mean closer monitoring and treatments to protect the baby and make full-term birth more likely, experts say.

Beyond writing code

The data used in the Cell Reports Medicine paper started "in good shape," Sirota noted, in tables that AI could easily read. "But we can speed that up as well — the cleaning part and normalization of data — with generative AI," she said.

Sirota's team is now exploring other LLM applications, including a new tool called Chat PTB (short for "preterm birth") that they've developed. The Chat GPT-based tool is embedded in papers published by the March of Dimes research network, part of a nonprofit aimed at improving maternal and infant health. Instead of manually combing through this literature, researchers can now query Chat PTB and get synthesized answers with references — a task that used to take hours, compressed into seconds.

But tools like Chat PTB and the code-writing approach in Sirota's study represent only the first wave. AI-enhanced medical research is moving toward "agentic" AI, meaning systems that don't respond to only one prompt but instead carry out multistep research workflows with increasing autonomy.

Instead of responding with only text, an agentic agent is capable of checking and iterating on its own work until it reaches its objective. It can also take action on a user’s behalf, like searching the internet and running code, rather than just writing it.

That shift toward greater AI autonomy and less human oversight brings both enormous potential and serious risk. In a January study published in the journal Nature Biomedical Engineering, researchers evaluated LLMs on 293 coding tasks drawn from 39 published biomedical studies, initially allowing the LLMs to come up with workflows on their own. They found that the overall accuracy came in below 40%.

Their solution was to separate planning from execution: They had the AI produce a step-by-step analysis plan that a human researcher reviewed before any code got written. The approach boosted the accuracy to 74%.

"The goal is not to ask researchers to blindly trust an AI system," study co-author Zifeng Wang, who was a doctoral student at the University of Illinois Urbana-Champaign at the time of the study, told Live Science in an email.

Instead, the goal is to "design frameworks where the reasoning, planning, and intermediate steps are visible enough that researchers can supervise and validate the process," said Wang, who is a co-founder of Keiji AI.

Why safeguards matter

These risks don't mean researchers should shy away from AI, but they do need to apply the same rigor to AI-generated work that they would to any other collaborator's output, scientists caution.

"The question is not whether LLMs accelerate science or create 'AI slop,'" Ian McCulloh, a professor of computer science at Johns Hopkins University's Whiting School of Engineering, told Live Science in an email. "The question is how we leverage this powerful technology within the scientific method."

But McCulloh also cautioned against holding AI to an impossible standard. People tend to assume AI is error-prone and downplay human error, he said, when, in reality, both humans and machines make mistakes. He anecdotally described a consulting client who lamented AI's 15% miss rate on a certain task, not realizing his human employees' miss rate was 25%.

"The goal of AI is not perfection," McCulloh said, "but to do better than people."

That effort will involve agreeing on how to measure AI's success. Dr. Ethan Goh, a physician-researcher at Stanford University, pointed out that health care still lacks standardized benchmarks for evaluating AI's performance. Goh recently published a randomized trial in JAMA Network Open that studied how LLMs influence doctors' reasoning in determining diagnoses.

Because LLMs are trained on such a vast amount of data, "benchmarks are so expensive to produce," Goh told Live Science. What's more, he said, AI improves so quickly that most commercial models start beating the few benchmarks that exist and rapidly render them useless. Amid these challenges, Goh's team at Stanford's AI Research and Science Evaluation (ARISE) Healthcare Network is working to develop such standards by the end of this year.

For all the uncertainty around standards and safeguards, the researchers who spoke with Live Science shared a common conviction: AI belongs in the lab, but not unsupervised.

"We have to be careful not to forget what we know in terms of the scientific process," Sirota said. "But I think the opportunity is tremendous."

Rates of diabetes are lower in high-altitude locations, but researchers have been unsure why. Now, a new study in mice reveals a possible explanation: Red blood cells, which play a pivotal role in transporting oxygen throughout the body, may lower blood sugar by converting glucose into a compound that helps release oxygen into tissues.

If the results can be replicated in people, they also hint that drugs in early-stage development could potentially mimic this pathway.

"The work highlights the important role that red blood cells can play in diabetes regulation," study lead author Isha Jain, a biochemist at the Gladstone Institutes and the University of California, San Francisco, told Live Science. "That's the concept to be targeted in the future."

Higher altitude, lower blood sugar

It's well-known that people living at high altitudes with low oxygen levels, such as the Andes and the Himalayas, tend to have lower rates of diabetes, but the reason for the link has not been clear. In a 2023 study, scientists observed the same phenomenon in mice: When the mice were exposed to low-oxygen conditions, they developed a condition called "hypoxia," which occurs when oxygen supply to tissues is insufficient, and their blood glucose also dropped.

But the disappearing glucose couldn't be explained by the amount of glucose absorbed by the muscles and other organs in the scans, so it wasn't clear where it was going.

From high altitudes to lab chambers

To test whether red blood cells were responsible for lowering glucose, the study authors exposed mice to low-oxygen chambers containing 8% oxygen. This mimicked high-altitude air, while another group of mice was kept in air with 21% oxygen, which mimicked normal atmospheric conditions, Jain said.

After several weeks, both groups of mice were given glucose injections, and their blood sugar levels were measured over time. Compared with the mice in normal oxygen environments, the mice in low-oxygen conditions showed a much smaller spike in their blood sugar levels, suggesting that they could clear glucose from their blood faster. This effect persisted for weeks, even after the animals were returned to normal oxygen levels, suggesting that a low-oxygen environment had a lasting impact on metabolism, experts said.

The researchers also took imaging scans to track how much glucose was being absorbed by major organs and tissues, such as the liver and muscles. However, a large fraction of the disappearing glucose could not be accounted for. This prompted them to investigate whether cells in the circulating blood itself might be consuming the glucose.

To test this idea further, they manipulated red-blood-cell numbers directly. The team periodically removed blood in oxygen-deprived mice to keep red-blood-cell levels near normal, and found that doing so eliminated the glucose-lowering effect of hypoxia. In contrast, transfusing red blood cells into mice breathing normal air caused blood glucose levels to fall, suggesting that the number of red blood cells alone drove down glucose levels.

Next, the team injected mice with labeled glucose and tracked it through the body. They found that red blood cells from the oxygen-deprived mice absorbed substantially more glucose than those from the comparison mice. The mice in low-oxygen conditions rapidly converted glucose into a molecule that binds to hemoglobin, the protein in red blood cells that carries oxygen. This binding forces hemoglobin to release oxygen more easily into tissues when oxygen levels are low.

Further analysis showed that red blood cells produced in the oxygen-deprived mice also contained higher levels of a protein called GLUT1, which sits on the cell membrane and helps glucose enter the cell. These red blood cells had about twice as much GLUT1 and took up roughly three times more glucose than normal red blood cells. By labeling existing red blood cells before exposing the mice to low-oxygen conditions, the researchers confirmed that only the new cells produced under low-oxygen conditions showed these adaptations.

Besides triggering an uptick in red blood cells, the study shows that the cells are structurally changed to consume more sugar in low-oxygen environments, said Daniel Tennant, a hypoxia and metabolism researcher at the University of Birmingham who was not involved in the work.

Lars Kaestner, a red blood cell biologist at Saarland University in Germany who was not involved with the study, noted that red blood cells are known to increase in number when the air is thin, to boost oxygen transport around the body. Red blood cells use glucose as fuel. Therefore, it's not surprising that low-oxygen conditions lead to lower blood glucose levels, as more red blood cells are there to clear it, he told Live Science.

"From a systemic point of view, this makes a lot of sense,” he said.

It's an "evolutionarily conserved corrective mechanism" to essentially better oxygenate the body at high altitudes, Tennant told Live Science.

The body increases its red-blood-cell count at high altitudes by changing the expression of genes that control metabolism and producing more of a hormone called erythropoietin, which triggers the bone marrow to churn out more red blood cells, said Sonia Rocha, a biochemist at the University of Liverpool who was not involved in the study.

This is why elite athletes train in high-altitude areas for their competitions: Their bodies make more red blood cells and thus achieve "more efficient circulation to distribute oxygen to their tissues," she told Live Science.

A diabetes drug that mimics oxygen deprivation?

In another experiment, the researchers treated mice with HypoxyStat, an experimental compound that was developed in Jain's lab and increases how strongly hemoglobin binds to oxygen, preventing its release and mimicking hypoxia. The idea is that mimicking oxygen deprivation with a drug could boost red-blood-cell counts and help regulate blood sugar levels.

However, much more testing is needed before a drug like HypoxyStat could be tested in humans, Rocha noted.

While transfusing red blood cells is not a practical therapy for diabetes, the findings suggest potential directions such as engineering RBCs that act as better glucose sinks, the authors suggest. “It opens the door to thinking about diabetes treatment in a fundamentally different way,” Jain said in a statement.

For 60 years, the earliest known human species has also been one of the most mysterious. Homo habilis was added to our family tree in 1964. But it's long been unclear exactly what the ancient species, which lived between about 2.4 million and 1.65 million years ago, looked like.

That's because, until recently, only three very incomplete fossilized skeletons had been unearthed.

Then in January, researchers described a fourth, more complete skeleton — and it revealed that H. habilis had an anatomy very unlike our own. The discovery has some researchers asking a big question: Is the earliest known human ancestor not human after all?

"As we have discovered more fossils, we've stretched the definition of the Homo genus," Bernard Wood, a paleoanthropologist at the George Washington University in Washington, D.C., told Live Science. "Maybe this time we just stretched it too far."

Obviously our species, Homo sapiens, belongs in the Homo genus. We also know that our closest living relatives, the chimpanzees and bonobos, don't. This means that the human genus evolved at some point after our evolutionary lineage, which includes humans and our closest extinct relatives, split away from the chimpanzee line, an event that occurred more than 5 million years ago. So, when exactly did the human genus evolve?

One approach would be to argue that it dates to the split with the chimpanzee lineage. But the first creatures that appeared after the split don't look much like we do. They include species, like Australopithecus afarensis, that had long, ape-like arms and relatively small brains.. This species existed in Africa between about 3.9 million and 2.9 million years ago, and includes the famous Lucy skeleton. Very few researchers consider Lucy to be human.

Most anthropologists, however, have historically considered H. habilis a member of the Homo genus.

Few skeletons

The first, very incomplete Homo habilis skeleton was discovered in Tanzania in the 1960s. The 1.75 million-year-old specimen included fragments of the skull, from which it was possible to estimate that they came from an individual whose brain had been roughly 45% the size of the average living person's. This may sound small, but that was substantially bigger than the average australopithecine's brain, which was about 35% the size of ours. Because of this evidence, the skeleton was placed in our Homo genus, and given the name Homo habilis, meaning handy or skillful human, it was a decision that most researchers have accepted.

But the H. habilis skeleton described in 2026 complicates things. This skeleton is 2 million years old, and was found in Kenya, about 500 miles (800 kilometers) north of where the first H. habilis remains were unearthed. Just like the first skeleton, the Kenyan skeleton is far from complete. But the bones that survived give us our best ever look at H. habilis's arms, said study co-author Carrie Mongle, a paleoanthropologist at Stony Brook University in New York. The problem is that those arms aren't like ours. Instead, they are long and ape-like, similar to the arms of our australopithecine relatives like Lucy.

"They are very much australopith-like," Ian Tattersall, a paleoanthropologist at the American Museum of Natural History in New York City, told Live Science. In an article published earlier this year, Tattersall argued that these ape-like arms are a clear indication that H. habilis wasn't a member of the human genus.

He isn't the first to make this suggestion. Wood and his colleague, Mark Collard, an archaeologist at Simon Fraser University in Canada, argued in 1999 that H. habilis wasn't a member of the human genus. By then, the second and third H. habilis skeletons had been discovered, and although extremely incomplete, they hinted that the species had limb proportions unlike ours — something that the fourth skeleton confirms.

Wood and Collard suggested transferring the species to the same genus as Lucy, which would mean renaming Homo habilis to Australopithecus habilis. Tattersall doesn't think that's a good solution, because the species had human-like brain size and teeth. He thinks habilis should be put in its own genus, although he hasn't yet come up with a name.

A different approach

Other researchers, meanwhile, suspect that Wood and Tattersall are both wrong.

They think there is no need to rename H. habilis despite its arms. "Those ape-like limb proportions don't necessarily tell us all that much," Carol Ward, an anthropologist at the University of Missouri, told Live Science. This is because of the way most evolutionary scientists operate when they are defining species and genera.

We know that our very earliest ancestors, living just after the split with the chimpanzee line, spent plenty of time climbing trees, where long, ape-like arms would have been useful. Gradually, they adapted to spend more time walking on the ground before ultimately evolving into humans.

These bipedal ancestors probably no longer needed long, ape-like arms. But crucially, Ward said, long arms were almost certainly not a hindrance to survival either. Under those circumstances, even the first species in the Homo genus might have retained the long arms of their ancestors, because there was no strong evolutionary pressure to shorten them. Why arms eventually shrank is still not entirely clear, although some researchers think shorter arms may have brought some subtle advantages while running and using tools. This suggests there was weak evolutionary pressure for shorter arms, meaning they shrank, but at a relatively slow rate.

There's a broader point here. "We want to think there was this big change with Homo, that we're different from everything else that came before," Ward said. "But this H. habilis skeleton supports the idea that maybe there was a more gradual transition from australopiths to Homo."

This idea highlights an awkward problem that scientists are still grappling with.

Evolution is so complicated that it's surprisingly difficult to divide living things into clear groups, such as species, which is one reason why there are now dozens of different ways to define species, and a heated debate on which one is the best. It turns out that genera are just as difficult to define, which means there isn't actually any agreement on what a genus is, Wood said.

In other words, researchers will probably continue to debate whether or not H. habilis is in the human genus, for the simple reason that they still can't fully agree what a genus actually is.

Human origins quiz: How well do you know the story of humanity?

Architectural constraints in today's most popular artificial intelligence (AI) tools may limit how much more intelligent they can get, new research suggests.

A study published Feb. 5 on the preprint arXiv server argues that modern large language models (LLMs) are inherently prone to breakdowns in their problem-solving logic, known as "reasoning failures."

Reasoning failures occur when an LLM loses track of key information needed to reliably solve a task, resulting in incorrect answers to seemingly straightforward problems. The paper, which was presented as a review of existing research, looked specifically at transformer models, a type of neural network architecture that underpins popular AI chatbots including ChatGPT, Claude and Google Gemini.

Based on LLMs' performance on evaluations such as Humanity's Last Exam, some scientists say the underlying neural network architecture can one day lead to a model capable of reaching human-level cognition. While transformer architecture makes LLMs extremely capable at tasks like language generation, the researchers argue that it also inhibits the kind of reliable logical processes needed to achieve true human-level reasoning.

"LLMs have exhibited remarkable reasoning capabilities, achieving impressive results across a wide range of tasks," the researchers said in the study. "Despite these advances, significant reasoning failures persist, occurring even in seemingly simple scenarios ... This failure is attributed to an inability of holistic planning and in-depth thinking."

Limitations with LLMs

LLMs are trained on huge amounts of text data and generate responses to user prompts by predicting, word by word, a plausible answer. They do this by stringing together units of text, called "tokens," based on statistical patterns learned from their training data.

Transformers also use a mechanism called "self-attention" to keep track of relationships between words and concepts over long strings of text. Self-attention, combined with their massive training databases, is what makes modern chatbots so good at generating convincing answers to user prompts.

However, LLMs don't do any actual "thinking" in the conventional sense. Instead, their responses are determined by an algorithm. For long tasks, particularly those that require genuine problem-solving across multiple steps, transformers can lose track of key information and default to the patterns learned from their training data. This results in reasoning failures.

"This fundamental weakness extends beyond basic tasks, to compositions of math problems, multi-fact claim verification, and other inherently compositional tasks," the researchers said in the study.

Reasoning failures are also why LLMs often circle the same response to a user query even after being told it's incorrect, or produce a different answer to the same question when it's phrased slightly differently, even when it's prompted to explain its reasoning step by step.

Federico Nanni, a senior research data scientist at the U.K's Alan Turing Institute, argues that what LLMs typically present as reasoning is mostly window dressing.

"People figured out that if you tell an LLM, instead of answering directly, to 'think step by step' and write out a reasoning process first, it often gets the right answer," Nanni told Live Science. "But that's a trick. It's not real reasoning in the human sense — it's still just next‑token prediction dressed up as a chain of thought," he said. "When we say these models 'reason,' what we actually mean is that they write out a reasoning process — something that sounds like a plausible chain of reasoning."

Gaps in existing AI benchmarks

Current ways to assess LLM performance fall short in three key areas, the researchers found. First, results can be affected by rewording a prompt. Second, benchmarks degrade and become contaminated the more they're used. And finally, they only assess the outcome, rather than the reasoning process a model used to reach its conclusion.

This means current benchmarks may significantly overstate how capable LLMs are and understate how often they fail in real-world use.

"Our position is not that benchmarks are flawed, but that they need to evolve," study co-author Peiyang Song, a computer science and robotics student at Caltech, told Live Science via email. Likewise, benchmarks tend to leak into LLM training data, Nanni said, meaning subsequent LLMs figure out how to trick them.

"On top of that, now that models are deployed in production, usage itself becomes a kind of benchmark," Nanni said. "You put the system in front of users and see what goes wrong — that's the new test. So yes, we need better benchmarks, and we need to rely less on AI to check AI. But that's very hard in practice, because these tools are now woven into how we work, and it's extremely convenient to just use them."

A new architecture for AGI?

Unlike other recent research, the new study doesn't argue that neural-network approaches to AI are a dead end in the quest to achieve artificial general intelligence (AGI). Rather, the researchers liken it to the early days of computing, noting that understanding why LLMs fail is key to improving them.

However, they do argue that simply training models on more data or scaling them up are unlikely to resolve the issue on their own. This means developing AGI may require a fundamentally different approach to how models are built.

"Neural networks, and LLMs in particular, are clearly part of the AGI picture. Their progress has been extraordinary," Song said. "However, our survey suggests that scaling alone is unlikely to resolve all reasoning failures … [meaning] reaching human-level reasoning may require architectural innovations, stronger world models, improved robustness training, and deeper integration with structured reasoning and embodied interaction."

Nanni agreed. "From a philosophy‑of‑mind point of view, I'd say we've basically found the limits of transformers. They're not how you build a digital mind," he said. "They model text extremely well, to the point that it's almost impossible to tell if a passage was written by a human or a machine. "But that's what they are: language models … There's only so far you can push this architecture."

Is the idea of the metaverse dead? Even without a much hotter technology in the form of artificial intelligence (AI) capturing the public conversation, most ordinary people have stopped talking about it, beyond reminiscing about the COVID-19-era hype and catty quips at the technology's expense.

The hype was once so loud that one of the biggest names in the technology industry capitalized on it by changing its own name — we're looking at you, Meta, the company formerly known as Facebook. Several years on, however, Meta has quietly divested itself of its interests in the area. After losing more than $70 billion since 2021 as of December 2025, the firm was preparing to cut metaverse development outfit Reality Labs' funding by 30%.

Recently, Meta announced it was shutting down its virtual reality environment Horizon Worlds in June 2026 — meaning the new paradigm we were all promised would have become one of technology's infamous short-lived flameouts. Then, days later, it reversed course — with company representatives saying the platform would remain available on Quest, Meta's VR headset.

The original iteration of virtual reality (VR) emerged in the 1990s, primarily for gaming, but the theory goes back further, with immersive digital worlds featured in science fiction as far back as the work of author Ray Bradbury in the 1960s. A real-world attempt at digital worlds, Second Life, made an early splash when released in 2003, but technical and copyright issues hobbled adoption.

The idea was resurrected by Facebook in the early 2020s. CEO Mark Zuckerberg believed we’d increasingly use shared, immersive experiences that go beyond flat screens as digital social interaction became central to life and work, and scientific studies confirmed users were ready, particularly after COVID lockdowns had made remote work the new normal.

But, several years later, things have changed. While Meta wasn't the only company touting a new generation of VR, it’s a fairly damning indictment of the faith the technology industry has in this concept. But does that mean it's dead and buried for good? Although it was a non-starter for many, for others, it's not dead but simply still finding its place in the modern world.

The failed promise of the metaverse

"The branding is wounded, or under correction from its initial expectation," said Lik-Hang Lee, assistant professor of augmented reality (AR) and VR at the Hong Kong Polytechnic University. He authored a study on the metaverse, which was published Oct. 21 in the journal Computers and Society. "The grand vision of a single, shared virtual universe where we all work, play, and socialize in headsets all day was always a bit of a sci-fi oversell."

Futurist Mark van Rijmenam, who writes and speaks publicly about future technologies, agrees. He added that the vision of cartoon avatars in virtual reality (VR) lobbies was never going to happen. But he thinks the metaverse is very much alive nonetheless. "It's maturing into something more meaningful than the hype once promised," he told Live Science. "What felt like abandonment was actually a pivot beneath the surface. It's being rebuilt with purpose, not PR, and with technology that's actually ready for the spatial internet."

There's no denying that something went seriously wrong with this idea after Meta popularized it during and after the COVID-19 pandemic. Spending our time in virtual utopias, at least in the short term, did not enjoy wide appeal, and what little interest there was evaporated. When asked why metaverse proponents failed to convince the masses, Lee said there was no "killer app" for non-early adopters.

"VR meetups and virtual offices were marketed as the future, but for most of us, they were just clunkier versions of Zoom, Slack, or games we already had," he said. "The friction of putting on a headset is high and the reward wasn't clearly higher."

For van Rijmenam, the metaverse that received all the hype also missed one crucial unique selling point — it was disconnected from real use cases. "It focused on virtual hangouts and flashy graphics instead of solving real-world problems or merging with everyday workflows. Early metaverse projects promised novelty before utility," he said.

A common theme that experts brought up was that the hardware wasn't seamless or mature enough to deliver the promised experience. "The current hardware suffers from limitations like a small field of view, heavy designs, motion sickness and poor graphics," Bob Gourley, CTO at intelligence and analysis firm OODA, told Live Science. "The future of the metaverse lies in the hands of technologies in AI, 5G, edge computing, and display like microLEDs and better optics that are still to come before it can be fully realized. The debate now isn't about the metaverse being dead or alive but how the technological support will catch up."

Presenting a particular stumbling block were the VR headsets touted by Oculus (later Meta) and Sony (for use with PlayStation). Not only are they bulky, and much harder to set up and use compared to a laptop or phone, but reports of headaches and nausea were widespread, thanks to something known as a "vergence-accommodation conflict."

We focus on an object when the brain uses muscles to pull the eyes in different directions so that their combined focal point converges on an object, no matter how far away. But when you wear a VR headset, your eyes constantly focus on a small flat screen just fractions of an inch from your eyes, an illusion that works — but only up to a point. Prolonged exposure causes a contradiction between the visual field and how your brain directs muscles in your eyes to focus in response, a phenomenon that was central to a 2024 study in the Journal of Optometry.

"Interestingly, humans aren't purely visual-first organisms," said Jennalyn Ponraj, founder of Delaire, a research lab focused on voice and human nervous system regulation in AI systems. "Presence is actually regulated through interconnected systems that include vestibular balance, proprioception, breath, and timing. When you flood vision with high-resolution but low-latency input, the rest of the sensory system receives conflicting or absent signals, and it often results in fatigue, nausea, dissociation, and cognitive strain. The technology functions, but the models of human perception are incomplete. Meta's divestment looks like an admission that immersion ultimately depends on attunement to biophysical regulation."

This lack of comfort is something of an insurmountable barrier, experts suggest, with virtual worlds likely going nowhere until accessing them is as easy as putting on a pair of glasses and forgetting they're there. We also can't forget one of the most common barriers to uptake of any new technology paradigm: tech fatigue.

"A lot of people already feel overwhelmed by digital life," said Lee. "Asking them to strap a gadget to their face for casual interaction is a big ask. There wasn't a clear "why" — why do this in VR instead of just using a phone or laptop? Outside of gaming and some enterprise use cases, developers struggled to find sustainable business models and reasons for people to come back."

The case against the metaverse's revival

But not everyone’s so certain of such manifest destiny for the metaverse. As long ago as 2023, a market research firm told the media: "The metaverse was briefly attractive to enterprises, but few invested seriously in moving the concept further within their organization."

Even before that, when metaverse hype was still at fever pitch, a Pew Research study reported that 46% of respondents said the metaverse won’t be a refined, truly immersive and functioning aspect of daily life by 2040.

Former software engineer at Facebook, Unity and Adobe, Mircea Dima, now currently founder, CEO, CFO and software engineer at AlgoCademy, thinks there’s little future in mass-market metaverse use. The limiting factor, he told Live Science, is human behavior — not graphics or computing power.

"Wearing a headset isolates you from other people, puts a physical strain on your body, and demands prolonged focus," he said. "No one will wear a device on their face for an extended period of time for tasks that can be done just as quickly with a laptop or smartphone. Hardware getting to a point where it’s comfortable to wear on your head isn’t going to remove that friction."

He added that the metaverse still has proponents because the concept is "emotionally appealing."

"Engineers love working within a closed system where almost anything is programmable. Investors hate to admit they’ve thrown good money after bad and founders are often emotionally attached to a publicly promoted vision of what they wanted to create, and won’t easily abandon it."

In one example, Mark Friend, Company Director at education IT provider Classroom365, delivered and ran VR classroom pilots. While such programs captured students’ attention, he said, the metaverse will never be mainstream in the field. He told Live Science the pilot data showed increases in the level of focus in students of up to 150%, but the model collapsed in the real world. The primary barriers to scale, he added, was hardware cost and the time to meet compliance requirements.

Saswata Baksi, co-founder of Local Glyph, tells a similar story, saying the metaverse is a non-starter because it solves a problem users don’t have. "A retail brand spent six months creating a virtual showroom for their customers to browse products as avatars," he told Live Science. "Peak usage was 200 people in the first week — but by the end of the first month, usage had plummeted to single digits. It simply took too long for customers to navigate the virtual space. They preferred scrolling through Instagram on their phones, which offered no setup friction."

As for why so many organizations continue to assert the metaverse will have its day, almost everybody Live Science spoke with highlighted the "sunk cost fallacy" phenomenon.

The metaverse is down, but is it out?

Many think the general philosophy of meeting, working and playing in virtual worlds in digital form is still coming. Lee said that if you strip away the hype, more and more of our lives are moving into persistent digital spaces: "Games, virtual collaboration tools, industrial simulations and augmented reality on phones and headsets are all growing."

But perhaps the ultimate irony is that AI, which left the metaverse in the dust on the hype scales, will be one of the cornerstones that enable it. "The original metaverse failed because content creation and responsiveness were manual and static. AI changes that," said van Rijmenam.

Joachim van der Meulen is the secretary and facilitator of DROPS Asia, an industry association to prevent harm from dropped objects in the workplace. He began building VR training content in 2017 and quickly learned "how difficult and expensive it is to deliver a genuinely good immersive experience," but now he can see a way to rekindle the effort by using AI.

"Much of our effort went into maintaining core systems and keeping the software up to date, leaving little budget or capacity for new features or content," he said. "AI-assisted development proved effective for routine, labour-intensive tasks, in some cases solving long-standing technical problems, and we’re now seeing experimentation with generative AI for 3D environments."

Lee agreed, adding: "If anything like the metaverse is ever going to work at scale, AI will probably be a huge part of what makes it viable when it comes to content creation, intelligent characters and agents, personalization and usability and real-time understanding of the 'world model' and translation."

We also can't forget the roll call of technologies, from the Apple Newton to Google Glass, that only failed because they were too early — a position the metaverse finds itself squarely in today, with almost everybody we spoke to for this article agreeing it still shows potential despite a subpar early hardware experience. We can also see echoes of the Google Glass in the Ray-Ban Meta devices and similar AR-powered smart glasses today, which face regulatory and privacy hurdles.

"The metaverse feels like it's in its middle phase," said Lee. "The initial "this will change everything tomorrow" phase is over. What we have now is slower, more incremental progress. It's a classic technology trough — less visible, less glamorous, but often where the most important engineering and design work actually happens."

Even though Meta has cut funding and effectively neutered the idea, for many others, the underlying technologies and architectural foundations are now emerging, says van Rijmenam. AI, spatial computing, AR and VR hardware, as well as real-world physics models driven by spatial intelligence and real-time rendering, are all converging to provide a foundation for a future attempt to resurrect the metaverse. "What once felt abandoned was actually incubation," he says.

Flying cars have long captured the imagination of science fiction writers, and in recent years, a host of tech startups have been rushing to make the dream a reality. Following a long period of research and development, more bullish voices in the industry say the launch of air taxi services is imminent — but significant technical, regulatory and economic hurdles may yet stymie their take off.

U.S. companies Joby Aviation and Archer have both announced plans to launch air taxi services in Dubai, in the United Arab Emirates (U.A.E.), later this year, which would mark a major milestone in the technology’s winding road to commercialization.

Despite the enduring appeal of this new form of urban mobility and efforts by aviation authorities to lay the regulatory groundwork, however, experts say they’re not yet ready for prime time. Concerns around safety, question marks about the financial viability of air taxi services, and the challenges involved in setting up the infrastructure and operational capacity to support a completely new transport network mean we could still be waiting at least a decade for an airborne alternative to Uber.

"We think that full-scale services is more of a middle of the next decade kind of thing, not anytime soon," Sergio Cecutta, founder and partner at SMG Consulting, which covers the advanced air mobility sector, told Live Science.

Despite lofty ambitions, plans to roll out flying taxis to date have so far fizzled out. There were, for example, plans to introduce flying taxis in time for the 2024 Paris Olympics — but they were scrapped due to delays over the certification for the vehicles' engines. Around the same time, Archer set out plans to introduce flying taxis ahead of the upcoming FIFA World Cup, for flights in LA. This now seems unlikely, with the company now targeting the 2028 LA Olympic Games for the introduction of new passenger transport services.

In recent weeks, however, the U.S. Federal Aviation Administration (FAA) has announced the latest phase of its advanced air mobility program that will evaluate air taxi performance in eight separate projects over the summer across 26 states. Does that mean we can expect flying taxis to take to the skies this year?

Certification challenges

The promise of electric vertical take-off and landing (eVTOL) aircraft is to create a quieter, cleaner and more efficient alternative to helicopters. While designs vary between companies, they typically rely on distributed electric propulsion (DEP) — spreading multiple electric motors and propellers across the airframe to create something that looks more like a drone than a conventional rotorcraft.

Some eVTOLs, like those made by German company Volocopter and Chinese company EHang, feature propellers fixed in a vertical position. But other companies like Joby and Archer are experimenting with propellers that can shift from a vertical to horizontal position, which allows them to combine vertical takeoff with more efficient forward flight.

Because they are powered by batteries, they could be both greener and quieter than conventional aircraft. In theory, at least, the increased efficiency and simplicity of electric motors could also make eVTOLs cheaper to build and operate, according to NASA. Proponents say this could make it feasible to operate large numbers of these aircraft in urban areas at a price point that many more people would be able to afford.

But getting a new class of aircraft off the ground is not simple. While eVTOL companies have carried out hundreds of hours of test flights, starting commercial operations will require them to go through onerous certification processes with aviation authorities. These include bodies like the FAA, the European Union Aviation Safety Agency (EASA) and the U.K. Civil Aviation Authority (CAA).

While the specifics vary from country to country, this is likely to involve around 1,000 hours of test flights overseen by regulators to ensure they can safely fly the kinds of missions the developers envisage, said Cecutta. Given that even the most advanced eVTOL developers have managed just a few hundred hours of test flights, normally in several different aircraft designs, he says certification is still some way off.

"There's not enough time on this planet for them to do it in a year," said Cecutta. "We think certification, even for the most advanced of the bunch, is a 2027 thing. And for some other companies, it might be a 2028 or 2029 milestone."

Early pilots launch — but technical challenges remain

Regulators in the U.A.E have committed to fast-tracking approvals for eVTOLs, but Cecutta suspects this won't be a full certification in the traditional sense. Normally, certification allows an aircraft to fly anywhere under various conditions, but he says the U.A.E's General Civil Aviation Authority (GCAA) is expected to issue an airworthiness certificate that allows flights only on very restricted routes that avoid flying over densely populated areas, such as from Dubai airport to the Palm Jumeirah resort island.

While this may provide companies with some early operational experience, aerospace analyst Bill Sweetman, principal at Valkyrie Strategic Solutions, says it's unlikely to accelerate the technology’s broader rollout. Leading aviation authorities typically have agreements whereby certifications are reciprocally recognized between countries. But because the U.A.E. has never certified an aircraft before, this is not likely to happen here.

Companies in the U.S. do have another avenue for getting services up and running before full certification. In September 2025, the FAA introduced the eVTOL Integration Pilot Program (eIPP) that will allow uncertified aircraft to operate in controlled environments in a series of pilots designed to test out various use cases for the technology, including air taxi services.

But Cecutta emphasizes that this is unlikely to accelerate the widespread deployment of the technology. The program is designed to help companies gain operational experience so that they can launch commercial services faster following certification, but it won’t speed up the certification process itself.

There may yet also be significant speed bumps in efforts to prove the airworthiness of these vehicles. Richard Brown, an aerodynamics consultant at Sophrodyne Aerospace in Scotland, says there are technical complexities to eVTOL operations that many in the industry have been reluctant to fully address.

One major concern involves what happens when these multi-rotor aircraft interact with the ground. Research Brown conducted for the CAA has revealed that the downwash from eVTOL rotors can create highly concentrated airflows that travel considerable distances across the ground with surprising strength. This could damage surrounding infrastructure or even send people flying unexpectedly.

Brown suggests that eVTOL company modeling focuses on average airflows and fails to capture these dynamics. But the phenomenon could restrict where these aircraft can safely operate and what kind of infrastructure they'll need.

Even more concerning, he says, is the risk of a vortex ring state, a dangerous aerodynamic condition that can cause rotors to suddenly lose thrust. This is already a major safety concern for helicopters, which experience vortex ring state accidents every year, he adds, but eVTOL designs with multiple interacting rotors may be even more susceptible to these effects.

Facing up to hard economic truths

Even if startups overcome the technical hurdles, significant questions remain about economic viability. Although eVTOLs are likely to be cheaper to operate than conventional aircraft — because electric propulsion is inherently more efficient and less prone to failure — they are likely to be considerably more expensive to purchase for the foreseeable future, Cecutta said.

Costs will come down as companies scale up production, and plans to enable autonomous flight could remove the considerable cost of training and employing pilots, he adds. But this will take time and Cecutta estimates that it could be a decade before eVTOL services become a transportation option for middle-class passengers, rather than just the ultra-wealthy.

Sweetman is more sceptical about whether eVTOL operations will ever achieve the scale required to significantly reduce costs — with per-vehicle costs decling with many more vehicles manufactured. He questions whether urban airspace can safely accommodate the hundreds or thousands of aircraft required to make the economic case stack up.

And long-term operational costs may be considerably higher than many have suggested. Unlike electric cars, which draw power relatively steadily, eVTOLs subject their batteries to very high discharge rates during takeoff and landing.

"You're beating the hell out of the battery," said Sweetman. He points to regulatory filings from U.S. eVTOL developer Beta Technologies that suggested they would need to replace their aircraft's batteries every year, a cost that "alone is enough to destroy the economics."

In the near term, cargo operations and emergency medical services may prove a more viable use case for eVTOL aircraft than air taxi services, according to Cecutta. These applications don't require flying over dense populations and face fewer public acceptance hurdles than large-scale urban transportation networks. Many eVTOL companies are also branching out into defense applications, he adds, where they can tap the deeper pockets of military customers.

These applications could provide the financial breathing room for eVTOL manufacturers to continue the development of their technology, said Cecutta. Despite outlandish projections that flying taxi services could commence this year, for flying taxis to truly take off in a meaningful way, the industry still faces years of work ahead, with no guarantee that all the pieces will ultimately come together.

What counts as an addiction — should compulsive shopping or gaming qualify?

Researchers are currently debating whether to classify problematic behaviors as addictions in the next edition of psychiatry's "bible" of mental health disorders. The current version of the American Psychiatric Association's Diagnostic and Statistical Manual of Mental Disorders, DSM-5, includes gambling disorder as a behavioral addiction, while all other addictions mentioned are tied to substances such as alcohol, tobacco, stimulants, marijuana and opioids

But a recent study suggested that compulsive shopping also may qualify as a behavioral addiction. That raises the question of whether other compulsive behaviors, such as excessive gaming or incessant social media use, could also be considered addictions.

Shopping addiction?

In the new study, researchers compared three types of shoppers in a group of 241 German adults: those with a pathological desire to shop, dubbed "compulsive buyers"; "risky" or impulsive shoppers, whose buying behavior was influenced by external factors like advertisements or life events; and people who engaged in intentional, planned shopping.

The study, published Feb. 16 in the journal Comprehensive Psychiatry, found that compulsive buyers scored higher on measures linked to addictive behaviors, including distress and the use of shopping for emotional gratification or compensation. They also reported lower self-esteem and higher levels of anxiety and depression compared with the other shoppers. Additionally, compulsive shoppers showed higher levels of impulsiveness and weaker performance on tasks that measure self-control.

Together, the findings suggest that compulsive shopping is more than just a bad habit and should be classified as a behavioral addiction, the study researchers argue. But making that case requires a definition of addiction.

Defining addiction

Not every intense habit qualifies as an addiction.

Dr. Nathan Carroll, a board-certified psychiatrist and author of "Internet Gaming Disorder: A Clinical Strategy Guide for Providers, Parents, and Players" (American Psychiatric Association Publishing, 2025), said addiction is defined by how much it disrupts a person's life.

"When a behavior becomes so all-encompassing and so obsessive that you get functional impairment across multiple domains — socially, occupationally, educationally — that's where it becomes an addiction," Carroll told Live Science.

According to the DSM, a person can be diagnosed with gambling addiction if at least four of the following happen within one year:

● Frequent thoughts about gambling (such as reliving past gambling or planning future gambling)

● Need to gamble with increasing amounts to achieve the desired excitement

● Repeated unsuccessful efforts to control, cut back on or stop gambling

● Restlessness or irritability when trying to cut down or stop gambling

● Gambling when trying to escape from problems or negative mood or stress

● After losing an item of value by gambling, feeling the need to continue to get even ("chasing" one's losses)

● Often gambling when feeling distressed

● After losing money gambling, often returning to get even

● Lying to hide the extent of gambling involvement

● Losing important opportunities, such as a job or school achievements or close relationships, due to gambling

● Relying on others to help with money problems caused by gambling

Mark Griffiths, a behavioral addiction researcher at Nottingham Trent University in the U.K., has studied a wide range of compulsive behaviors, including gambling, video gaming, internet use, exercise, work, and social media or smartphone use. He has also investigated excessive tanning, TV binge-watching, dance obsession, and even the obsessive viewing of eating videos, called mukbang.

Through four decades of research, Griffiths identified these six key criteria that help define addiction, which applies to both substance addictions and behavioral addictions:

1. Salience: The behavior becomes the most important activity in a person's life.
2. Mood modification: The person uses the behavior to change how they feel.
3. Tolerance: Increasing amounts of the activity are needed to achieve the same effect.
4. Withdrawal: Stopping the behavior triggers negative psychological or physical symptoms.
5. Conflict: The behavior harms relationships, work or education.
6. Relapse: The person repeatedly returns to the behavior after trying to stop.

Griffiths noted that people need to meet all six of his criteria for their behavior to be classified as an addiction. According to his criteria, very few people would be classified as having a true shopping or gaming addiction, he told Live Science.

More often, people meet several criteria but not all of them, in which case the behavior may be better described as "problematic" rather than truly addictive.

Risks and benefits of classifying behaviors as addictions

Expanding the list of recognized behavioral addictions has both potential risks and benefits.

One concern is the possibility of pathologizing normal behavior. People who spend many hours working out, tabletop gaming or video gaming could be labeled as addicted when they are simply dedicated to their hobby, Carroll said. If diagnostic criteria for gaming addiction are too rigid, there is a risk of classifying professional gamers as addicts when they are not, he noted, and this can undermine the credibility of the diagnosis.

On the flip side, classifying more behaviors as addictions means having a useful framework for clinicians to distinguish normal behavior from pathological behavior, Carroll said.

These criteria would also help people recognize a problem they often can't see. "The nature of addiction is that a lot of people don't even see it in themselves when the addiction is occurring," Carroll said.

Identifying an addiction is the first step toward treatment, which often involves addressing underlying conditions such as anxiety or depression, he said.

While Carroll is a self-confessed "huge fan of gaming," he is in favor of recognizing behaviors like gaming disorder as addictions. "I do think the pendulum is swinging too far into the realm of obsession, and I think people are missing out on real-life experiences in pursuit of virtual experiences," he said. "So I'm definitely in the camp where we should look at this as a behavioral addiction."

Will the DSM-6 include new behavioral addictions?

The recognition of behavioral addictions has evolved gradually over time. When the DSM-3 was introduced in 1980, pathological gambling was listed as an impulse-control disorder, like kleptomania and pyromania. It wasn't until the release of the DSM-5 in 2013 that it was renamed gambling disorder and reclassified as a behavioral addiction, based on brain imaging studies and neurochemical tests that showed gambling activates the reward system in a similar way that a drug does.

Internet gaming disorder also appears in the DSM-5 but as a condition that requires further research. The main reason it wasn't classified as an addiction "was the lack of research backing it up," Carroll said. "But I think we've produced quite a bit of research supporting that now."

Griffiths agrees that the evidence base has grown substantially. "DSM-5 was published 13 years ago, and there's been so much research now," he said. "In DSM-6, gaming disorder will definitely be a fully recognized disorder, just like gambling disorder."

Other diagnostic systems have already moved in that direction. The International Classification of Diseases 11th Revision, compiled by the World Health Organization and finalized in 2019, includes gambling disorder, gaming disorder and compulsive sexual behavior disorder.

Still, experts disagree about how quickly additional behavioral addictions will be recognized.

Aside from gaming addiction, Carroll believes there is strong evidence for shopping addiction to be included in the DSM. Beyond that, "there's some good evidence for social media addiction and pornography addiction," he said.

Griffiths takes a more cautious stance. While he believes addiction to social media, shopping, exercise, work and sex does occur, he argues that the current evidence base is still limited. "Most of the research is of relatively low quality," he said. "We just haven't got the quality research that we've got for gambling and gaming."

Another issue is that it doesn't make sense to classify something as a disorder without evidence that there is effective treatment, Griffiths said.

Without stronger epidemiological studies, biological research and treatment evidence, it's not a good idea for other behavioral addictions to be included in the DSM, he said.

Perhaps, in 30 years, the situation will change, Griffiths said. But for now, he predicts that gaming disorder will be the only new behavioral addiction likely to be included in the next DSM.

Microplastics that accumulate in the body can disrupt the immune system by preventing immune cells from gobbling up microbes and clearing dead cells, a new study in mice and lab dishes shows. Although it's still unclear how this might play out in the human body, experts say the study has potential implications for human health.

Microplastics are tiny particles that break off of plastic products, like water bottles, food storage containers and freezer bags. Although these undigestible particles accumulate in the body, scientists remain unsure how they contribute to illness. The new study revealed that these tiny particles can clog up and disrupt the function of macrophages, cells that engulf and destroy potentially harmful viruses, bacteria or fungi, as well as dead cells.

Microplastics contaminate air, soil and water; are abundant in food and drinks; and have been detected in various human tissues and fluids, including blood, breast milk, the brain, the liver and reproductive organs. Eliseo Castillo, who researches how microplastics affect health at the University of Mexico but was not involved with the study, told Live Science in an email that "because plastic production continues to increase and these materials slowly fragment over time, microplastic pollution is expected to keep rising in the future."

The buildup of microplastics in the body has been correlated with atherosclerosis, in which plaque builds up in the inner walls of arteries, causing them to become narrow and rigid; neurodegeneration; and cancer, said Justin Perry, an immunologist at Memorial Sloan Kettering Cancer Center and senior author of the study, which was published March 10 in the journal Immunity.

Nonetheless, researchers have not proved that microplastics directly cause disease. "I think that really set off myself and other colleagues in the field to try and understand this," Perry told Live Science. For the new study, he and his colleagues focused on polystyrene, a common plastic that is used in food containers and breaks down into microplastic particles that can be as small as a virus or as large as a grain of sand.

The team focused on macrophages, which engulf material, such as dead cells and infectious microbes, from their surroundings and break them down into sugars, amino acids and fats that the cells can recycle. Scientists have shown in mice and humans that macrophages actively take up microplastics but can't break them down.

"That's kind of scary, because that means that we really haven't evolved a response to them," Perry said, so it's possible that cells hoard increasing volumes of microplastics throughout a person's lifetime, he added.

Perry and his colleagues administered this nonbiodegradable plastic to human macrophages cultured in lab plates. They also inserted the microplastic particles into the airways of mice, euthanized them and extracted their lungs to image on a microscope.. Cells that took up the particles struggled to engulf and destroy dead cell matter or infectious microbes like bacteria or fungi, the team discovered.

To determine if microplastics could prevent macrophages from fighting infections, the team administered the fungus Aspergillus fumigatus, which causes respiratory infections in immunosuppressed people, to the mice's lungs. They found that the mice that were given microplastics struggled to clear the infection and experienced worse disease.

"Whether this leads to higher infection rates in people is still unknown," Castillo said. "At this time, there is no clear public health evidence directly linking microplastic exposure to increased infections in humans."

Microplastics also stopped macrophages from engulfing and destroying dead cell matter, potentially causing litter to build up in tissues. Perry said people lose 2% of their body mass per day due to cell death, and it's up to macrophages to clear away the dead litter. "It's about 3 million [dying] cells per second. If you really do the math, it's quite a large burden [for macrophages]," Perry added

A pileup of dead cell material like DNA can alarm the immune system because it signals that a threat such as an infection may be causing tissue damage. This could trigger inflammation, potentially precipitating autoimmune disorders, Castillo said.

Microplastics might also reduce fertility when taken up by macrophages in the testes. Perry and his colleagues exposed male mice to regular doses of microplastics and found that their sperm counts decreased over 18 weeks. He speculated that microplastics may partly account for the drop in sperm count seen globally in the human population. The study didn't explore the potential impact of these particles on female fertility.

Although the researchers demonstrated that human macrophages engulf microplastics when cultured in lab dishes, researchers still don't know how readily the cells would do so in the body.

Perry plans to study samples taken from humans ‪—‬ specifically, how microplastics might contribute to atherosclerotic plaques that clog up blood vessels. It's already known that macrophages loaded with undigested material accumulate in these blood-flow-stopping plaques, so he and his colleagues hypothesize that microplastics may exacerbate the problem, he said.

This article is for informational purposes only and is not meant to offer medical advice.

For much of the 20th century, artificial intelligence (AI) struggled not because researchers lacked ambition, but because the hardware available to power it simply wasn't powerful enough. Early AI systems hit hard limits on processing speed and memory, contributing to repeated "AI winters" as progress stalled and funding dried up.

That problem is mostly gone now. Today, AI models are trained on specialized chips in huge data centers, and they can scale up in weeks instead of years. Compute, which used to be the main bottleneck, is now something that can be bought with enough money. Companies like Nvidia or AMD are also mass-producing even more powerful graphics processing units (GPUs) — components conventionally used for gaming or visualization but also well suited to processing AI calculations — as each year goes by.

So, beyond the fundamental architectures at the heart of these models, what's keeping AI from becoming even more advanced? The new limit is far more physical in nature — and far harder to work around. It’s electricity.

Why AI’s energy appetite is exploding

Modern AI models don’t just train once and then stop. They run all the time, powering things like chatbots, search tools, image generators and more autonomous agents. This change has made AI a constant, large-scale user of electricity.

According to Sampsa Samila, academic director of the AI and the Future of Management Initiative at Barcelona’s IESE Business School, the problem isn’t a lack of energy in absolute terms. "It’s not the overall supply of energy, but having reliable, firm capacity at the right place and the right time that is in short supply," he told Live Science.

Predictions for AI energy consumption show this strain clearly. The International Energy Agency (IEA) expects data centers to consume more than twice as much electricity by the end of the decade, reaching levels similar to those in major industrial economies. In some parts of the U.S, data centers already use as much power as heavy industry.

How AI is actually used matters just as much as how it’s trained. Training large language models (LLMs) still consumes a lot of power, but it tends to occur in large, infrequent runs. What’s growing faster is the everyday work — models responding to users, over and over again. Samila notes that newer "reasoning" systems, which spend more time working out an answer, push energy use into normal operations rather than occasional training bursts.

A grid built for a slower world

Power grids were designed for gradual growth, not for city-sized loads appearing almost overnight.

Juan Arismendi-Zambrano, an assistant professor at Ireland's University College Dublin (UCD) Michael Smurfit Graduate Business School, said the main issue is timing. Large AI campuses grow faster than grid upgrades or government approvals can keep up with. This creates a real bottleneck: getting enough power, when and where it’s needed.

"The ‘short supply’ of AI electricity is, in my view, less about an absolute global lack of electricity and more about local bottlenecks created by fast deployment of large data centres," Arismendi-Zambrano told Live Science.

"These campuses scale quicker than electricity grid upgrades, or bureaucracy can respond. Especially when they land in rural areas chosen for cheap land and political ‘lobbying’ for states, but not engineered for sudden, concentrated load. The result is a very physical constraint: access to a lot of electricity power, on time, at the right node," he said.

Clustering data centers in one area makes the problem worse. Jens Förderer, a professor at the University of Mannheim Business School in Germany, pointed to Northern Virginia’s "Data Center Alley," where many facilities draw huge amounts of power from the same grid. Power plants, transmission lines and substations take years to build, but AI companies often start using compute much sooner, sometimes even before their buildings are finished.

"When many city-scale loads draw from the same local grid, scaling electricity provision becomes far harder," Förderer said.

How the industry is scrambling to respond

There is no single fix for AI’s energy problem. Instead, companies are pursuing several strategies at once.

One is building power closer to the data centers themselves. Large tech firms have signed long-term contracts to support new power generation, including nuclear plants, and are exploring on-site power where grid upgrades move too slowly.

Google, for example, has been doing this in Texas through its acquisition of energy developer Intersect, which builds large-scale solar and storage projects alongside data center demand rather than waiting for grid upgrades. Microsoft, meanwhile, has signed a long-term deal with Constellation Energy tied to the planned restart of a nuclear reactor at Pennsylvania’s Three Mile Island site to supply power for its data centers.

Another is choosing locations based on electricity, rather than users. As Förderer noted, data centers are increasingly sited where power is easiest to scale, even if that means moving further from major population centers.

Then there is reuse — including a surprising source. Former cryptocurrency mining facilities are emerging as candidates for AI workloads. Once criticized for their energy use, these sites already have what AI needs most: large grid connections, cooling systems and experience running power-hungry hardware around the clock. The crossover between Bitcoin and AI may look strange, but the underlying physics is the same.

"These facilities already have large grid connections, and some former miners may pivot toward AI workloads," Förderer said.

Canadian miner Bitfarms has recently announced plans to transition its facilities away from Bitcoin mining toward high-performance computing and AI data centers, while Hut 8 — originally a Bitcoin mining company — struck a major $7 billion lease deal in late 2025 to provide data-center capacity for AI computing

Some ideas look even further afield. Space-based data centers are sometimes pitched as a way to sidestep Earth’s grid entirely, using constant solar energy and the cold of space for cooling. Samila said the idea works on paper, but the numbers get intimidating fast.

A single 5-gigawatt facility would require around 2.5 by 2.5 miles (4 by 4 kilometers) of solar panels in orbit. It’s "in principle doable," he added, but only with some serious engineering. Latency, upkeep and launch logistics remain open questions.

Efficiency may be the fastest lever of all. Förderer pointed out that advances in chips, model design and system architecture have already reduced the energy required per unit of intelligence. Some recent efforts include an MIT breakthrough that aims to cut energy use by stacking components vertically, as well as a "rainbow-on-a-chip" that uses lasers to transmit data in components.

Such gains won’t eliminate the need for more power, but they can slow the rate at which demand grows.

Does unlocking energy unlock smarter AI?

The growing demand placed upon the electricity grid by AI also raises environmental concerns. Engineer Aoife Foley, professor and chair in Net Zero Infrastructure at the University of Manchester in the U.K., pointed out that the wider IT sector already makes up about 1.4% of global carbon emissions.

AI workloads use much more energy than regular cloud computing, and while big tech companies are investing in renewables and better cooling, Foley said these efforts alone are not enough."These impacts can be reduced through smarter model optimisation and a closer alignment between data centre strategy and regional renewable generation," she told Live Science.

Despite the scale of the challenge, none of the experts see electricity as a shortcut to artificial general intelligence (AGI) —a hypothetical form of AI that can simulate behaviour as intelligent as, or more intelligent than, that of a human being. More energy makes it easier to build and run bigger systems, but it doesn’t solve the harder problems. Instead, Förderer argued that the real limits sit elsewhere — in access to data, in new model architectures and in genuine advances in reasoning.

"Energy is necessary but not sufficient,” Samila said in agreement, adding that today’s dominant approach to improving AI relies on massive amounts of power, but more electricity alone will not magically produce AGI.

More energy doesn’t guarantee smarter machines, but it does change who gets to participate. Access to power will shape where AI is built, who can afford to run it and how broadly it’s deployed. The bottleneck has shifted away from silicon and toward the physical world, where grids, permits, and power plants move at a very different pace than code.

Planting trees along coastlines with human-made shore defenses, such as dikes, could protect more than 140,000 people from flooding and save up to $800 million from flood damage globally each year, a new study finds.

Places that have mangroves, such as parts of Florida, are better able to withstand the ravages of storms and their powerful waves. But although there is a push to restore mangroves around the world, there are several challenges.

In 2022, Hurricane Ian slammed into southwest Florida. Storm-powered waves reached up to 18 feet (5.5 meters) and devastated coastal communities and infrastructure. The hurricane killed 158 people and generated $110 billion in damage in the state. Authorities say the storm surge, which is when the storm pushes water ashore, was the main cause of deaths.

But places in Florida with mangroves saw 30% less damage than areas without mangroves, saving about $13 billion. "Mangroves act as a sponge to incoming waves," Daniel Friess, an environmental scientist at Tulane University, told Live Science. "Their dense tangle of aboveground roots are great at soaking up incoming wave energy."

Mangroves are forests that exist in the intertidal zone between the ocean and land. Their trees can live in the salty water, and they are found in tropical and subtropical coastal zones.

Climate change is expected to make hurricanes more frequent, and rising sea levels will drive higher storm surges. Mangroves protect communities and infrastructure from these surges.

They could also help to combat climate change. A 2025 study found that restoring 1.1 million hectares (2.7 million acres) of mangroves globally would remove about 0.93 gigatons of carbon dioxide from the atmosphere. That's almost triple the emissions from cars in the U.S. It would cost about $10.73 billion to restore those mangroves, according to the study.

Despite their importance, the world's mangroves are in danger. More than half of Earth's mangrove ecosystems are at risk of collapse by 2050, according to a 2024 assessment by the International Union for Conservation of Nature. They are being replaced by agriculture and aquaculture.

Balancing costs and benefits

Researchers wanted to see how mangrove restoration around the world could protect people and prevent costly damage from floods, as well as determine where these measures might have the greatest impact.

In the study, published Jan. 20 in the journal PNAS, they modeled the effects of mangroves when the forests were combined with flood defenses, such as dikes or seawalls. Dikes are human-made structures that run alongside the ocean or rivers to stop water from overflowing onto land.

"We used a published mangrove restoration tool, which looks into where the mangroves have been lost based on satellite data, the hydrological conditions of those areas now" to determine whether those mangroves could be restored, study lead author Timothy Tiggeloven, a climate adaptation specialist at Vrije Universiteit Amsterdam, told Live Science. Then, the team combined that information with flood risk, future climate scenarios, changes in GDP and population, and sea-level rise.

They found that mangrove-dike systems could save a total of $800 million globally and protect 140,000 people from flooding each year. These numbers increased under different climate scenarios linked to human carbon emissions.

Their cost-benefit analysis suggests that in a high-emissions scenario, in which Earth's climate changes dramatically, every dollar spent on mangrove dike systems globally could ultimately generate — or save — $6. That could translate to as much as $125 billion by 2100.

The benefits were not the same everywhere, though. Countries in Southeast Asia would see the greatest benefits — about $270 million a year and 70,000 people shielded from flooding. West Africa was a close second, saving about $221 million and protecting 38,000 people. Nationally, Nigeria, India and Indonesia would benefit the most from restoring mangroves in front of human-made coastal defenses.

In the U.S., Florida would see significant benefits from restoring its mangroves, but Louisiana would get even greater returns, the study found.

Jonah Busch, an environmental economist and a former senior research fellow at the Center for Global Development who was not involved in the research, welcomed the study. "It combines the biophysical analysis of mangrove restoration with the engineering of dikes, and then economics," he said.

However, he would have liked to see a breakdown of the financial benefits of mangroves on their own. "They're assuming that places already have dikes and then you can add mangroves on top of that," he said.

The authors flagged this as a limitation of the study. The analysis relies on a flood-protection database, which lists existing infrastructure, and cannot say whether the dikes are strong enough or even still standing.

Gray-green strategies

Adaptation strategies that combine nature-based solutions and engineered infrastructure are sometimes called gray-green infrastructure. This area is "a new, open and important topic," Busch said.

Other examples include combining forest management with home hardening (which involves retrofitting or building houses with flame-retardant materials) to lower fire risk, and marrying dam maintenance with upstream watershed restoration.

"There is no doubt that a hybrid approach can be a pragmatic and effective approach" to coastal management, Thomas Westhoff, a nature-based solutions officer at the conservation nonprofit Wetlands International, told Live Science. That was especially the case along heavily urbanized, subsiding coastlines that have lost much of their mangrove cover, he added.

Westoff cautioned that there is no one-size-fits-all solution. "Whether this is a feasible solution is very context specific," he said, adding that dikes may not exist in many areas.

However, "in many regions, healthy mangrove belts can still provide enough of a buffer for coasts and communities as the climate changes," Westhoff said.

Challenges of restoring mangroves

There is a global push to restore mangroves, but a majority of these projects — up to 80% — fail.

"Restoring mangroves is a good idea, but these projects are difficult to implement," Tiggeloven said. Mangroves are sometimes planted in unsuitable places, or the wrong types of trees are planted.

Successful projects require community buy-in, Westhoff said. "When communities benefit directly from restored ecosystems — whether through sustainable harvesting or ecotourism — they are more likely to protect them for the future."

Plus, when restoring or preserving a mangrove, people may want to develop the land in other, more profitable ways, Busch noted.

"Mangroves have to compete with that from an economic perspective," he said. The new paper "is a key part of that, because it shows the economic value of mangroves' storm protection."

If you know what to listen for, a person's voice can tell you about their education level, emotional state and even profession and finances — more so than you could imagine. Now, scientists posit that technology in the form of voice-to-text recordings can be used in price gouging, unfair profiling, harassment or stalking.

While humans might be attuned to more obvious cues such as fatigue, nervousness, happiness and so on, computers can do the same — but with far more information, and much faster. A new study claims intonation patterns or your choice of words can reveal everything from your personal politics to the presence of health or medical conditions.

The research, published Nov. 19, 2025 in the journal Proceedings of the IEEE, highlights a grave concern for the technology's capability in privacy and unfair profiling.

While voice processing and recognition technology present opportunities, Aalto University's speech and language technology associate professor Tom Bäckström, lead author of the study, sees the potential for serious risks and harms. If a corporation understands your economic situation or needs from your voice, for instance, it opens the door to price gouging, like discriminatory insurance premiums.

And when voices can reveal details like emotional vulnerability, gender and other personal details, cybercriminals or stalkers can identify and track victims across platforms and expose them to extortion or harassment. These are all details we transmit subconsciously when we speak and which we unconsciously respond to before anything else.

Jennalyn Ponraj, Founder of Delaire, a futurist working in human nervous system regulation amid emerging technologies, told Live Science: "Very little attention is paid to the physiology of listening. In a crisis, people don't primarily process language. They respond to tone, cadence, prosody, and breath, often before cognition has a chance to engage."

Watch your tone

While Bäckström told Live Science that the technology isn't in use yet, the seeds have been sown.

"Automatic detection of anger and toxicity in online gaming and call centers is openly talked about. Those are useful and ethically robust objectives," he said. "But the increasing adaptation of speech interfaces towards customers, for example — so the speaking style of the automated response would be similar to the customer's style — tells me more ethically suspect or malevolent objectives are achievable."

He added that although he hasn't heard of anyone caught doing something inappropriate with the technology, he doesn't know whether it's because nobody has, or because we just haven't been looking.

We must also remember that our voices are everywhere. Between every voicemail we leave and every time a customer service line tells us the call is being recorded for training and quality, a digital record of our voices exists in comparable volumes to our digital footprint, comprising posts, purchases and other online activity.

If, or when, a major insurer realizes they can increase profits by selectively pricing cover according to information about us gleaned from our voices using AI, what will stop them?

Bäckström said even talking about this issue might be opening Pandora's Box, making both the public and "adversaries" aware of the new technology. "The reason for me talking about it is because I see that many of the machine learning tools for privacy-infringing analysis are already available, and their nefarious use isn't far-fetched," he said. "If somebody has already caught on, they could have a large head start."

As such, he's emphatic that the public needs to be aware of the potential dangers. If not, then "big corporations and surveillance states have already won," he adds. "That sounds very gloomy but I choose to be hopeful I can do something about it."

Safeguarding your voice

Thankfully, there are potential engineering approaches that can help protect us. The first step is measuring exactly what our voices give away. As Bäckström said in a statement, it's hard to build tools when you don't know what you're protecting.

That idea has led to the creation of the Security And Privacy In Speech Communication Interest Group, which provides an interdisciplinary forum for research and a framework for quantifying information contained in speech.

From there, it's possible to transmit only the information that's strictly necessary for the intended transaction. Imagine the relevant system converting the speech to text for the raw information necessary; either the operator at your provider types the information into their system (without recording the actual call), or your phone converts your words to a text stream for transmission.

As Bäckström said in an interview with Live Science: "The information transmitted to the service would be the smallest amount to fulfill the desired task."

Beyond that, he said, if we get the ethics and guardrails of the technology right, then it shows great promise. "I'm convinced speech interfaces and speech technology can be used in very positive ways. A large part of our research is about developing speech technology that adapts to users so it's more natural to use."

"Privacy becomes a concern because such adaptation means we analyze private information — the language skills — about the users, so it isn't necessarily about removing private information, it's more about what private information is extracted and what it's used for."

The biggest trees in the Peruvian Amazon store disproportionately more carbon than smaller trees do, a new study finds. But these larger trees are also the ones most likely to be harvested, which means more carbon is being released into the atmosphere, thereby reducing these forests' ability to act as carbon sinks, the researchers say.

Almost 60% of Peru’s land is covered by forests, the vast majority in the Amazon region where it accounts for around 11% of the total Amazon rainforest. Peru's current forestry law allows trees to be selectively harvested when they reach a minimum diameter, which ranges from 16 to 24 inches (41 to 61 centimeters) depending on the species. The terrain in the Peruvian Amazon makes accessing and removing trees difficult, so forestry companies prefer larger trees because they yield the most wood, which reduces transportation costs, labor and time. These trees also tend to be older and more mature, with denser, harder and more stable wood.

But these large trees also store the most carbon, and removing them returns much of that carbon to the atmosphere, study co-author Geomar Vallejos-Torres, an agricultural scientist at the National University of San Martín in Peru, told Live Science in a translated email.

To determine how much carbon is stored in these large trees, Vallejos-Torres and his colleagues measured hundreds of trees across five forests in the country, recording variables such as diameter, height, crown area and wood density to estimate aboveground and belowground biomass and stored carbon.

They found that carbon storage, both aboveground and belowground, increased disproportionately with increasing trunk diameter, with 16 inches being an important threshold. The forests they studied sequestered up to 331 metric tons per hectare (148 U.S. tons per acre) aboveground and 47 metric tons per hectare (21 U.S. tons per acre) belowground. The majority of that carbon — between 88% and 93%, depending on the species — was concentrated in trees larger than 16 inches in diameter. Breadnut trees (Brosimum alicastrum), for example, stored 11.4% of their total aboveground carbon in trees smaller than this cutoff point, compared with 88.7% in larger trees.

The work was published Jan. 25 in the journal Frontiers in Forests and Global Change.

The results show that Peru's forestry policy targets the trees that store the most carbon and suggest that the country should shift its forest policy to protect them, Vallejos-Torres said.

"Given the urgency of keeping carbon reserves out of the atmosphere, it is necessary to conserve trees larger than 41 cm," he said. "This also allows for the conservation of forest biodiversity and microfauna, buffering the microclimate in the face of future climate changes."

Size isn't everything

But some researchers say tree size is not always the most important consideration for carbon sequestration. Ulf Büntgen, a professor of environmental systems analysis at the University of Cambridge who was not involved in the research, said the length of time the carbon is stored is more relevant.

"The paper talks not much about tree ages and thus ignores carbon residence time, which is overall low in the tropics," he told Live Science.

Vallejos-Torres countered this argument, however, saying that the largest trees continue to accumulate carbon for centuries, while smaller ones grow too slowly to make up the difference.

"The regeneration of smaller trees is slow, uncertain and often limited by degradation, disturbance and microclimatic changes, so the lost carbon is not recovered on timescales relevant for climate mitigation," he said.

The relationship between tree diameter and carbon stocks highlighted in the study is both empirically valid and relatively intuitive, said Martin Perez Lara, director of forest climate solutions impact and monitoring for the World Wildlife Fund. But he added that focusing only on tree size is not the best way to design a climate-friendly forestry system.

"A large share of research shows that well-designed management systems, including selective harvesting forests — including some trees around 40 cm — can contribute positively to climate mitigation and reducing degradation risks, and support long-term carbon dynamics," he told Live Science.

Despite the urgent need to preserve carbon sinks in the Amazon, Vallejos-Torres does not have much faith that the necessary changes to Peru's forestry policy will be forthcoming. "A legal reform that protects the largest trees would directly affect the economic interests of the timber sector, which depends on the extraction of these individuals of high commercial value and has an important weight in forestry policy decisions in the country," he said.

A vaccine designed using "DNA origami" activated more of the key immune cells needed to fight HIV than did traditional vaccines built upon protein scaffolds, a new mouse study found.

"DNA origami" refers to a precisely engineered, three-dimensional scaffold made of folded DNA that can hold and display viral antigens — bits of viruses that the immune system can recognize and attack.

The new mouse study's results, published Feb. 5 in the journal Science, suggest a "potential breakthrough" that could "transform the way we think about active immunotherapies and vaccine design," study co-author Mark Bathe, a professor of biological engineering at MIT, said in a statement.

Why could this origami approach be transformational? The answer may lie in how the DNA-based vaccines are seen by the immune system, compared to traditional vaccines.

How the DNA origami vaccine works

Conventionally, vaccines have relied on weakened or killed viruses to rouse immune cells to make antibodies against proteins found on that virus's surface. By binding to the proteins, the antibodies block the virus from invading human cells and flag the germ for destruction by other immune cells.

This process imparts immunity by prompting the body to make "memory B cells," which linger and get activated much faster if the same pathogen is encountered again.

But nowadays, rather than using whole viruses, many vaccines use only the surface antigens attached to synthetic, virus-like particles. These nanostructures mimic the size and geometry of viruses but cannot cause infection.

Most virus-like particles used today are built using protein scaffolds that the immune system sees as "foreign," so they trigger an "off-target" antibody response against the scaffold itself. In some contexts, this may dilute the responses against the antigen, previous studies suggest.

In the new study, scientists replaced the protein scaffolds with a DNA-based scaffold and, by doing so, sharply reduced those off-target responses. This new vaccine design produced up to three times more of the important memory B cells than state-of-the-art protein nanoparticle vaccines did.

John Moore, an HIV researcher at Weill Cornell Medicine who was not involved in the work, described the study as "elegant." It clearly demonstrates how eliminating scaffold-related immune responses pushes the immune response "in the right direction," he told Live Science.

He cautioned, however, that it remains to be seen whether the same degree of immune focusing will occur in humans.

An edge on the competition?

HIV evades the immune system by constantly reshaping its surface proteins so that antibodies that work against one strain often fail against others. That is why HIV vaccine design has been "incredibly challenging," said Adam Wheatley, an immunologist at the University of Melbourne who was not involved in the study.

What the vaccine needs to produce are "broadly neutralizing antibodies" against the virus, he said. These antibodies lock onto parts of the virus that barely change from strain to strain.

One example of such an antibody is called VRC01, which has been identified in a small number of people living with HIV whose bodies produce broad antibody responses. VRC01 targets a vulnerable region on HIV's outer envelope called the CD4-binding site. This is the "key" the virus uses to enter human immune cells, and it doesn't differ much between strains.

The challenge is that B cells that are capable of producing VRC01-like antibodies are extraordinarily rare in the human body, said Raiees Andrabi, an immunologist at the University of Pennsylvania who was not involved in the work. Activating those elusive cells "becomes an engineering problem," he told Live Science.

To target these rare B cells, the researchers carefully engineered the vaccine using an HIV antigen on the DNA scaffolding. Developed about a decade ago, the antigen they used mimics the CD4-binding site and selectively binds the rare B cells' receptors, thus initiating the production of broadly neutralizing antibodies.

The researchers got the idea to combine the antigen with DNA origami after testing the origami approach in an experimental COVID-19 vaccine. They'd found that the immune system showed virtually no response to the DNA scaffold.

"This property seemed especially useful for a case like HIV, where the B cells of interest are exceptionally rare," first study author Anna Romanov, an immunology researcher at MIT, said in the statement.

They hypothesized that delivering the antigen on a silent scaffold could reduce the competition with other irrelevant B cells, thereby boosting the "on-target" response against HIV. And in the study, they found the silent-scaffold approach had indeed amplified the B cells that produce broadly neutralizing antibodies. (That said, they have not yet assessed how many broadly neutralizing antibodies actually get made; that should be addressed in future work.)

"We were all surprised" that DNA origami outperformed the standard virus-like particles used in eliciting the desired B cell responses, Bathe said.

In general, it's unclear how bad it is for the body to generate an immune response against the scaffold, Wheatley said. But in the case of HIV, the desired B cells are so rare that even a modest off-target response seems to undermine the response against the target antigen.

The road ahead

The engineering of the DNA-origami vaccine was not straightforward; early versions produced weak immune responses. That was partly because, after injection, those vaccines failed to reach specialized immune cells inside the lymph nodes, where B cells train.

To correct this, the team redesigned the DNA particles to pack the HIV antigens more precisely and tightly. This enabled them to be carried into the right regions within the lymph nodes. The researchers also added a molecule to help activate T cells — immune cells help the critical immune response grow. This T-cell recruitment happens naturally with protein-scaffold vaccines.

"I think it is quite striking how efficiently they modified their DNA scaffold in a number of ways to get it to work," Wheatley said. "I guess the major utility of it is [that] it's really tunable."

Beyond HIV, the study authors suggest that DNA origami could be applied to make vaccines against other rapidly mutating viruses, such as influenza, where the effectiveness of the vaccine might be improved by focusing the immune response it triggers.

However, it remains to be seen how well this technique will translate to humans. HIV vaccination is "very difficult" and might have multiple components to help develop the immune response over time, Andrabi explained, adding, "It's not just going to be one or two shots."

However, he said, "they have figured out the first step."

Some people hit their stride in the morning, while others feel most awake and productive at night. But a growing body of evidence suggests that being part of the latter group, the "night owls," could be linked to poorer heart health.

Now, a large study published Jan. 28 in the Journal of the American Heart Association adds weight to the idea.

Pulling from the UK Biobank, a massive repository of data from U.K. adults, researchers followed more than 322,000 adults ages 39 to 74 for nearly 14 years. They examined how "chronotype" — whether someone is a morning or evening person — relates to their cardiovascular disease risk.

The participants reported their preferred sleep-wake schedules and were classified as morning, intermediate or evening types. Heart health was assessed using the American Heart Association's Life's Essential 8 score, which rates heart health on a scale of 0 to 100 based on eight factors: smoking, diet, exercise, blood sugar, body weight, blood pressure, cholesterol and sleep quality. Higher scores indicate better cardiovascular health.

Compared with intermediate sleepers, "definite evening" types were 79% more likely to have poor overall scores and had a 16% higher risk of heart attack or stroke over about 14 years of follow-up. "Definite morning" types were 5% less likely than intermediate sleepers to have poor scores.

But what's behind this link? Does something about night owls' lifestyles set them up for worse heart health, or is there something unique about their biology that plays a role?

How lifestyle may factor in

The night owls scored worse on various lifestyle-related factors, compared to the intermediate sleepers in the biobank study. For example, they had 54% higher nicotine exposure, 42% poorer sleep, 19% lower physical activity, and 8% poorer diet. In contrast, the morning people generally had lower nicotine use and better diets than intermediate types did.

When these differences were taken into account, about 75% of the increased cardiovascular risk observed in night owls appeared to be linked to lifestyle factors, rather than to the chronotype itself. Among these factors, smoking contributed the most (34%), followed by poor sleep (14%), high blood sugar (12%), and diet and body weight (11% each).

"The drivers of overall cardiovascular health are modifiable, including nicotine exposure," study lead author Sina Kianersi, a postdoctoral research fellow at Brigham and Women's Hospital and Harvard Medical School, told Live Science in an email. Other influential lifestyle factors included physical activity levels, blood pressure and cholesterol, he added.

"So for people who identify as night owls, it is a good reason to be even more proactive about those basics," he said.

Prior studies have suggested that night owls often experience social jet lag, a mismatch between their internal clock and their daily routines. This mismatch may be associated with habits like late-night eating, skipped breakfasts, irregular sleep and a greater reliance on caffeine or nicotine. In turn, these behaviors have been linked to factors that raise heart disease risk, such as higher body mass index (BMI), elevated blood sugar, higher insulin resistance, and lower levels of good cholesterol.

Past research also indicates that, even if they get a decent amount of sleep overall, adults with irregular sleep patterns have a 26% higher risk of stroke, heart attack, or heart failure than those with regular sleep patterns.

"Fragmented or insufficient sleep can raise blood pressure, cholesterol, and blood sugar, often with weight gain," Dr. Bibhu Mohanty, an associate professor at the University of South Florida who was not involved in the research, told Live Science in an email. "This is largely driven by [the stress hormone] cortisol, which mobilizes sugar and fat in ways that can make disease harder to reverse, once it is established."

The biology of night owls

Being a night owl isn't a choice — it's partly written in people's genes.

Studies have linked hundreds of genetic variants to chronotype. Many of the highlighted genes are involved in regulating the body's 24-hour clocks, or circadian rhythms. These rhythms help control sleep, appetite, hormone cycles and heart function throughout the day and night.

People with late chronotypes often have delayed circadian rhythms. That means their levels of melatonin, the sleep-promoting hormone, rise later in the evening than they do in early chronotypes. In turn, their cortisol, which helps wake up the body, rises later in the morning than usual.

This shift can throw the body out of sync with the normal day-night cycle, making the body’s fight-or-flight system overactive. As a result, blood pressure can stay high at night, and inflammation and oxidative stress increase. Over time, these changes can damage blood vessels, disrupt blood sugar, and promote plaque buildup in the arteries, which may manifest as heart disease, studies suggest.

Many of the genes involved in determining chronotype also directly affect lipid and glucose metabolism, or how the body breaks down fats and sugars. Studies show that evening types show altered expression of these "clock genes" alongside elevated inflammation markers and triglycerides, a type of fat, compared to morning types. These patterns may contribute to metabolic issues like insulin resistance and poor cholesterol, which are known heart disease risks, though more research is needed to confirm the direct link.

Early evidence also suggests subtle differences in heart structure and function in night owls, though the long-term effects of these differences are still being studied.

Questions to address

Some data from the UK Biobank study suggest the link between night owls and poorer heart health may be stronger in females than in males. Female night owls were 96% more likely, and male night owls 67% more likely, than their intermediate chronotype counterparts to have a poor overall cardiovascular health profile, according to the LE8 score.

Kianersi said, "This difference could reflect social pressures, biology, or both. It's a signal that deserves further study."

Even though the UK Biobank study found a link between the chronotype and poorer heart health, being a night owl doesn't doom you to cardiovascular disease.

"We need studies that get closer to cause and effect, using approaches like genetics as well as intervention research," Kianersi said. "And it will be important to see if these patterns hold in younger adults and in other populations with different work schedules and environments."

For night owls looking to protect their hearts, Mohanty advises focusing on improving lifestyle habits whenever possible.

"That means developing healthier sleep routines, exercising, making healthy dietary choices, avoiding unnecessary snacking and smoking, and potentially using medication to reduce risk if needed," he said. Approaching heart health holistically is "probably the best approach to limit risk."

This article is for informational purposes only and is not meant to offer medical advice.

Transgender women who have undergone hormone therapy show comparable physical fitness to cisgender women, according to the most comprehensive analysis of its kind to date.

The review, published Tuesday (Feb. 3) in the British Journal of Sports Medicine, looked at 52 published studies that assessed the body composition, muscular strength and aerobic capacity of nearly 6,500 individuals, including about 2,900 transgender women and 2,300 transgender men.

But what does that mean for actual performance in women's sports?

While the study counters the claim that trans women have an inherent, unfair advantage due to certain baseline physical traits, it did not look at elite athletes and does not capture all the elements that go into sports performance, experts said.

"Sport is multifactorial," senior study author Bruno Gualano, an associate professor at the University of São Paulo's Center of Lifestyle Medicine, told Live Science in an email. "Training quality, access to facilities, psychological stress and exposure to discrimination all influence performance, and these factors are rarely captured in physiological studies."

But it does argue against a blanket ban on trans women in sports competition, experts told Live Science.

Systematic analysis

The analysis was motivated by recent efforts unfolding around the world to ban transgender people from athletic competitions.

"We've seen increasingly restrictive rules on transgender participation in sport, often justified by claims of large and unavoidable physical advantages," Gualano said.

While circulating testosterone levels seem to increase muscle mass, strength and aerobic capacity, these bans, typically aimed at transgender women and girls, often argue that even past exposure to testosterone during puberty gives individuals a permanent and inherent physical advantage over cisgender women.

To see if that was the case, the researchers pooled data from many studies that employed different approaches and measures to compare physical fitness in transgender and cisgender people. The study participants ranged from 14 to 41 years old, and most were adults.

Among transgender women, hormone therapies included different forms of estrogen and antiandrogens, which suppress the effects of testosterone, while transgender men used various forms of testosterone. Most studies followed participants for about one to three years of therapy.

When normalized for height, "transgender women, after gender-affirming hormone therapy, do not show greater strength or aerobic capacity than cisgender women," Gualano said. That included both upper and lower-body strength.

Transgender women also had comparable fat mass to that of cisgender women.

They have slightly higher lean mass, he noted, but that doesn't translate into greater strength or maximal oxygen consumption. Studies rarely look at metrics of specific athletic performance, so the team could not assess that, Gualano added.

Most of the study participants included in the analysis were not competitive athletes, so "we should be cautious about extrapolating directly to elite sport," Gualano noted. But that said, "if there were large, intrinsic physical advantages, we would expect to observe them even in non-athletic populations, and we do not."

"What's new here is the consistency of these findings across many datasets," Ada Cheung, an endocrinologist and head of the Trans Health Research Group at the University of Melbourne, who was not involved in the study, told Live Science in an email. The results challenge the notion that trans women have intrinsic athletic advantages, she added.

Phoebe Toups Dugas, an associate professor of human-centered computing in the Exertion Games Lab at Monash University in Australia who was not involved in the study, agreed.

"Counter to narratives being used to push transgender athletes out of sport, there's no evidence that transgender women have any kind of advantage," she told Live Science in an email. "There's no reason to ban us from playing."

Olympic competition

While the U.S. Olympic Committee has banned transgender women from competing in women's events to align with an executive order from President Donald Trump, the International Olympic Committee (IOC) has not yet done so. But recent reporting suggests that might change, as the IOC has said it will issue new rules this year. The IOC medical and scientific director has reportedly stated that, even after hormone therapy, transgender women who have gone through male puberty retain physical advantages over cisgender women. She has said the new rules also may apply to cisgender women with "male" characteristics, such as Y chromosomes or "male levels" of testosterone.

The results of the recent meta-analysis do not support those rules, Gualano said. While the new analysis can't tell us anything about the fitness of cisgender women with relatively high testosterone, other datasets suggest that "performance is not determined by testosterone alone," he added.

"A major gap in the literature is our limited understanding of how hormonal effects interact with long-term training and social context, particularly in women and gender diverse populations," Gualano said. "Another gap lies in the assumption that testosterone thresholds neatly separate 'fair' from 'unfair' competition, an idea that is far less scientifically robust than often assumed."

For now, the previous regulations are in place for the 2026 Winter Olympics, which kick off Feb. 6 in Milan. Elis Lundholm of Sweden will compete in mogul skiing as the Winter Games' first openly transgender athlete. As this discussion unfolds, it's important to note that few studies track transgender athletes in high-stakes competitions, in part because there are very few competing.

Given a lack of data on elite sports, an analysis like Gualano's can be very informative, Toups Dugas said.

"These findings are of tremendous value to the IOC," she said. "There is a big opportunity here for the IOC to make the Olympics more diverse and really support athletes."

Youth sports

Most athletes, of course, aren't competing in the Olympics. Instead, many are children playing youth sports. At least 29 U.S. states currently ban transgender youth from competing on teams aligned with their gender identity, and although some of these bans have been legally challenged, they are expected to be upheld by the U.S. Supreme Court in 2026.

Though the new meta-analysis included some data from adolescents, it was not enough to make robust conclusions about kids' body composition, Toups Dugas said. "But I don't think it needs to for us to make sensible choices."

Cheung agreed. "There is no evidence here to justify categorical bans on trans youth in sport."

This article is for informational purposes only and is not meant to offer medical advice.

For close to a century, geoscientists have pondered a mystery: Where did Earth’s lighter elements go? Compared to amounts in the Sun and in some meteorites, Earth has less hydrogen, carbon, nitrogen and sulfur, as well as noble gases like helium — in some cases, more than 99 percent less.

Some of the disparity is explained by losses to the solar system as our planet formed. But researchers have long suspected that something else was going on too.

Recently, a team of scientists reported a possible explanation — that the elements are hiding deep in the solid inner core of Earth. At its super-high pressure — 360 gigapascals, 3.6 million times atmospheric pressure — the iron there behaves strangely, becoming an electride: a little-known form of the metal that can suck up lighter elements.

Study coauthor Duck Young Kim, a solid-state physicist at the Center for High Pressure Science & Technology Advanced Research in Shanghai, says the absorption of these light elements may have happened gradually over a couple of billion years — and may still be going on today. It would explain why the movement of seismic waves traveling through Earth suggests an inner core density that is 5 percent to 8 percent lower than expected were it metal alone.

Electrides, in more ways than one, are having their moment. Not only might they help solve a planetary mystery, they can now be made at room temperature and pressure from an array of elements. And since all electrides contain a source of reactive electrons that are easily donated to other molecules, they make ideal catalysts and other sorts of agents that help to propel challenging reactions.

One electride is already in use to catalyze the production of ammonia, a key component of fertilizer; its Japanese developers claim the process uses 20 percent less energy than traditional ammonia manufacture. Chemists, meanwhile, are discovering new electrides that could lead to cheaper and greener methods of producing pharmaceuticals.

Today’s challenge is to find more of these intriguing materials and to understand the chemical rules that govern when they form.

Electrides at high pressure

Most solids are made from ordered lattices of atoms, but electrides are different. Their lattices have little pockets where electrons sit on their own.

Normal metals have electrons that are not stuck to one atom. These are the outer, or valence, electrons that are free to move between atoms, forming what is often referred to as a delocalized “sea of electrons.” It explains why metals conduct electricity.

The outer electrons of electrides no longer orbit a particular atom either, but they can’t freely move. Instead, they become trapped at sites between atoms that are called non-nuclear attractors. This gives the materials unique properties. In the case of the iron in Earth’s core, the negative electron charges stabilize lighter elements at non-nuclear attractors that were formed at those super-high pressures, 3,000 times that at the bottom of the deepest ocean. The elements would diffuse into the metal, explaining where they disappeared to.

The first metal found to form an electride at high pressure was sodium, reported in 2009. At a pressure of 200 gigapascals (2 million times greater than atmospheric pressure) it transforms from a shiny, reflective, conducting metal into a transparent glassy, insulating material. This finding was “very weird,” says Stefano Racioppi, a computational and theoretical chemist at the University of Cambridge in the United Kingdom, who worked on sodium electrides while in the lab of Eva Zurek at the University at Buffalo in New York state. Early theories, he says, had predicted that at high pressure, sodium’s outer electrons would move even more freely between atoms.

The first sign that things were different came from predictions in the late 1990s, when scientists were using computational simulations to model solids, based on the rules of quantum theory. These rules define the energy levels that electrons can have, and hence the probable range of positions in which they are found in atoms (their atomic orbitals).

Simulating solid sodium showed that at high pressures, as the sodium atoms get squeezed closer together, so do the electrons orbiting each atom. That causes them to experience increasing repulsive forces with one another. This changes the relative energies of every electron orbiting the nucleus of each atom, Racioppi explains — leading to a reorganization of electron positions.

The result? Rather than occupying orbitals that allow them to be delocalized and move between atoms, the orbitals take on a new shape that forces electrons into the non-nuclear attractor sites. Since the electrons are stuck at these sites, the solid loses its metallic properties.

Adding to this theoretical work, Racioppi and Zurek collaborated with researchers at the University of Edinburgh to find experimental evidence for a sodium electride at extreme pressures. Squeezing crystals of sodium between two diamonds, they used X-ray diffraction to map electron density in the metal structure. This, they reported in September 2025, confirmed that electrons really were located in the predicted non-nuclear attractor sites between sodium atoms.

Just the thing for catalysts

Electrides are ideal candidates for catalysts — substances that can speed up and lower the energy needed for chemical reactions. That’s because the isolated electrons at the non-nuclear attractor sites can be donated to make and break bonds. But to be useful, they would need to function at ambient conditions.

Several such stable electrides have been discovered over the last 10 years, made from inorganic compounds or organic molecules containing metal atoms. One of the most significant, mayenite, was found by surprise in 2003 when material scientist Hideo Hosono at the Institute of Science Tokyo was investigating a type of cement.

Mayenite is a calcium aluminate oxide that forms crystals with very small pores — a few nanometers across — called cages, that contain oxygen ions. If a metal vapor of calcium or titanium is passed over it at high temperature, it removes the oxygen, leaving behind just electrons trapped at these sites — an electride.

Unlike the high-pressure metal electrides that switch from conductors to insulators, mayenite starts as an insulator. But now its trapped electrons can jump between cage sites (via a process called quantum tunnelling) — making it a conductor, albeit 100 to 1,000 times less conductive than a metal like aluminum or silver. It also becomes an excellent catalyst, able to surrender electrons to help make and break bonds in reactions.

By 2011, Hosono had begun to develop mayenite as a greener and more efficient catalyst for synthesizing ammonia. Over 170 million metric tons of ammonia, mostly for fertilizers, is produced annually via the Haber-Bosch process, in which metal oxides facilitate hydrogen and nitrogen gases reacting together at high pressure and temperature. It is an energy-intensive, expensive process — Haber-Bosch plants account for some 2 percent of the world’s energy use.

In Haber-Bosch, the catalysts bind the two gases to their surfaces and donate electrons to help break the strong triple bond that holds the two nitrogen atoms together in nitrogen gas, as well as the bonds in hydrogen gas. Because mayenite has a strong electron-donating nature, Hosono thought mayenite would be able to do it better.

In Hosono’s reaction, mayenite itself does not bind the gases but acts as a support bed for nanoparticles of a metal called ruthenium. First, the nanoparticles absorb the nitrogen and hydrogen gases. Then the mayenite donates electrons to the ruthenium. These electrons flow into the nitrogen and hydrogen molecules, making it easier to break their bonds. Ammonia thus forms at a lower temperature — 300 to 400° C — and lower pressure — 50 to 80 atmospheres— than with Haber-Bosch, which takes place at 400 to 500° C and 100 to 400 atmospheres.

In 2017, the company Tsubame BHB was formed to commercialize Hosono’s catalyst, with the first pilot plant opening in 2019, producing 20 metric tons of ammonia per year. The company has since opened a larger facility in Japan and is setting up a 20,000-ton-per year green ammonia plant in Brazil to replace some of the nation’s fossil-fuel-based fertilizer production. The company estimates that this will avoid 11,000 tons of CO₂ emissions annually — about equal to the annual emissions of 2,400 cars.

There are other applications for a mayenite catalyst, says Hosono, including a lower-energy conversion of CO₂ into useful chemicals like methane, methanol or longer-chain hydrocarbons. Other scientists have suggested that mayenite’s cage structure also makes it suitable for immobilizing radioactive isotope waste in nuclear power stations: The electrons could capture negative ions like iodine and bromide and trap them in the cages.

Mayenite has even been studied as a low-temperature propulsion system for satellites in space. When it is heated to 600° C in a vacuum, its trapped electrons blast from the cages, causing propulsion.

Organic electrides

The list of materials known to form electrides keeps growing. In 2024, a team led by chemist Fabrizio Ortu at the University of Leicester in the UK accidentally discovered another room-temperature-stable electride made from calcium ions surrounded by large organic molecules, together known as a coordination complex.

He was using a method known as mechanical chemistry — “You put something in a milling jar, you shake it really hard, and that provides the energy for the reaction,” he says. But to his surprise, electrons from the potassium he had added to his calcium complex were not donated to the calcium ion. Instead, what formed “had these electrons that were floating in the system,” he says, trapped in sites between the two metals.

Unlike mayenite, this electride is not a conductor — its trapped electrons do not jump. But they allow it to facilitate reactions that are otherwise hard to get started, by activating unreactive bonds, doing a job much like a catalyst. These are reactions that currently rely on expensive palladium catalysts.

The scientists successfully used the electride on a reaction that joins two pyridine rings — carbon rings containing a nitrogen atom. They are now examining whether the electride could assist in other common organic reactions, such as substituting a hydrogen atom on a benzene ring. These substitutions are difficult because the bond between the benzene ring carbon and its attached hydrogen is very stable.

There are still problems to sort out: Ortu’s calcium electride is too air- and water-sensitive for use in industry. He is now looking for a more stable alternative, which could prove particularly useful in the pharmaceutical industry to synthesize drug molecules, where the sorts of reactions Ortu has demonstrated are common.

Still questions at the core

There remain many unresolved mysteries about electrides, including whether Earth’s inner core definitely contains one. Kim and his collaborators used simulations of the iron lattice to find evidence for non-nuclear attractor sites, but their interpretation of the results remains “a little bit controversial,” Racioppi says.

Sodium and other metals in Group 1 and Group 2 of the periodic table of elements — such as lithium, calcium and magnesium — have loosely bound outer electrons. This helps make it easy for electrons to shift to non-nuclear attractor sites, forming electrides. But iron exerts more pulling power on its outer electrons, which sit in differently shaped orbitals. This makes the increase in electron repulsion under pressure less significant and thus the shift to electride formation difficult, Racioppi says.

Electrides are still little known and little studied, says computational materials scientist Lee Burton of Tel Aviv University. There is still no theory or model to predict when a material will become one. “Because electrides are not typical chemically, you can’t bring your chemical intuition to it,” he says.

Burton has been searching for rules that might help with predictions and has had some success finding electrides from a screen of 40,000 known materials. He is now using artificial intelligence to find more. “It’s a complex interplay between different properties that sometimes can all depend on each other,” he says. “This is where machine learning can really help.”

The key is having reliable data to train any model. Burton’s team only has actual data from the handful of electride structures experimentally confirmed so far, but they also are using the kind of modeling based on quantum theory that was carried out by Racioppi to create high-resolution simulations of electron density within materials. They are doing this for as many materials as they can; those that are confirmed by real-world experiments will be used to train an AI model to identify more materials that are likely to be electrides — ones with the discrete pockets of high electron density characteristic of trapped electron sites. “The potential,” says Burton, “is enormous.”

This article originally appeared in Knowable Magazine, a nonprofit publication dedicated to making scientific knowledge accessible to all. Sign up for Knowable Magazine’s newsletter.

Satellite images have revealed that dozens of dams across the U.S. — including the biggest one in Texas — may be at risk of collapse due to the ground shifting beneath them. Inspections do not typically account for these movements, suggesting many dams in the country are in worse condition than previously understood.

The new findings raise the prospect that thousands of dams we haven't been monitoring closely due to high costs and staff shortages could be damaged and at risk of failure. But how big is the problem, and is it worth using satellite data to provide early warnings?

Satellites could provide a simple way to monitor dams, potentially revealing places where ground movements are destabilizing their structure, experts told Live Science. However, the new findings also highlight another looming problem: Changing climate conditions could push several of the aging dams in the U.S. to fail, endangering tens of thousands of lives, experts added.

"Identifying issues at dams is critical in preventing failures," John Roche, a dam regulator in Maryland and the president of the Association of State Dam Safety Officials (ASDSO), who did not participate in the new research, told Live Science in an email. "Faced with the threat of increased stress from a changing climate, the lack of timely rehabilitation of dams will lead to increased public safety risks, negative economic impacts, and environmental hazards to citizens of the country."

Shifting ground

In a presentation to the American Geophysical Union in December 2025, scientists used 10 years of radar images from the Sentinel-1 satellite to identify dams that have shifted due to sinking or elevating ground. Depending on the material of the dam, this can lead to cracks forming, especially if different parts of the structure are moving in opposite directions or at varying rates.

"This technology helps us to find potential issues, and then inform the people who are in charge," lead researcher Mohammad Khorrami, a postdoctoral geotechnical engineer at Virginia Tech and The United Nations University Institute for Water, Environment and Health, told Live Science.

The results are based on 41 high-hazard hydroelectric dams that are higher than 50 feet (15 meters), and whose condition is "poor" or "unsatisfactory" under the National Inventory of Dams' classification. These are dams with known defects that compromise the safety of operations and require repairs.

The results are preliminary and have not been peer reviewed. Nevertheless, they show previously unknown weaknesses in dams across 13 U.S. states and Puerto Rico — including Roanoke Rapids Dam in North Carolina and Livingston Dam, the biggest dam in Texas.

Some of these high-risk dams are shifting considerably. For example, the northern portion of Livingston Dam — which feeds two water purification plants supplying more than 3 million people in Houston — is sinking at a rate of about 0.3 inches (8 millimeters) per year, while the southern portion is simultaneously rising by the same amount.

"That doesn't mean that part of the dam is collapsing," Khorrami said. But such elevation differences warrant further investigation, because they might turn out to be a problem, he added. Given that these dams are decades old, potentially faulty and affect both people downstream and energy supplies, deformations in the structure could be disastrous.

A tragic incident in Libya in 2023 suggests land elevation changes are not something to overlook. On Sept. 11, two dams collapsed following extreme rainfall from Storm Daniel. The failures unleashed 1 billion cubic feet (30 million cubic meters) — or 10,000 Olympic swimming pools — of water upon the city of Derna, destroying buildings and bridges, and killing up to 24,000 people.

Deformations in the dams resulting from land elevation changes likely contributed to the collapses, a 2025 study found. "The results of satellite imagery showed a constant and persistent deformation on both these dams during the last decade," Khorrami said. "So those dams were already vulnerable."

Khorrami and his colleagues are finalizing the results of their study. The next step will be to produce an interactive map or database that policymakers can use to assess the safety of U.S. dams.

"It's not a replacement for inspections," Khorrami said. "We're providing another tool to help find early warning signs if there is any issue, or potential issue, with the dam."

Aging infrastructure, changing climate

But ground shifts are just one factor that can compromise dams. The U.S. has almost 92,600 dams — more than 16,700 of which have a "high-hazard potential," meaning that if they collapsed, they could cause loss of human life and significant property destruction, according to ASDSO. Most were designed more than 50 years ago, and around 2,500 show signs of damage that would collectively take billions to fix.

Not all of these are behemoths like the Hoover Dam; in fact, thousands are small watershed dams designed to prevent flooding, provide drinking water and preserve wildlife habitats.

When they were built in the 1960s and 1970s, these dams posed very little risk to people because few lived nearby. But several decades on, communities have mushroomed around them, meaning a failure could be devastating.

What's more, most of these dams were designed to withstand the environmental conditions that existed when they were built, but global warming and land-use changes have altered the picture.

Some rivers are dwindling due to drought, while others have higher water levels and flows than they did 50 to 60 years ago due to increases in rainfall and urbanization, which reduces the amount of water stored in the soil, Ebrahim Ahmadisharaf, an assistant professor of civil and environmental engineering at Florida State University, who was not involved in the research, told Live Science.

Weather is also becoming more extreme and unpredictable, raising the risk of sudden floods, Ahmadisharaf said. In a 2025 study, he and his colleagues found that the likelihood of dam overtopping — when water is so high, it exceeds the capacity of spillways and gushes over the dam — has increased at 33 dams over the past 50 years.

The dams with the highest overtopping probabilities in that study were big dams with relatively large populations living in cities and small towns downstream — including Whitney Dam in Texas, Milford Dam in Kansas and Whiskeytown Dam in California. The population centers that could be impacted include Waco, Texas, with a population of 150,000, and Junction City, Kansas, with 22,000 residents.

"Overtopping is a possible failure mechanism of a dam," Ahmadisharaf explained. "It can lead to catastrophic flooding downstream, and then structural failure. The larger the dam and the shorter the distance to the infrastructure and people downstream, the more dangerous [overtopping is]."

Money problems

One of the biggest hurdles in the way of making dams in the U.S. safer is funding — and the older dams get, the bigger the bill grows.

"Operation, maintenance, and rehabilitation of dams can range in cost from the low thousands to millions of dollars, and responsibility for these expenses lies with owners, many of whom cannot afford these costs," Roche said. "To rehabilitate just the most critical dams was estimated at $37.4 billion, a cost that continues to rise as maintenance, repair, and rehabilitation are delayed."

Rolling out satellite monitoring for dams would increase the financial burden — but it may be worth the cost if it helps prioritize fixes and prevent failures, Roche said. According to a forensic report about the Oroville Dam spillway incident in 2017, which caused more than 180,000 evacuations but no deaths, traditional inspections of dams do not always identify important structural issues.

With only preliminary results available so far, it is hard to tell whether using satellite data to prioritize dam repairs is useful, Roche said. But in theory, "deformation of dam structures may be indicative of a problem or worsening condition," he said.

David Bowles, a dam safety risk expert and professor emeritus of civil and environmental engineering at Utah State University, is more skeptical. "There are many ways a dam can breach," Bowles told Live Science in an email. "Foundation settlement is not a major root cause of dam breach in my experience, but it could be a factor particularly if it is not being monitored and managed."

There might also be a role for satellites in assessing dam overtopping risks, Ahmadisharaf said. Satellite radar images could provide better estimates of water levels and flooding, which in turn could help disseminate warnings earlier.

Overall, satellites could provide a broader overview than what we currently have of risks at dams, Ahmadisharaf said. "We can't monitor everywhere," he said, "but satellites provide this opportunity."

Same-sex sexual behavior among primates could be shaped partly by specific environmental and social conditions, according to a new study comparing 59 species.

Same-sex sexual behavior (SSB) in animals is increasingly recognized in the scientific community as widespread, as it's documented across the animal kingdom in about 1,500 species. Untangling how and why it emerged across so many diverse species is challenging, however. The new study, published Jan. 12 in the journal Nature Ecology & Evolution, found SSB may be beneficial — at least for primates — when ecological conditions are harsh and social conditions are complex.

While the study authors emphasized that this research on SSB in nonhuman primates does not address modern human sexual orientation or identity, they hope the findings will foster open discussion about the benefit of sexual diversity in nature and society.

"[The findings] do suggest that the orientation toward individuals of the same sex has a very strong evolutionary history, and it's nothing bizarre or derived or unnatural," said Durham University primatologist Zanna Clay, who was not involved with the new study. "In fact, it's likely part of the evolutionary fabric of our societies."

Primate same-sex bonding to navigate harsh conditions

Research suggests SSB facilitates bonding in socially complex animals. In bonobos (Pan paniscus) and chimpanzees (Pan troglodytes), it is associated with reduced tension, conflict resolution and the strengthening of alliances. For golden snub-nosed monkeys (Rhinopithecus roxellana), SSB and grooming strengthen social bonds in harsh, cold climates with scarce resources.

Genes also seem to play a role. In a 2023 study of rhesus macaques (Macaca mulatta), Vincent Savolainen, a biologist at Imperial College London, found SSB to be about 6.4% heritable, meaning the tendency toward this behavior can be passed down genetically from parents to offspring. But such a small percentage leaves a lot of uncertainty around what else might cause it to arise.

To explore the ecological and social context, Savolainen and colleagues conducted a meta-analysis of SSB studies across primates. Of the 491 species, they found the behavior documented and prevalent in 59 species. SSB is more likely to occur when species face a drier environment, scarcer resources, and a lot of predators, according to the study. It is also more common in species with complex social systems, greater size differences between males and females, and longer lifespans.

These trends suggest SSB might serve as a social strategy to reinforce bonds, manage conflict or build alliances, driven by the ecological and social pressures a group faces. "Species that have particularly challenging environmental and social pressures have evolved, independently of common ancestry, same-sex sexual behavior as a way to manage the pressure and navigate the social dynamic," Savolainen said, "forming coalitions, bonding, helping them deal with the challenges they face."

If predators are especially abundant, for example, having a socially close group that can trust each other's alarm calls is beneficial, Savolainen said; SSB offers one way to form or maintain relationships.

Chimpanzees and bonobos are known to engage in SSB when navigating ecological challenges, said Clay. "In a resource-poor situation, you need to cooperate and learn to tolerate each other," Clay said. "If there's food scarcity, having techniques to maintain and keep social bonds going is important."

However, while the trends are notable, it's not so simple to draw parallels across species that display SSB. "It points to some common explanations that might cut across deep taxonomic divisions, but there is a risk you're obscuring some of the nuance in individual lineages," said Nathan Bailey, an evolutionary biologist at the University of St Andrews who was not involved with the new study. "Does this behavior emerge for different functional reasons, under different selective pressures, in different lineages? They're starting to scratch the surface of that."

Savolainen said SSB has been traditionally underreported, so the new findings highlight its significance in a broader understanding of primate behavior. "Same-sex behavior is as important as feeding, fighting, or looking after young," he said.

The research could help paint a more complete picture of social and sexual behavior in primates. "People tend to separate reproductive sex and social sex, whereas actually I think the social element of both is very important and should be integrated," Clay said.

But can these findings shed any light on human behavior? Our early hominin ancestors likely experienced various ecological and social pressures, including those linked to SSB for the primates in this study, the study authors pointed out –– but it's unclear whether these pressures would have contributed to the evolution of same-sex sexual orientation in hominin species in a similar way.

Generalizing the results to humans is tricky, according to the researchers and experts, without behavioral data from our hominin ancestors, and considering modern human culture and identity is so complex.

"I don't think this tells us much about what's going on in humans," Bailey said. "There seems to be a huge diversity of explanations [of SSB] across animals, even in closely related lineages, so it doesn't stand to reason to me that any one particular explanation in animals would map onto human beings."

The study also points to a key reason primates, including humans, have succeeded so well across the globe: adaptability. "We're not fixed to one mating system, one behavioral system," Clay said. "To me, the fact that sexual behaviors can expand [under different conditions] reflects that behavioral flexibility which is really important for primate success."

Our personalities as humans are shaped through interaction, reflected through basic survival and reproductive instincts, without any pre-assigned roles or desired computational outcomes. Now, researchers at Japan's University of Electro-Communications have discovered that artificial intelligence (AI) chatbots can do something similar.

The scientists outlined their findings in a study first published Dec. 13, 2024, in the journal Entropy, which was then publicized last month. In the paper, they describe how different topics of conversation prompted AI chatbots to generate responses based on distinct social tendencies and opinion integration processes, for instance, where identical agents diverge in behavior by continuously incorporating social exchanges into their internal memory and responses.

The researchers evaluated individual chatbots' responses to questions using psychological tests and answers to hypothetical scenarios, revealing varied opinions and behavior patterns. They modeled answers on the Maslow hierarchy of human needs (physiological, safety, social, esteem and self-actualization).

Graduate student Masatoshi Fujiyama, the project lead, said the results suggest that programming AI with needs-driven decision-making rather than pre-programmed roles encourages human-like behaviors and personalities.

How such a phenomenon emerges is the cornerstone of the way large language models (LLMs) mimic human personality and communication, said Chetan Jaiswal, professor of computer science at Quinnipiac University in Connecticut.

"It's not really a personality like humans have," he told Live Science when interviewed about the finding. "It's a patterned profile created using training data. Exposure to certain stylistic and social tendencies, tuning fallacies like reward for certain behavior and skewed prompt engineering can readily induce 'personality', and it's easily modifiable and trainable."

Author and computer scientist Peter Norvig, considered one of the preeminent scholars in the field of AI, thinks the training based on Maslow's hierarchy of needs makes sense because of where AI's “knowledge” comes from.

"There's a match to the extent the AI is trained on stories about human interaction, so the ideas of needs are well-expressed in the AI's training data," he responded when asked about the research study.

The future of AI personality

The scientists behind the study suggest the finding has several potential applications, including "modeling social phenomena, training simulations, or even adaptive game characters."

Jaiswal said it could provide a shift away from AI with rigid roles, and towards agents that are more adaptive, motivation-based and realistic. "Any system that works on the principle of adaptability, conversational, cognitive and emotional support, and social or behavioral patterns could benefit. A good example is ElliQ, which provides a companion AI agent robot for the elderly."

But is there a downside to AI generating a personality unprompted? In their recent book "If Everybody Builds It Everybody Dies," (Bodley Head, 2025) Eliezer Yudkowsky and Nate Soares, past and present directors of the Machine Intelligence Research Institute, paint a bleak picture of what would befall us if agentic AI develops a murderous or genocidal personality.

Jaiswal acknowledges this risk. "There is absolutely nothing we can do if such a situation ever happens," he said. "Once a superintelligent AI with misaligned goals is deployed, containment fails and reversal becomes impossible. This scenario does not require consciousness, hatred, or emotion. A genocidal AI would act that way because humans are obstacles to its objective, or resources to be removed, or sources of shutdown risk."

So far, AIs like ChatGPT or Microsoft CoPilot only generate or summarize text and pictures — they don't control air traffic, military weapons or electricity grids. In a world where personality can emerge spontaneously in AI, are those the systems we should be keeping an eye on?

"Development is continuing in autonomous agentic AI where each agent does a small, trivial task autonomously like finding empty seats in a flight," Jaiswal said. "If many such agents are connected and trained on data based on intelligence, deception or human manipulation, it's not hard to fathom that such a network could provide a very dangerous automated tool in the wrong hands."

Even then, Norvig reminds us that an AI with villainous intent need not even control high-impact systems directly. "A chatbot could convince a person to do a bad thing, particularly someone in a fragile emotional state," he said.

Putting up defences

If AI is going to develop personalities unaided and unprompted, how will we ensure the benefits are benign and prevent misuse? Norvig thinks we need to approach the possibility no differently than we do other AI development.

"Regardless of this specific finding, we need to clearly define safety objectives, do internal and red team testing, annotate or recognize harmful content, assure privacy, security, provenance and good governance of data and models, continuously monitor and have a fast feedback loop to fix problems," he said.

Even then, as AI gets better at speaking to us the way we speak to each other — ie, with distinct personalities — it might present its own issues. People are already rejecting human relationships (including romantic love) in favour of AI, and if our chatbots evolve to become even more human-like, it may prompt users to be more accepting of what they say and less critical of hallucinations and errors — a phenomenon that's already been reported.

For now, the scientists will look further into how shared topics of conversation emerge and how population-level personalities evolve over time — insights they believe could deepen our understanding of human social behavior and improve AI agents overall.

The online admonition to "touch grass" to soothe your emotional state may be backed by science — at least in lab mice.

A recent study finds that mice that live outside are less anxious than those that spend their days in safe, shoebox-sized cages. And that may highlight a fundamental flaw in laboratory research, including that used to test the safety and effectiveness of drugs eventually intended for people.

Medications that seem to work in lab mice don't necessarily work in human patients, and some scientists think that they might fail, in part, because of how bizarre and isolated laboratory life is for the rodents.

"Why is there that huge gap in results between the animal models in the labs and the real-life experiences when we test [many] drugs in humans?" said first study author Matthew Zipple, a postdoctoral researcher at Cornell University "We think much of this effect may be explained by this really artificial, standardized environment in which lab animals are kept."

The findings were published in December in the journal Current Biology.

Less anxious in the outdoors

Both wild mice and humans have rich social environments, and wild mice are constantly on the go, foraging, burrowing and facing risks, including the many predators that like to snack on them.

In comparison, lab mice sit in small cages with two or three same-sex siblings. There, food and water are delivered on a regular schedule. Studying medications in those mice may be akin to limiting research to prisoners in solitary confinement, Zipple told Live Science.

Zipple and his colleagues set out to compare the psychology of two groups of lab mice: a group that remained in a laboratory and a group that lived with other mice in an outdoor enclosure, complete with grass, dirt and exposure to the sky. They did so using a standard maze, called the "elevated plus maze," which has two enclosed arms and two open, catwalk-style arms.

On their first exposure to this maze under bright lab lights, lab mice typically explore the open arms, find them terrifying, and basically never venture out on them again. Instead, they remain in the comparatively safe, enclosed portion of the maze. This reaction is so consistent that researchers use the open arms to induce and measure anxiety in lab mice.

But mice living in a wild-type environment weren't freaked out by the open arms at all, Zipple and his team found. They spent just as much time exploring these areas on subsequent visits to the maze as they had the first time, all while under bright light.

Meanwhile, cage-dwelling mice that were sent to live outside also saw their maze anxiety evaporate; animals that already had demonstrated an apparent fear of the open arms and then spent a week outside subsequently spent twice as much time exploring the open arms compared with animals that kept living in cages.

The use of the standardized maze was a "very powerful way to show the limits of business as usual," said Andrea Graham, an evolutionary ecologist at Princeton University who was not involved in the research.

Caged mice have other key differences

Graham's lab has shown that mice that live in lab cages are also immunologically different from mice who live outside and encounter dirt, plants and large numbers of other mice. That matters, she said.

In one famous 2006 case, a medication called TGN1412 seemed to boost the immune system against leukemia in lab mice but caused a near-fatal immune reaction in the first six healthy human volunteers exposed to the drug. Subsequent research revealed that, in the lab mice, the medication activated immune cells that regulate and calm the immune response. However, in mice living in wild-type enclosures, the medication instead activated cells that ramp up the immune response to the point that the body attacked itself.

"If we restrict ourselves to only studying a couple of different genotypes [genetic profiles] of lab mouse in the same immunologically boring, psychologically boring environments, we're not going to really be able to study the full spectrum of human immune or nervous system response to the environment," Graham told Live Science.

Using wild-style enclosures requires some upfront cost and effort, and it also reduces the rigid control that's placed on study animals in order to limit confounding variables in experiments. As such, they pull biomedical scientists out of their comfort zone, Zipple said.

But adding in tests of these less-confined mice could save a lot of effort and money on the human trials side by pinpointing the medications that are most likely to translate from the lab to the clinic, the study authors argue. Zipple and his colleagues are now looking at ways that caged and wild-living mice age differently.

"The broader goal is to make a list of biomedically relevant behaviors, phenotypes [observable traits] and psychological traits that look the same in the lab and the field," he said, to help with the issue of translating results to humans. They also want to compile a "list of traits that look quite different," he said.

A 30-minute stroll across New York's Central Park separates Trump Tower from the American Museum of Natural History. If the US president ever found himself inside the museum he could see the Cape York meteorite: a 58-tonne mass of iron taken from northwest Greenland and sold in 1897 by the explorer Robert Peary, with the help of local Inuit guides.

For centuries before Danish colonisation, the people of Greenland had used fragments of the meteorite to make tools and hunting equipment. Peary removed that resource from local control, ultimately selling the meteorite for an amount equivalent to just US$1.5 million today. It was a transaction as one-sided as anything the president may now be contemplating.

But Donald Trump is now eyeing a prize much larger than a meteorite. His advocacy of the US taking control of Greenland, possibly by force, signals a shift from dealmaking to dominance. The scientific cost would be severe. A unilateral US takeover threatens to disrupt the open scientific collaboration that is helping us understand the threat of global sea-level rise.

Greenland is sovereign in everything other than defence and foreign policy, but by being part of the Kingdom of Denmark, it is included within Nato. As with any nation, access to its land and coastal waters is tightly controlled through permits that specify where work may take place and what activities are allowed.

Over many decades, Greenland has granted international scientists access to help unlock the environmental secrets preserved within its ice, rocks and seabed. US researchers have been among the main beneficiaries, drilling deep into the ice to explain the historic link between carbon dioxide and temperatures, or flying repeated Nasa missions to map the land beneath the ice sheet.

The whole world owes a huge debt of thanks to both Greenland and the US, very often in collaboration with other nations, for this scientific progress conducted openly and fairly. It is essential that such work continues.

The climate science at stake

Research shows that around 80% of Greenland is covered by a colossal ice sheet which, if fully melted, would raise sea level globally by about 7 metres (the height of a two storey house). That ice is melting at an accelerating rate as the world warms, releasing vast amounts of freshwater into the North Atlantic, potentially disrupting the ocean circulation that moderates the climate across the northern hemisphere.

The remaining 20% of Greenland is still roughly the size of Germany. Geological surveys have revealed a wealth of minerals, but economics dictates that these will most likely be used to power the green transition rather than prolong the fossil fuel era.

While coal deposits exist, they are currently too expensive to extract and sell, and no major oil fields have been discovered. Instead, the commercial focus is on "critical minerals": high-value materials used in renewable technologies from wind turbines to electric car batteries. Greenland therefore holds both scientific knowledge and materials that can help guide us away from climate disaster.

Unilateral control could threaten climate science

Trump has shown little interest in climate action, however. Having already started to withdraw the US from the Paris climate agreement for a second time, he announced in January 2026 the country would also leave the Intergovernmental Panel on Climate Change, or IPCC, the global scientific body that assesses the impacts of continued fossil-fuel burning. His rhetoric to date has been about acquiring Greenland for "security" purposes, with some indications of accessing its mineral wealth, but without mention of vital climate research.

Under the 1951 Greenland defence agreement with Denmark, the US already has a remote military base at Pituffik in northern Greenland, now focused on space activities. While both countries remain in Nato, the agreement already allows the US to expand its military presence if required. Seeking to guarantee US security in Greenland outside Nato would undermine the existing pact, while a unilateral takeover would risk scientists in the rest of the world losing access to one of the most important climate research sites.

Lessons from Antarctica and Svalbard

Greenland's sovereign status and its governance is different to some other notable polar research locations. For example, Antarctica has, for more than 60 years, been governed through an international treaty ensuring the continent remains a place of peace and science, and protecting it from mining and other environmental damage.

Svalbard, on the other hand, has Norwegian sovereignty courtesy of the 1920 Svalbard treaty but operates a largely visafree system that allows citizens of nearly 50 countries to live and work on the archipelago, as long as they abide by Norwegian law. Interestingly, Norway claims that scientific activities are not covered by the treaty, to almost universal disagreement among other parties. Russia has a permanent station at Barentsburg, Svalbard's second-largest settlement, from which small levels of coal are mined.

Unlike Antarctica or Svalbard, Greenland has no treaty that explicitly protects access for international scientists. Its openness to research therefore depends not on international law, but on Greenland's continued political stability and openness – all of which may be threatened by US control.

If it is minded to take a radical approach, Greenland could develop its own treaty-style approach with selected partner states through Nato, enabling security cooperation, mineral assessment and scientific research to be carried out collaboratively under Greenlandic regulations.

The future for Greenland should lie with Greenlanders and with Denmark. The future of climate science, and the transition to a safe prosperous future worldwide, relies on continued access to the island on terms set by the people that live there. The Cape York meteorite – taken from a site just 60 miles away from the US Pituffik Space Base – is a reminder of how easily that control can be lost.

This edited article is republished from The Conversation under a Creative Commons license. Read the original article.

On Dec. 16 2025, the Trump administration announced plans to dismantle the National Center for Atmospheric Research (NCAR) — one of the world's leading weather and climate research centers. In a statement to USA Today, Russell Vought, director of the White House Office of Management and Budget, said NCAR is "one of the largest sources of climate alarmism in the country."

But dismantling NCAR could seriously impact the nation's capacity to ensure public safety and its ability to protect economic stability, Holly Gilbert, the institute's interim deputy director, told Live Science. "The research we do here directly applies to protecting the public," she said.

The National Science Foundation (NSF) established NCAR in 1960 to shield human life and property from atmospheric and solar hazards. At its core, the center's research translates vast atmospheric data from satellites, aircraft and ground sensors into forecasts that can potentially save lives.

For example, taking a flight used to be a riskier endeavor. Pilots encounter an abundance of challenges — from turbulence and tornadoes to ice on wings and rapidly varying pockets of wind known as microbursts. But thanks in no small part to the research conducted at the NCAR, flying has become safer, helping pilots avoid these hazards: There hasn't been a single crash on a commercial airline caused by sudden, sharp changes in wind speed in more than 25 years.

But the center's impacts stretch well beyond aviation safety. NCAR models help forecast tornadoes, hurricanes, floods, extreme heat, drought and the potential for wildfires. "Whenever there has been an early warning, NCAR science and models have been part of bringing that warning to the people who need to hear it," Daniel Swain, a climate scientist at University of California Agriculture and Natural Resources, told Live Science.

Swain, a research partner of NCAR, also stressed that without these early warnings, people could lose their properties or die unnecessarily in weather-related disasters.

NCAR also partners with insurance and reinsurance industries to assess the risks these hazards pose to lives and property. "NCAR provides leading weather research that helps keep people safe and protect homes, businesses, and jobs," Nicole Austin, senior vice president and director of federal affairs at the Reinsurance Association of America, told Live Science in an email. "Its long-term studies of hail and wildfires, along with real-time weather data, help reduce damage and help communities recover faster after disasters."

One center, multiple impacts

NCAR's models currently save the national aviation system an estimated $27 million annually. For terrestrial disaster resilience, NCAR's WRF-Hydro system powers the National Water Model (NWM), which monitors flood risk for more than 2.7 million stream locations.

NCAR has also developed GPS dropsondes—packages of instruments, tethered to little parachutes, that measure atmospheric conditions. These dropsondes reduced hurricane track forecast errors by up to 30%. Modern economic analyses estimate that such improvements in forecast accuracy save the nation up to $2 billion per major storm by optimizing protective decisions and reducing unnecessary evacuations.

NCAR also engages with private entities such as The Weather Company, which operates key digital consumer brands like The Weather Channel and Weather Underground. "Our proprietary forecasting capabilities — which are vital to national safety and security, economic resiliency, and aviation efficiency — greatly benefit from the basic and applied science and technology produced at NCAR and widely used across the entire weather enterprise, including directly at The Weather Company," Peter Neilley, senior vice president of science and forecasting operations at The Weather Company, told Live Science in an email.

NCAR's models also forecast phenomena beyond the weather. Its Wildland Fire model simulates how wildfires create their own weather, giving first responders critical on-the-ground intelligence, while its Fire INventory from NCAR (FINN) model tracks toxic smoke transport, alerting health officials to air quality hazards thousands of miles downwind.

NCAR's High Altitude Observatory (HAO) also anticipates threats to U.S. infrastructure from events originating in space, such as geomagnetic storms. Predicting space weather is essential for safeguarding the nation's power grid, GPS satellites and communication networks from solar flares that can trigger widespread blackouts and equipment failure.

Greater than the sum of its parts

A statement from the NSF said it "is reviewing the structure of the research and observational capabilities" of NCAR and will "explore options to transfer stewardship" of several of its strategic projects "to concentrate on needs such as seasonal weather prediction, severe storms, and space weather."

But the unique value of NCAR lies in its capacity to treat Earth as a single system, where all the different components — the atmosphere, the oceans, the land and water — are connected to one another, Gilbert said. "Risks to public health and safety are inherently complex and multidisciplinary," she said, "which makes it important to have the expertise and capabilities in one organization to look at where the Earth systems come together."

Modeling the path of a catastrophic wildfire, for example, takes more than data around the fuel available and local wind patterns; it requires simultaneously tracking how heat from the fire creates its own violent, localized winds and determining how multi-year drought has desiccated the fuels. Understanding the public health impact also requires modeling how the resulting toxic smoke will travel. NCAR's expertise is unique because it ensures that the intertwined components of fire, air temperature, dry land, water availability and air quality are modeled simultaneously using shared tools and computational resources, Gilbert said.

The complexity and interconnectedness of these types of disasters means that splitting the center into multiple distributed research locations could have serious consequences, Swain said.

Different areas of research at NCAR used to be structurally siloed, but they've evolved to be more integrated and better suited to addressing complex environmental challenges, Swain said. To better understand weather, you have to understand it in the context of climate, he said. "Otherwise it is like being an emergency room doctor without access to 20% to 40% of a patient's health records."

Returning to siloing would be akin to disassembling the central operating system from a custom, high-performance car and distributing the components to three different mechanics. Each part might function individually, but the dismantling process would impact the ability to drive the vehicle at high speeds safely, Gilbert said.

In a similar vein, splitting up NCAR would hinder the capacity to predict highly complex, multi-hazard events.

"Looking at individual parts is not nearly as important as looking at this integrated system," Gilbert said. NCAR researchers look at the "whole" using established observation, modeling and computational resources and drawing on multiple partnerships established with industries and universities. "It would be so costly to pull these pieces apart," Gilbert added.

Genetic variants believed to cause blindness in nearly everyone who carries them actually lead to vision loss less than 30% of the time, new research finds.

The study challenges the concept of Mendelian diseases, or diseases and disorders attributed to a single genetic mutation. The idea is that Mendelian diseases — such as the neurological disease Huntington's and the bleeding disorder hemophilia — are passed down in predictable ways in families, and if a given person carries a disease-causing mutation, they will have it.

These diseases stand in contrast to those caused by multiple genes and environmental factors, which makes their occurence harder to predict in family lines.

"What we suggest is that there is overlap there," senior study author Dr. Eric Pierce, director of the Ocular Genomics Institute at Mass Eye and Ear and an ophthalmologist at Harvard Medical School, told Live Science. In other words, many diseases thought to have simple, Mendelian causes might be a lot more complex than previously thought.

And this doesn't only apply to inherited blindness. Similar results have been found for other genes once thought to be strongly linked to health conditions. A 2023 study on ovarian insufficiency, a condition that causes infertility and early menopause, found that 99.9% of supposedly disease-causing variants were actually present in healthy women. And certain kinds of inherited diabetes also have more complex genetics than previously believed, according to 2022 research.

"We're in an era of discovering a lot more about the complexity of our genomes," said Anna Murray, a geneticist at the University of Exeter who led the ovarian insufficiency research.

Simple or complex?

Pierce and his colleagues focused on inherited retinal disorders (IRDs), a group of diseases that cause significant vision loss, sometimes as early as age 10 but certainly by age 40, said study co-author Dr. Elizabeth Rossin, a vitreoretinal surgeon and scientist in Mass Eye and Ear’s Retina Service and a Harvard ophthalmologist. Researchers have teased out the genetic roots of these diseases by doing genetic testing on affected patients and their families.

But that method can lead to a problem called ascertainment bias, Pierce said. True, you'll learn that some genetic variants are associated with the disease. But because you're studying only people with the disease and their relatives, you don't get a clear notion of how many people have the same gene variants and don't go blind.

To widen their view, the researchers used data from two large biobanks that contain genetic sequencing data from people, as well as their medical diagnoses and demographic information. One, the All of Us biobank, is a program run by the National Institutes of Health and included nearly 318,000 individuals with both genetic and electronic health record data at the time of the study. The other, the UK Biobank, is comparatively less diverse but contains data from 500,000 individuals, including about 100,000 with images of their retinas submitted to the database.

The researchers picked the 167 genetic variants thought to have the strongest causal link to IRDs and searched for them in the All of Us database. They then used the health record data to see if the people with the variants had vision loss. To their surprise, depending on which diagnostic codes they used, only 9.4% to 28.1% of people with the variants had any indication of a retinal disorder or vision problems.

"You would expect, given what we know about these diseases, that nearly 100% of the people would have blindness," Rossin told Live Science. "But it was far fewer than that."

To validate their findings, the researchers turned to the UK Biobank, this time using the included retinal imagery to seek out evidence of IRDs themselves. They found that only between 16.1% and 27.9% carriers of the gene variants had indications of possible retinal disease.

People who were older who carried these retinal disease genes weren't any likelier to have gone blind. And there was no other evidence that their results were because they were catching people who might later lose their vision. Instead, Pierce says, it seems that the complexity of these presumed Mendelian diseases has been underestimated.

"The mutation we used to think caused disease 100% of the time doesn't exist in isolation," he said. Instead, people carry tens or hundreds of thousands of other genes, some of which may protect against retinal disease, he added.

New avenues for treatment

In theory, those protective gene variants could lead to ways to treat these retinal disorders.

"It's going to take a lot of data in order to find these types of low-effect variants," Pierce said. "There are likely many of them, each contributing a little bit to the protection against disease."

There are good reasons to study the genes of patients with particular disorders, Murray said. For instance, finding genes associated with a condition — even if they don't always cause it — can help researchers pinpoint the biology underlying the disease. In ovarian insufficiency, these kinds of patient-centered studies have shown that genes associated with DNA repair are important for the disorder. But such studies should still be taken with a grain of salt.

"It is only now that we have the ability to look at the granular detail of the genetic sequence in hundreds of thousands of people," she said. To learn more, these databases need to become more diverse, she added. And at the same time, she added, biomedical researchers need better lab models of diseases in which to test certain gene mutations and their effects.

"There are likely some [diseases] where it really is a one-to-one correspondence," Pierce said. "But my prediction would be [that] the majority of these disorders are going to share this new complexity."

The new findings appeared Jan. 8 in the American Journal of Human Genetics.

Editor's note: This story was updated on Jan. 16, 2026, to add mention of Dr. Eric Pierce's and Dr. Elizabeth Rossin's affiliations with Mass Eye and Ear, where the work was conducted.

People and institutions are grappling with the consequences of AI-written text. Teachers want to know whether students’ work reflects their own understanding; consumers want to know whether an advertisement was written by a human or a machine.

Writing rules to govern the use of AI-generated content is relatively easy. Enforcing them depends on something much harder: reliably detecting whether a piece of text was generated by artificial intelligence.

Some studies have investigated whether humans can detect AI-generated text. For example, people who themselves use AI writing tools heavily have been shown to accurately detect AI-written text. A panel of human evaluators can even outperform automated tools in a controlled setting. However, such expertise is not widespread, and individual judgment can be inconsistent. Institutions that need consistency at a large scale therefore turn to automated AI text detectors.

The problem of AI text detection

The basic workflow behind AI text detection is easy to describe. Start with a piece of text whose origin you want to determine. Then apply a detection tool, often an AI system itself, that analyzes the text and produces a score, usually expressed as a probability, indicating how likely the text is to have been AI-generated. Use the score to inform downstream decisions, such as whether to impose a penalty for violating a rule.

This simple description, however, hides a great deal of complexity. It glosses over a number of background assumptions that need to be made explicit. Do you know which AI tools might have plausibly been used to generate the text? What kind of access do you have to these tools? Can you run them yourself, or inspect their inner workings? How much text do you have? Do you have a single text or a collection of writings gathered over time? What AI detection tools can and cannot tell you depends critically on the answers to questions like these.

There is one additional detail that is especially important: Did the AI system that generated the text deliberately embed markers to make later detection easier?

These indicators are known as watermarks. Watermarked text looks like ordinary text, but the markers are embedded in subtle ways that do not reveal themselves to casual inspection. Someone with the right key can later check for the presence of these markers and verify that the text came from a watermarked AI-generated source. This approach, however, relies on cooperation from AI vendors and is not always available.

How AI text detection tools work

One obvious approach is to use AI itself to detect AI-written text. The idea is straightforward. Start by collecting a large corpus, meaning collection of writing, of examples labeled as human-written or AI-generated, then train a model to distinguish between the two. In effect, AI text detection is treated as a standard classification problem, similar in spirit to spam filtering. Once trained, the detector examines new text and predicts whether it more closely resembles the AI-generated examples or the human-written ones it has seen before.

The learned-detector approach can work even if you know little about which AI tools might have generated the text. The main requirement is that the training corpus be diverse enough to include outputs from a wide range of AI systems.

But if you do have access to the AI tools you are concerned about, a different approach becomes possible. This second strategy does not rely on collecting large labeled datasets or training a separate detector. Instead, it looks for statistical signals in the text, often in relation to how specific AI models generate language, to assess whether the text is likely to be AI-generated. For example, some methods examine the probability that an AI model assigns to a piece of text. If the model assigns an unusually high probability to the exact sequence of words, this can be a signal that the text was, in fact, generated by that model.

Finally, in the case of text that is generated by an AI system that embeds a watermark, the problem shifts from detection to verification. Using a secret key provided by the AI vendor, a verification tool can assess whether the text is consistent with having been generated by a watermarked system. This approach relies on information that is not available from the text alone, rather than on inferences drawn from the text itself.

Limitations of detection tools

Each family of tools comes with its own limitations, making it difficult to declare a clear winner. Learning-based detectors, for example, are sensitive to how closely new text resembles the data they were trained on. Their accuracy drops when the text differs substantially from the training corpus, which can quickly become outdated as new AI models are released. Continually curating fresh data and retraining detectors is costly, and detectors inevitably lag behind the systems they are meant to identify.

Statistical tests face a different set of constraints. Many rely on assumptions about how specific AI models generate text, or on access to those models’ probability distributions. When models are proprietary, frequently updated or simply unknown, these assumptions break down. As a result, methods that work well in controlled settings can become unreliable or inapplicable in the real world.

Watermarking shifts the problem from detection to verification, but it introduces its own dependencies. It relies on cooperation from AI vendors and applies only to text generated with watermarking enabled.

More broadly, AI text detection is part of an escalating arms race. Detection tools must be publicly available to be useful, but that same transparency enables evasion. As AI text generators grow more capable and evasion techniques more sophisticated, detectors are unlikely to gain a lasting upper hand.

Hard reality

The problem of AI text detection is simple to state but hard to solve reliably. Institutions with rules governing the use of AI-written text cannot rely on detection tools alone for enforcement.

As society adapts to generative AI, we are likely to refine norms around acceptable use of AI-generated text and improve detection techniques. But ultimately, we’ll have to learn to live with the fact that such tools will never be perfect.

This edited article is republished from The Conversation under a Creative Commons license. Read the original article.