New AI system can 'predict human behavior in any situation' with unprecedented degree of accuracy, scientists say

A new artificial intelligence (AI) model can predict and simulate human thought and behavior with a surprising degree of accuracy. The language model, called Centaur, could help researchers improve our understanding of human cognition.

The model was trained on more than 10 million real decisions made by participants of psychological experiments. Using this dataset, Centaur predicted and simulated how people would think and behave with 64% accuracy, researchers reported July 2 in the journal Nature.

Previous AI models could predict some human behavior, but not to this degree. "Centaur is a step beyond in its predictive capabilities," Brenden Lake, who studies psychology and data science at New York University and was not involved in the study, told Live Science in an email.

What is Centaur?

Centaur is an AI model of human cognition that was trained on a curated dataset called Psych-101, which included data from 60,000 people who made more than 10 million individual choices over 160 psychology experiments, according to the paper. The researchers say it might be the world's largest human behavior dataset.

"Essentially, we show the model a full transcript of a psychological experiment from a participant — everything they were told, have seen and have done," study lead author Marcel Binz, a research scientist at the Helmholtz Institute for Human-Centered AI in Germany, told Live Science in an email.

The team then let the model predict what experiment participants chose in specific contexts. If the model predicted a choice that a person didn't make in real life, the researchers would fine-tune the model by correcting its choice. They repeated this process over and over, until Centaur was regularly making correct predictions.

Related: Scientists discover major differences in how humans and AI 'think' — and the implications could be significant

When tested against several established AI models of human thought, Centaur predicted behavior more accurately every time, according to the study. The model is unique because it can anticipate human choices in circumstances that it had never encountered during training, according to a statement from the Helmholtz Association of German Research Centres.

Notably, Centaur can also adapt to changing situations and predict human reaction times.

"We've created a tool that allows us to predict human behavior in any situation described in natural language — like a virtual laboratory," Binz said in the statement.

Studying the brain "at an unprecedented scale"

Binz and his team plan to continue improving the new model. They hope to expand Psych-101 to include demographic and psychological data, including age, socioeconomic status and personality traits, to strengthen its training capabilities. These details would allow Centaur to predict behavior based on an individual's characteristics.

Their next goal is to use Centaur as a proxy for the human brain to determine whether certain patterns they see in the program's data processing correlate with specific decisions. This information could help answer questions about how people process information and how decision-making differs between those with and without mental health conditions, for example, according to the statement.

"Right now, we essentially have a big black box model that predicts human behavior really well," Binz told Live Science. Thanks to Centaur, the researchers can now predict how humans will act, but don't yet have insight into how they arrive at a given decision.

Currently, one big unknown is whether Centaur just predicts human thought or whether it reproduces the underlying cognitive processes, Lake explained. "Is it modeling human mental processes, or just mimicking the outcomes?"

Centaur opens the door to many improvements in psychology and health research. Applications could include predictions of human behavior in clinical contexts or prototyping and analysis of psychology experiments. Scientists could, for example, use the model to figure out how to design an experiment that demonstrates a phenomenon most clearly or requires the fewest participants.

Lake is most excited about the possible educational applications. "Long term, if we can predict how a student learns and reasons through a problem, then we could simulate the effects of different teaching strategies — a potential game changer," he said.

Brain quiz: Test your knowledge of the most complex organ in the body