Your AI-generated image of a cat riding a banana exists because of children clawing through the dirt for toxic elements. Is it really worth it?

Behind the output of Large Language Models like Chat GPT lies a journey with complex environmental and social impacts, from the extraction of minerals by children in the Democratic Republic of the Congo, to training systems that expose people to violent and degrading imagery in countries like Nigeria, and to vast, resource-guzzling data centers in regions where energy, water and access to transmission infrastructure is cheap. This means that the AI boom has the potential to create new resource production and consumption economies — likely in communities which are already marginalized or have been subject to previous resource booms and busts.

Yet, these costs are rarely recognized and they raise profound questions about sustainability, not just from a mineral resource point of view, but also in the broader, moral sense — do we want to build a society that profits off the suffering of the world's most marginalized? Will this end up fracturing societies and lead to the politics of resentment?

Akhil Bhardwaj

Akhil Bhardwaj is an Associate Professor of Strategy and Organization at the University of Bath, UK. He studies extreme events, which range from organizational disasters to radical innovation. 

Grete Gansauer

Dr. Grete Gansauer is an Assistant Professor  in the Haub School of Environment and Natural Resources at the University of Wyoming. She is an economic geographer and interdisciplinary public policy researcher focused on regional policy and effects of sustainability transitions in rural and natural-resource producing contexts.

The journey that powers the texts and images produced by AI begins with rare earth minerals used in computer chips. Rare earth minerals are "rare" because they are found in small isolated pockets in Earth’s crust and are difficult to extract through physical and chemical processes.

China currently dominates global rare earth production in mining and processing; the U.S. is second for mining, but it lacks the infrastructure to process rare earths once they are out of the ground.

Many critical minerals, such as lithium and cobalt, are also important for AI processing and storage. Unlike rare earths that are designated due to their chemical properties, the critical minerals designation is a political one assigned to minerals of key strategic, geopolitical, or national security importance.

Many of these minerals are found in regions that are currently ravaged by war (for example, Ukraine has some of Europe’s largest stores of lithium and Russia, is the world’s largest producer of uranium. Others, such as cobalt, are in regions like the Congo, where many of the mines are controlled by Chinese interests.

Setting aside the geopolitical concerns — although these are certainly very important — concerns also arise about labor practices. Many of these mines use artisan mining, which often is a euphemism for child labor — artisan mining can involve children digging for minerals with their hands. These minerals are then mixed with those extracted from industrial mining, making traceability impossible. Labor conditions can be horrific, with high death rates, often due to exposure to air and water pollutants that cause terminal illness.

Consequently, intensified resource production driven by the demands of AI and a highly digitalized economy could produce a new "resource curse" in peripheries of the Global North and the Global South. Wealth produced by local labor is extracted and used to prop up some of the most lucrative digital services industries in the world. The trap, then, is that communities whose material contributions are integrated into AI’s Global Value Chain will be again vulnerable to the same boom-and-bust dynamics that plague economies built on the production or extraction of other resources, like oil or diamonds..

Beyond the extraction of mineral wealth, many AI models require considerable training — and humans need to do that training. LLMs are trained on an increasingly large corpus of "tagged" data that contains violent and pornographic content. Setting aside the precariousness of gig work, the content itself can be highly disturbing and can result in traumatizing workers. Much of this work is carried out in countries like Nigeria and India where the cost of labor is low and where workers have little protection.

Once these models are trained, running them involves using massive data centers to cool the servers that process them. These server farms/data centers consume tremendous resources — both energy and water. Such centers are an emerging business frontier with major implications for land use change and resource impacts.

Private landholding firms are rapidly seeking resource frontiers with the most affordable mix of cheap land, cheap water, cheap energy, and cheap access to transmission infrastructure, proximity to population-dense centers and cheap-but-skilled labor. However, such a geographic unicorn is difficult to find.

Many data centers are located in or being prospected in water-scarce regions such as Nevada and Arizona, where labor and land is cheap. This trend appears to holds true globally. In addition to cheap land, deserts have low humidity, thus reducing the likelihood of metal corrosion. These centers also challenge the capacity of local electricity grids, and because energy is often purchased wholesale or "premarket," can drive rates up for the average consumer. Researchers have estimated that using AI to write an email consumes half a liter of water.

While there is a massive push to embrace the use of LLMs worldwide, especially with the lure of economic and labor efficiency gains and other potential benefits — including employing it for seeking information and writing as well as automating repetitive tasks, we must be fully aware of the material and social costs it imposes. Do you need ChatGPT to write that email? Do you really need to generate an image of a cat riding a banana?

Regardless of how we might answer these questions, it does appear that we need to fundamentally re-asses what it means to be sustainable — claiming to be sustainable while embracing and promoting LLMs is, to say the very least, suspect.

And do we really want the progress LLMs can bring if it's built on the suffering of others? This is a question that we, as a society, need to answer urgently.

Opinion on Live Science gives you insight on the most important issues in science that affect you and the world around you today, written by experts and leading scientists in their field.