Artificial intelligence is often presented as something close to magic: a system that writes, draws and reasons on its own. Behind that image sits a workforce of millions whose names rarely appear in product announcements, and a physical supply chain that the word “cloud” does a good job of hiding.
Models are built on human labor
AI models do not become capable by themselves. They are trained on billions of images, videos, sentences and audio clips, and much of that material has to be labeled, ranked or filtered by people. Data workers around the world tag images, grade chatbot responses and, in the hardest version of the job, review violent, hateful and sexually abusive content so that it can be kept out of the systems users interact with.
Much of this work is outsourced to countries with lower wages and weaker labor protections. Workers in Kenya, India, the Philippines, Venezuela, Brazil and elsewhere have described doing this work for a few dollars an hour on behalf of some of the most valuable companies in the world.
The toll on workers
The best documented case involves Kenyan workers employed through the outsourcing firm Sama to build content filters for OpenAI. According to reporting by Popular Science and others, workers were paid roughly $1.32 to $2 an hour to read and label descriptions of child sexual abuse, torture, murder and self-harm. Several later described depression, insomnia, anxiety and symptoms consistent with post-traumatic stress disorder.
In 2023, a group of these workers petitioned Kenya’s parliament to investigate conditions in the sector, saying they had been given inadequate psychological support and had been let go abruptly when the contract ended. Some also said confidentiality agreements kept them from telling even family members what their work involved.
These accounts are not isolated. Workers moderating content for social media platforms have raised similar complaints for years, and the AI boom has expanded demand for the same kind of labor.
The material footprint
The cost of AI is not only human. Training and running large models requires data centers, specialized chips, large amounts of electricity and water for cooling, and land. The hardware depends on lithium, cobalt, nickel, copper, gold and rare earth elements, much of it mined under conditions that carry their own labor and environmental problems. The technology commonly described as living in the cloud rests on extracted metals, heavy energy use and growing pressure on local resources wherever data centers cluster.
The question that follows
Technology companies argue that AI will help cure diseases, address climate change and raise living standards. Those claims may prove partly true. But they sit alongside a less comfortable set of facts about who is doing the unglamorous work of making these systems usable, how they are paid and what the work does to them.
The debate about AI tends to focus on how intelligent these systems can become. An equally important set of questions concerns who pays the cost of building them, who collects the profit and who absorbs the harm. Those questions do not have technical answers, and they are unlikely to resolve themselves without public attention and regulation.
