A young woman sits in front of a laptop at a co-working space in Nairobi’s Westlands neighborhood, surrounded by the kind of open-plan office furniture that could belong to any startup anywhere in the globe. When pressed, she refuses to describe the image she is staring at. She presses one button, proceeds to the next, and then presses once more. She spends eight hours doing this.
The images originate from a queue created by software run by a subcontractor employed by an American AI business whose name appears frequently in the media. There is not a single mention of her name. Her work is what makes the AI usable, not in a metaphorical sense, but literally: the model cannot learn to differentiate between harmful and safe content, real and imagined objects, and accurate information from hallucinogenic noise without people looking at these images and labeling what they contain. She is a part of the system in a functional sense.
Data is the foundation of the AI sector. Not raw data, but data that has been cleansed, labeled, categorized, filtered, and ranked by millions of unique human judgments applied to text, photos, audio, and video. In order to make ChatGPT and related models conversationally fluent and behaviorally acceptable, the technique known as Reinforcement Learning from Human Feedback requires human raters to assess AI outputs and determine whether responses are better, more accurate, more appropriate, or less destructive. The model must be observed, evaluated, and trained to perform better. that a significant portion of the work produced by the AI sector is not done by Silicon Valley engineers. They labor under contract for a few bucks an hour in Manila or Nairobi.
The pipeline passes via businesses that act as middlemen between the workforce and the AI developers. Platforms like Scale AI, Sama, and Remotasks have created companies that aggregate this labor, hiring workers in Kenya and the Philippines where internet connectivity is sufficient, English proficiency is high, and wages are significantly lower than what the same work would cost in the US or Europe. These businesses then provide clients like OpenAI, Meta, Google, and Microsoft with labeled datasets, content moderation queues, and human feedback pipelines. The AI firms can accurately characterize their own workforce without specifying who is actually performing the task, and the workers are unaware of whose AI product their annotations are training. As a result, the client relationship is at arm’s length in both directions.
The category with the highest reported human cost is content moderation work. Human reviewers must be exposed to violent content, extremist material, and images of child sexual abuse in huge quantities in order to train a model to refuse to produce such content. This is because the model must learn what to block, which involves human labeling. According to a January 2023 TIME investigation, employees at a Sama-run facility in Nairobi were examining violent violence and child sexual abuse footage for OpenAI. They were paid about $2 per hour and had access to a counselor one day a week. Interviewees reported experiencing intrusive visions, recurrent nightmares, and symptoms typical of post-traumatic stress disorder. After that, Sama terminated its agreement with OpenAI. The work went on in other places.
The psychological impacts of large-scale content moderation job are not exclusive to the AI sector; Facebook, YouTube, and Twitter have all come under fire for how their primarily outsourced workforce handles dangerous content on their platforms. However, the AI training context introduces a dimension—the volume required—that is absent from traditional content moderation. Human raters must be exposed to poisonous outputs in amounts that platform content moderation, which reacts to user reports, usually doesn’t provide in order to train a big language model to identify and reject those outputs. The training process incorporates the task’s scale.
What links this to more general discussions regarding the global distribution of the AI industry’s economics is the pay issue. Large AI model firms are worth hundreds of billions of dollars, have access to funding on a scale that would have been unthinkable for the IT sector ten years ago, and are developing products that are already bringing in significant sums of money. Instead of receiving compensation commensurate with the value of the job they provide, the workers who provide the human input that enables those items to function are paid according to the local cost of living in Nairobi or Manila. Arbitrage is not specific to AI; it is the same mechanism that has organized software testing, call centers, and global manufacturing for decades. The public story of the AI sector stresses independent intelligence, while the real product relies on human judgment at the lowest possible cost. This is where the differences lie.
