The Hidden Cost of AI: African Workers Paid a Dollar an Hour to Train Your Models

I’ve been following the AI industry for years, and one thing that always gets glossed over in the breathless press releases about “democratizing AI” is the actual human labor behind it. We talk about models, data pipelines, and compute budgets, but rarely about the people who make it all possible—the ones sitting in windowless offices in Nairobi and Gulu, manually labeling data and scrubbing toxic content from social media feeds.

A recent investigation by James Muldoon, Mark Graham, and Callum Cant lays it bare. The piece, published in The Guardian, profiles workers at outsourced centers in Kenya and Uganda who do the grunt work for Meta and other tech giants. And the stories are brutal.

Take Mercy, a content moderator in Nairobi. She was tasked with reviewing flagged Facebook content—videos of violence, gore, and other violations. Her quota? One ticket every 55 seconds over a 10-hour shift. One day, a video of a fatal car crash popped up. She watched it, then recognized the victim: her grandfather. The company’s response? Her supervisor told her she could have the next day off, but since she was already at work, she might as well finish her shift. Then more videos of the same crash appeared. Different angles. More angles. More images of the dead. She had to keep watching.

This isn’t an outlier. Workers reported witnessing suicides, torture, and rape “almost every day.” They’re expected to process 500 to 1,000 tickets per shift. When they break down—crying, shaking, collapsing—they’re told they violated company policy for not logging an “idle” or “bathroom break” code, which hurts their productivity scores. One worker described themselves as a “walking zombie.”

And what do they get paid? Just over a dollar an hour. In some cases, less.

The cognitive dissonance here is staggering. We celebrate AI’s ability to generate art, write code, and summarize documents, but the foundational data that makes all that possible is often produced under conditions that would be illegal in the countries where these models are designed. The tech industry loves to talk about “alignment” and “safety,” but it seems to stop at the border of the global south.

I’m not naive—I know outsourcing has been a thing for decades. But there’s something particularly insidious about the way AI companies have offloaded the most psychologically damaging work to vulnerable populations, all while maintaining a pristine public image. Meta’s content moderation policies are strict about what users see, but apparently not about how the people enforcing those policies are treated.

The article also mentions data annotators—the ones who label images, text, and audio to train machine learning models. Same story: long hours, tight quotas, exposure to disturbing material, and minimal support. A 30-minute “wellness” session with a colleague who has no psychological training? That’s not support, that’s theater.

I’ve written before about how the AI industry’s labor practices are a ticking time bomb. This piece confirms it. The workers are burning out, quitting, and in some cases attempting suicide. And the companies? They just find new workers. The supply of desperate labor in the global south is essentially infinite.

What frustrates me is that this isn’t even hidden. It’s an open secret. Tech journalists have been covering it for years. But the industry keeps marching forward, raising billions, and patting itself on the back for “advancing humanity.” Meanwhile, people like Mercy are left to process their own grandfather’s death on a screen, for a dollar an hour.

I don’t have a neat solution here. Regulation? Maybe. But the real change has to come from within the industry—a willingness to pay a living wage, provide real mental health support, and treat these workers as human beings rather than interchangeable cogs in a data pipeline. Until then, every time you use an AI tool, remember: there’s a good chance someone in Africa saw something terrible so you didn’t have to.