The Human Cost of AI: African Workers Annotating Data for a Dollar an Hour

We hear a lot about the wonders of AI, but rarely about the people who make it work. Not the engineers in Silicon Valley, but the thousands of workers in places like Nairobi and Gulu, Uganda, who spend their days staring at the worst the internet has to offer, for about a dollar an hour.

Take Mercy. She works as a content moderator for Meta, in an outsourced office in Nairobi. One day, a flagged video pops up on her screen: a fatal car crash. She’s supposed to process one of these every 55 seconds, during a 10-hour shift. She leans in, recognizes a face — it’s her grandfather. She runs out crying. Her supervisor comforts her, then reminds her she needs to hit her targets for the day. She can have tomorrow off, but she might as well finish her shift. Then the new tickets keep coming: the same crash, from different angles, over and over. Her shift feels endless.

This isn’t an outlier. James Muldoon, Mark Graham, and Callum Cant spoke with dozens of workers at three data annotation and content moderation centers run by one company across Kenya and Uganda. The demands are brutal. Moderators witness suicides, torture, and rape “almost every day,” as one put it. They normalize the abnormal because they have to. There’s no zoning out — you have to watch every second of every video to correctly tag the highest violation, according to Meta’s policies. Violence and incitement rank higher than bullying, so you can’t stop after spotting one thing. You watch the whole thing, in case it gets worse.

One worker described feeling like a “walking zombie.” They’re expected to process between 500 and 1,000 tickets a day. The psychological toll is severe: “Most of us are damaged psychologically, some have attempted suicide … some of our spouses have left us.” Management offers a 30-minute break with a “wellness counsellor” — a colleague with no formal training. Workers who run from their desks after seeing a beheading are marked down for not entering the right “idle” or “bathroom break” code. One said, “I collapsed in the office.”

This is the hidden labor behind AI training data and social media moderation. It’s not just the graphic content; it’s the relentless pace, the low pay, and the lack of support. These workers aren’t building the algorithms — they’re the ones who make them learn, by labeling data and filtering out the garbage. And they’re doing it for wages that wouldn’t cover a meal in the cities where the tech companies are headquartered.

I’ve been following this story for a while, and it’s deeply unsettling. The AI industry loves to talk about “alignment” and “safety,” but it rarely acknowledges the human cost of its data pipelines. If your AI model is trained on data annotated by people who are traumatized, exhausted, and underpaid, what does that say about the system? It’s not a bug — it’s a feature of the current outsourcing model.

These workers deserve better. Not just better pay, but real psychological support, shorter shifts, and the right to walk away without penalty. Until the industry faces this, every AI model built on this labor carries a hidden stain.