The AI that's supposed to automate knowledge work is being trained by people who don't know if they'll get paid this week.
The Summary
- Refugees in Kenya's Kakuma camp are doing micro-task gig work for tech companies, often using AI tools like ChatGPT to complete tasks without knowing the client, purpose, or guaranteed payment
- Workers operate through platforms like AOP Connect, owned by British AI enterprise software company RWS, while legally barred from formal employment in Kenya
- The irony is sharp: vulnerable workers use AI to train AI, creating the systems that will eventually replace them, all while earning wages that continue to fall
The Signal
Grace, working from a portable building in Kakuma refugee camp, translated Christian hymns into east Asian languages she doesn't speak. Her method: prompt ChatGPT for translations, verify against existing versions online, extract metadata about translators and denominations, then feed it all back into AOP Connect. She had no idea who wanted this work or why. She wasn't sure she'd be paid. But as a refugee legally prohibited from working in Kenya, she took the task anyway.
This is the ground floor of the agent economy. Not San Francisco. Not even Bangalore. Kakuma, a camp near the South Sudan border where people with no legal work status become the invisible hands training enterprise AI systems.
"She didn't know the client, the purpose of her work or whether she would be paid, but as a refugee unable to legally work in Kenya, she needed any opportunity she could get."
The platform model here reveals something important about Web4's labor structure:
- Workers are atomized, contracted through apps, bound by NDAs that silence them
- Payment is uncertain, arrival is delayed, amounts are declining as competition increases
- The actual clients, purposes, and end uses of the work are deliberately obscured
- AI tools are both the product being built AND the means of production for the builders
Grace signed a nondisclosure agreement to do research work she may not get paid for. Think about that power asymmetry. The refugee can't talk about the work. The company doesn't have to guarantee payment. The legal status that prevents formal employment becomes the leverage that makes informal exploitation possible.
RWS, the British company running AOP Connect, sits several layers above this. Enterprise AI software. Crowdsourced research platform. Clean corporate language that obscures the fact that their training data and knowledge work is being extracted from people living in tents who are using ChatGPT because they have no other options.
The race to the bottom isn't theoretical anymore:
- "Someone else will do it for less" is the operating principle
- Pay rates drop as more desperate workers flood these platforms
- AI simultaneously creates the work (tasks that need human verification) and destroys it (automation that replaces those same humans)
- The people building the future have no ownership stake in it
The Implication
If you're building agents, you need to know this is the foundation they're standing on. Not because of ethics, though that matters. Because this model is structurally unsustainable. You cannot build reliable intelligence systems on top of a workforce that doesn't know if they'll eat next week. Quality degrades. Resentment builds. The hidden human labor that makes "autonomous" systems work becomes the breaking point.
For workers everywhere: this is the test case. If companies can extract labor from the most vulnerable people on earth with zero guarantees, what makes you think your job has a floor? Web4 needs a different model. Ownership. Revenue sharing. Transparent contribution tracking. Not because it's nice, but because the alternative is a digital economy built on disappearing wages and broken promises.