Refugees in Kakuma use ChatGPT for gigs, but pay is unclear

Grace, who asked not to be identified because she signed a nondisclosure agreement, works through AOP Connect, a crowdsourced research platform owned by RWS, a British AI and enterprise software company. To qualify, she filled in a questionnaire and agreed to keep her work secret and to do every task herself.

On the hymn project, she asked ChatGPT to translate the English lyrics into an East Asian language. She then used that translation to look for existing versions of the hymn and to track down details about its translator, author and church denomination, which she submitted to the platform.

This is microwork: small online tasks split off from large, confidential projects and handed to remote workers through a platform. The worker rarely knows who the client is or what the finished product will be used for.

The hymn project was listed with $9,000 in total rewards. That pot, however, is split among an unknown number of contributors, and a notice on the listing made clear the money is a discretionary reward, not a wage.

RWS documents say a single piece of work judged high quality on six criteria could earn up to $500. Asha Luka Abdallah, who mentors trainees at a refugee-run group called the Solidarity Initiative for Refugees (SIR), says most refugees earn far less than that and are never told why.

AOP Connect says it tells researchers upfront how rewards work and what the copyright research involves, and that pay depends on the quality, relevance and timeliness of what they submit. A spokesperson added that the platform is not AI-enabled and is not used to train AI models, and that clients require confidentiality.

A professor at the Oxford Internet Institute put the problem bluntly: calling payments rewards pushes all the risk onto the workers, who put in time and skill with no promise of fair pay.

Kakuma began as a temporary camp for people fleeing Sudan’s second civil war. More than 30 years later it is one of the largest refugee camps in the world, home to more than 300,000 people who rely almost entirely on aid, since the dry land and water shortages make farming nearly impossible.

SIR, set up in 2016 by Congolese refugee Bahana Hydrogene, has trained more than 2,000 refugees in digital skills and connected over half of them with remote tech work. Some now build websites and apps for outside clients from a co-working space with Starlink internet and steady power. Most, though, ended up doing low-paid data labeling for anonymous clients through platforms such as Remotasks. That Scale AI-owned service pulled out of Kenya in 2024.

Since then, trainers say, that kind of entry-level work has collapsed. A joint UN and World Trade Organization agency that has trained digital workers in Kakuma estimates that transcription, data entry, translation and web research jobs have fallen by about half. One refugee trained in 3D modeling said he used to get several jobs a month and now can’t find one. AI also speeds up the jobs that remain, so clients pay for fewer hours.

Hydrogene says the growing secrecy around these arrangements keeps him up at night, because refugees can’t get straight answers from the companies involved.

Some of the most powerful AI tools that people in the US, UK and Canada use every day were built partly on data handled by low-paid, insecure workers in countries like Kenya. Many of those workers never knew who they were working for. Some in Kakuma reported labeling violent and disturbing images, and some unknowingly worked on projects they would have refused.

As AI takes over the simplest tasks, the people who helped train it are often the first to lose income. That is a useful thing to keep in mind when a company describes its AI as automated: humans are usually somewhere in the supply chain.

Groups including SIR and Generation Aid, another refugee-run training group, are pushing companies for better pay and conditions, though workers fear being replaced if they ask for more. Generation Aid’s founder says a simple fix is transparency: workers should know who the client is, so they can judge what their work is worth.

Training groups are also shifting refugees toward software engineering, digital marketing and e-commerce, which they see as harder to automate. Researchers point to the clothing industry, where sweatshop scandals pushed big brands to take responsibility for their supply chains, as proof that pressure can work.