In a portable building in Kakuma, a refugee camp near Kenya’s northern border with South Sudan, Grace used ChatGPT to research translations of Christian hymns in languages she doesn’t speak. 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.

Grace, who requested anonymity due to a nondisclosure agreement, prompted ChatGPT to translate the hymn’s English text into the assigned east Asian language. She then used the translation to search for existing versions and find information about its translator, author and Christian denomination. She fed those details to AOP Connect, a crowdsourced research platform owned by British AI enterprise software company RWS.

If RWS deems her work high quality based on six listed criteria, it could award her up to $500 as a “discretionary reward”, according to RWS researcher documents and manuals reviewed by the Guardian. But refugees aren’t guaranteed wages, and most earn significantly less than the maximum without being told why, said Asha Luka Abdallah, a mentor at the Solidarity Initiative for Refugees (SIR), an organization that trains refugees to work in the tech outsourcing industry.

Remote jobs such as content moderation and data annotation have been promoted as a win-win by United Nations agencies as well as global tech companies: refugees get income opportunities, AI and social media companies get a source of cheap labor.

But interviews with more than two dozen refugees in Kakuma show that only one side has clearly benefited. Tech companies have built some of their most powerful tools on datasets handled by precarious workers, including refugees, in low-income countries. But their pay, which was never high, appears to be vanishing as entry-level tech work disappears. The shrinking number of opportunities is driving some to accept gigs for potential “rewards” rather than guaranteed wages, many said.

“It’s one of the things that keeps me awake at night,” Bahana Hydrogene, a Congolese refugee who set up the SIR’s digital skills programs in 2016, said of the increasingly opaque arrangements for refugees. “They can’t get transparent answers from these companies.”

A spokesperson for AOP Connect said the company offers researchers, regardless of location or background, the opportunity to participate in a wide range of projects. The reward structure is communicated “transparently upfront”, as is the nature of the copyright research work, but confidentiality is critical to AOP’s clients, the spokesperson said.

Pay is based on the quality, relevance and timeliness of research submitted, and the AOP Connect platform is “not AI-enabled or used for AI-model training work”, the spokesperson added.

The ‘invisible architects’

Before Grace could become a qualified researcher, she filled out a questionnaire, agreeing to keep her work confidential and complete tasks by herself. Then she was able to potentially earn money through so-called “microwork”, digital tasks broken off of large confidential projects and distributed to remote workers via online platforms.

She logged into AOP Connect to see a list of opportunities and found the hymns research project advertised at $9,000 in rewards available. That figure, however, would be divided among an unknown number of contributors as discretionary “rewards” rather than guaranteed wages, according to a notice displayed on the listing. RWS said higher-quality work leads to higher pay, but there is little transparency about how assessments are made, according to program participants and payment notices reviewed by the Guardian.

Chatbot-assisted research on behalf of mysterious clients offering the chance to win rewards was not what Hydrogene had in mind for when he set up the SIR’s digital skills programs in 2016. He hoped the booming global outsourcing industry would help refugees earn a guaranteed income.

Some highly skilled refugees in Kakuma have found remote work through outsourcing and freelancer platforms to be a gamechanger. Those who did say they worked as web developers and software engineers interacting directly with clients. But most refugees found themselves limited to performing low-paid microwork such as data annotation for anonymous clients abroad, mediated through platforms like Remotasks, the Scale AI subsidiary that left Kenya in 2024.

Life online in Kakuma

Kakuma was originally established as a temporary home for people forced to flee the second Sudanese civil war, later expanding to accommodate people fleeing Somalia, Ethiopia and the Democratic Republic of Congo. More than 30 years later, it is one of the world’s largest refugee camps, with sprawling networks of tin-roofed buildings that make it look more like a permanent town.

On a Thursday afternoon in June, temperatures topped 95F at the SIR’s compound, about 20 minutes away from Kakuma’s main entrance. The arid landscape and severe water shortages around the camp make farming unfeasible, so the more than 300,000 residents are almost entirely dependent on humanitarian aid.

Inside an air-conditioned shipping container computer lab, five students practiced everything from basic digital literacy to full-stack programming. Nearby, three SIR graduates built websites and apps for clients outside the camp in a coworking space where they can access Starlink internet service and reliable electricity.

One South Sudanese software engineer said that, while he has managed to get well-paid work through the SIR and Konexio, a French organization that sources remote opportunities for refugees, he worries about future workers. Entry-level work is disappearing due to AI, he said.

“Senior engineers … are not hiring interns or junior roles,” he added.

The SIR has trained more than 2,000 refugees, connecting more than half with remote tech work ranging from long-term employment to short-term gig work. A few years ago, SIR staff said they could easily find data annotation, transcription and translation work for their students on platforms like Remotasks. Hydrogene said workers regularly took home 500 to 1,000 Kenyan shillings, or $3 to $6, a day before Remotasks shuttered in 2024. While pay was low, it was valued in a place where less than half of households have income, and those that do earn less than $50 per month.

But the microwork economy is often exploitative, said Joan Kinyua, founding president of the Data Labelers Association. The Nairobi-based organization advocates for the rights of what it calls the tech sector’s “invisible architects”, workers foundational to the industry but hidden by opaque supply chains.

Kinyua, who worked many of these jobs, said she noticed wages and conditions for entry-level data annotation eroding as early as 2018. On some platforms, she noticed that even making a small mistake, like outlining an object imperfectly, could cost her pay. Further reductions in pay and opportunities for data labeling in 2020 meant that Kinyua had to work faster for less, and had to be available anytime gigs appeared on outsourcing platforms. Eventually, she was sleeping only a few hours a night.

“You can’t afford to stay away from your computer,” she said.

In Kakuma, refugee data annotators said that at times they were tasked with labeling images and videos of weapons, gruesome bodily wounds, and pornography on various outsourcing platforms. As a result, some workers, especially those fleeing wars, experienced post-traumatic stress disorder symptoms after annotating violent material, Hydrogene said. And the opacity of some arrangements meant workers often unknowingly contributed to products they were opposed to. The Bureau of Investigative Journalism found that Somali speakers in Kakuma had done transcription work likely used by the US military, which refugees told the bureau they didn’t consent to.

A spokesperson for RWS said that AOP Connect doesn’t offer data annotation work. Scale AI did not respond to the broader findings on the industry as a whole, saying it was not in a position to comment on the practices of other companies.

Dwindling opportunities

Work became more scarce between 2022 and 2024 as Remotasks closed and automated tools such as ChatGPT took off. Na’amal, a non-profit that links refugees with global employers, noticed clients requesting skilled specialists rather than workers with basic digital literacy.

“As AI became mainstream, some jobs started to dry up,” said Lorraine Charles, Na’amal’s founder.

The International Trade Centre (ITC), a joint World Trade Organization and United Nations agency that has trained digital workers in Kakuma, estimates jobs such as transcription, data entry, translation and web research have declined by about 50% since 2022, said ITC spokesperson Susanna Pak. Oxford Internet Institute research similarly shows that growth of popular AI tools coincided with a measurable drop in many of the tech jobs promoted as paths to refugee self-sufficiency.

Merci Biamungu, a refugee trained in 3D modeling, said he regularly received opportunities with guaranteed pay throughout 2023 and 2024.

However, in one of his final projects, he and 20 other workers competed in teams of five to design a 3D model of a potato. The best-performing group received a reward of 15,000 Kenyan shillings, or about $116, split among the five. Now, the AI tools that Biamungu believes he helped train can build 3D images faster, eliminating the need for workers like him, he said.

It’s now been two years since Biamungu has found remote work in tech.

“We used to get three to five jobs per month,” he said. “Now you cannot get one.”

Hubert Sanga, a Congolese refugee who established Generation Aid, another social enterprise that trains and helps refugees find work in tech, said he’s seeing entry-level work for refugees collapse as AI expands.

“One hundred per cent of the jobs we are doing right now are really at risk,” Sanga said.

Rewards, not wages

For many, the alternative to these precarious, lower-paid remote gigs is no income at a time when many families in Kakuma face starvation – partly the result of Donald Trump and Elon Musk’s dismantling of USAID last year.

“The language of ‘rewards’ instead of wages shifts risk entirely on to workers, who invest their time and expertise without any guarantee of fair compensation,” said Mark Graham, professor of internet geography at the Oxford Internet Institute.

Hydrogene, of the SIR, said clients within the microwork economy routinely threaten to take jobs to the Philippines or India if he tries to negotiate higher pay.

“AI is taking some of these jobs … and you have more talent available than there are jobs,” he said. “They say, ‘Someone else will do it for less.’”

Na’amal’s Charles has also noticed declines in wages, and as a result is shifting the organization away from microwork entirely. Graphic design, web design and search engine optimization trainings are also being phased out in favor of software engineering, digital marketing and e-commerce, and other jobs believed to be less vulnerable to automation.

But most refugees will struggle to become competitive candidates for such highly skilled jobs, said Konexio’s director, Fabien de Castilla, due to disruptions to their schooling and work experience. If annotation jobs are becoming obsolete, Castilla still believes that other entry-level jobs, such as data cleaning and reviewing AI outputs, might provide opportunities for refugees in the future.

Building better protections

As AI speeds up the refugees’ work, it’s also diminishing their pay. Jobs that used to take days now take hours, reducing billable time, Sanga said.

“If I’m going to take three days to do a job, [the client] says, ‘I’m going to find someone who can do it in two hours,’” he said. “It’s killing us” on pricing, he added.

Diminishing pay has prompted the SIR, Generation Aid and Action for Refugee Life, another organization that trains refugees and sources tech jobs, to push companies to offer higher wages and better conditions. But tech workers told the Guardian they feared pushing for more over concerns they’d be replaced by AI or another worker desperate for a job.

Multilateral organizations like the ITC are trying to understand how digital workers can be better protected. Pamela Coke-Hamilton, the ITC’s executive director, told the Guardian that the loss of the digital economy’s lower- and middle-level workers demands new ways of thinking about collective bargaining.

“We are at a tipping point,” she said. “AI has taken people by surprise with its speed and impact, and we need to ask what labor organizing looks like now. We need guardrails to prevent mass exploitation.”

Though workers and advocates are trying to build those guardrails, they said power ultimately rests with global tech companies, which benefit from an oversupply of cheap labor. But a “race to the bottom” is not inevitable, Graham argues. He’s pointed to the textile industry as a precedent for how pressure can change industry conditions. There, “sweatshop” scandals prompted many companies at the top of the supply chain to commit to responsible business practices at every level.

Sanga believes one practical fix is simple transparency: workers need to know who they are working for so they can understand the value of their labor. When Generation Aid sources work directly from companies, it puts workers in direct contact with the client – a sharp contrast to anonymous platform work.

Milagros Miceli, principal investigator at Data Workers’ Inquiry and research lead at the Distributed AI Research Institute, argues that the supposed choice between no job and a poorly paid one is false.

“Without data workers, tech companies do not have a product,” she said. “So there should be space for bargaining power.”