Refugees in Kakuma, a camp in northern Kenya, used to earn a few dollars a day from online tasks like labeling images and transcribing audio. As that work has dried up, many of the tasks left pay in "rewards", which means the client decides afterward whether to pay and how much. It looks like a story about one camp. It is a preview of how AI changes the terms of a job before it removes the job.
A prize moves the risk onto the worker
A refugee in the camp, who asked to be named only as Grace, used ChatGPT to translate a Christian hymn into an east Asian language she does not speak. She then searched for existing versions to find who translated and wrote it. The project advertised $9,000 in rewards, shared among an unknown number of contributors, and the best work could earn up to $500. She did not know the client, and she did not know whether she would be paid.
Paying crowds in prizes is an old model: RWS, the company behind that platform, has paid patent searchers in rewards for years. What is new is who is in the queue. People who used to get a wage are moving into it, because the wage work is going. Remotasks, a platform many in the camp relied on, shut down in Kenya in March 2024.
Spread the potato contest across everyone who took part and each person earned about $6. That is the top of what a single day of the old wage work paid, $3 to $6. A prize lets the buyer see many attempts and pay for one, and when there are more workers than jobs, the buyer never runs out of people willing to try. Clients in this market also tell the trainers of refugees that the work can move to the Philippines or India.
The client prices the new speed
The prize is one change. The other is quieter. Hubert Sanga, a Congolese refugee who runs a training group for refugee workers, describes what clients now say: if a job takes three days, they will find someone who does it in two hours. Three working days is about 24 hours, so the price is being set against less than a tenth of the time.
The worker did not get worse. The client cannot see the difference between three days of careful work and two hours of AI-assisted work, so the price moves to the faster number. The same will happen to anyone who bills by the hour. A freelance bookkeeper who charges for ten hours a month to reconcile a shop's accounts will meet an owner who has heard that AI does it in one. She can bill one hour. Or she can sell the monthly close at a fixed price: accounts checked, odd entries explained, questions answered. The second only works if the owner knows and trusts her, and that is the one thing anonymous platforms deny. Sanga's group finds work directly with companies and puts its workers in contact with the client, so they can see what their work is worth to the buyer.
The first rung is going first
Jobs like transcription, data entry, translation and web research have fallen by about 50% since 2022. Biamungu used to get three to five projects a month. A South Sudanese engineer in the camp sees the same thing one level up: senior engineers are not hiring juniors. Entry-level jobs most exposed to AI are shrinking elsewhere too, and Kenya's essay writers lost their trade to ChatGPT.
Entry-level work is where people learn what good work looks like and earn their first references. When it goes, the people with the least schooling and the fewest contacts are cut off first. The training groups are moving away from microwork toward software engineering and e-commerce, but most refugees will struggle to reach those jobs. Reviewing AI output and cleaning data may become the next entry-level work.
Contest pay and anonymous clients reached the camp first because the workers there have the least room to refuse. Anyone offered a prize in place of a wage should ask two things before starting: who the client is, and what the guaranteed minimum is. If the answers are nobody and nothing, the worker carries all the risk.