Teresios Bundi's first online assignment paid $7 for three hours: a two-page essay on the health benefits of a fruit he had never heard of, for a student he knew nothing about [4]. That works out at about $2.33 an hour [2]. He estimates he wrote more than 2,500 essays over twelve years, sometimes three in a day, while finishing his own public health degree [5]. By the time he was running his own operation he charged $40 to $70 a paper, six to ten times his first fee [6][3], and had rented a house near a university with desks and mattresses so the writers he hired could work between classes [6]. It paid at least five times what public health would have [7].
The trade did not disappear so much as change what it sells. The jobs that remain are mostly for people who edit AI-written papers so they will not trip the detection software universities now run [8], and one writer told the New York Times that business is as good as ever: she drafts with a model and rewrites by hand [9]. The buyer's test moved from a reader to a checker, and the paid skill moved with it.
There is a rehearsal for this in the same city. Frida Mwangi took up transcription in 2015 and watched the work start drying up from 2017 [10]. Her employers' fix was to have her run the audio through AI and correct the output; her wages fell, and she found the correcting slower than typing from scratch, because the models mangled medical terms [11]. The contracts went anyway. "My job went completely," she told the Times [11].
The forward-looking number in the reporting is an index run by Scale AI and the Center for AI Safety, which scores how much online freelance work leading models can actually complete across categories like product design and data analysis. It read 2.5% last October and 16% by July [12], a 6.4x increase in nine months [1]. That measures what models can finish, not what clients stopped buying from humans, and Scale AI is not a neutral party to the market it is scoring: it discontinued some of its own Kenyan operations, and Sama, which hired Kenyans to moderate content for Meta, cut back after a backlash over pay and conditions [17].
The sizing of the essay collapse is thinner than it first looks. There is one researchers' estimate at the peak and a qualitative "almost nothing" after it, with no second count [3]. What the evidence does carry is the sequence, and the sequence is specific: a deliverable graded by software, then a post-editing rate, then no work.
The policy sits on top of that. Kenya began training graduates for freelance platforms in 2016 [18], and more than 100,000 students leave its universities every year [15]. John Tanui, principal secretary at the responsible ministry, told the Times he stands by Kenya's potential as an outsourcing hub and said it would unlock opportunities for young people [14]. Mark Graham, the Oxford professor who studies the global gig economy, put it to the Times as "A.I. has come in and taken big chunks of these labor markets," and said Kenya will not be the only place [13].
For anyone buying or selling a text service line, what settles it is who inspects the deliverable: a person who will notice, or software that returns a score, and whether the staffing plan is now described as fixing what the model got wrong. Software plus post-editing put Mwangi in that position by 2017: her pay fell first, and the contract ended anyway [11]. A human inspector who will notice keeps a bill you can defend. Nairobi's surviving niche is the other case: it exists because universities keep paying for detectors [8], which is a thinner thing to depend on than the students.
