The 40,000 Ghostwriters AI Erased: Inside the NYT's Kenya Investigation
ChatGPT collapsed Kenya's essay-mill economy in two years — gig writer pay fell from $1,200 to $500 a month as AI capability on freelance tasks jumped from 2.5% to 16%. The NYT's Teresios Bundi profile is the first full autopsy of an industry AI actually killed.
For two decades, Kenya ran one of the internet’s stranger shadow economies: a genuine, professionalized industry of academic ghostwriting, staffed by tens of thousands of educated young Kenyans producing essays for students in the United States and Britain. On September 5, The New York Times published the most complete account yet of how fast that industry died — and why its collapse is a warning for every country betting on online gig work as an employment strategy.
The piece centers on Teresios Bundi, a writer from a farming village who built a career writing more than 2,500 academic papers for Western students at $40 to $70 each. His arc — from a profitable if ethically fraught livelihood to a market that “shrank to almost nothing” — is the human front page for a labor story with unusually hard numbers behind it.
The scale of what disappeared
Researchers estimated the Kenyan contract-cheating industry paid at least 40,000 people at its early-2020s peak. This was not a cottage trade; it was infrastructure. According to Rest of World, roughly 70% of online freelancers in Kenya offered writing and translation services as of 2023, and the International Labour Organization’s 2024 report on digital labour platforms in Kenya found that 36% of surveyed online freelance workers mainly performed academic writing, with another 17% doing article writing and 15% editorial services.
Then ChatGPT arrived on November 30, 2022, and the cheapest buyer in the market suddenly had an even cheaper option. The decline was not gradual. The Times reports that gig-writer pay in Kenya has fallen from $900–$1,200 per month to $500–$800, with assignments dwindling toward zero — a collapse compressed into roughly two years. One ghostwriter cited by Rest of World, identified only as Collins, saw his monthly income fall through exactly that band after generative AI tools spread among students.
The most striking data point in the coverage comes from Scale AI and the Center for AI Safety, which measured leading models’ capability on real online freelance tasks rising from 2.5% completion in October 2025 to 16% by July 2026. That sixfold jump in a single year is the mechanism: each incremental model release converted another slice of “hire a human” tasks into “just prompt the model.”
Why this industry, why this fast
The essay-mill case is close to a controlled experiment in AI labor displacement, because the work sat at the extreme end of several risk factors at once:
- The output was generic text. A passable undergraduate essay is exactly the kind of artifact large language models produce by default. No physical presence, no local knowledge, no licensed tools required.
- The buyer had zero loyalty. Students buying essays cared about price, speed, and not getting caught — not about who wrote the paper. The moment an acceptable substitute cost $20 a month, switching was instant.
- The workers had no protection. Anonymous accounts, offshore brokers, no contracts worth the name. There was no severance, no retraining fund, no notice. The demand curve simply moved.
- The ethics were already broken. Unlike nursing or teaching, nobody was going to legislate to save contract cheating. The industry had no constituency.
That combination explains the speed. But the Times’ deeper point is that the same geometry — text in, text out, distant buyer, unprotected worker — describes an enormous share of legitimate online gig work: transcription, basic translation, web copy, first-draft research summaries, customer-support scripts. The essay mills simply went first because they were the least defensible.
The uncomfortable Kenyan paradox
What makes the story genuinely painful rather than merely ironic is that Kenya had spent years building exactly this pipeline as a matter of national policy. The country has a large English-speaking graduate population and chronically patchy formal employment, and programs like Ajira Digital and the Jitume hubs were explicitly sold as routes to youth income via online work. The ILO found the workers were educated and underemployed — real skills chasing fake careers.
The timing is brutal. In May 2026, Kenya’s ICT ministry announced a “Digital Outsourcing Jobs for Kenya’s Youth” initiative aligned with the National Digital Master Plan 2022–2032. In August 2026, the Kenya Investment Authority and the Outsourcing Alliance of Kenya signed a cooperation agreement to attract global business services and BPO investment. The government is courting the next wave of outsourcing work in the same month the Times documented the last wave drowning.
There is a viable path in the new landscape — human review of model output, domain-specific data work, quality assurance, AI operations, customer roles requiring judgment — and notably, Bundi himself embodies the transition. His LinkedIn now lists him as a Gig and Data Economy Advisor at GIZ, Germany’s development agency, working on digital economy and youth employment policy. The man who wrote 2,500 essays for strangers now writes policy about the market that replaced him.
What to watch next
Three numbers define the trajectory. First, the Scale AI/CAIS task-capability metric: at 2.5%→16% in nine months, the models are eating freelance work from the bottom of the difficulty curve upward, and every frontier release since — GPT-6 Astra, Claude Fable 5.1, Gemini’s current generation — pushes further up that curve. Second, the pay band: $500–$800 a month is not zero, which means the residual work is the part AI still does badly, and that residue is shrinking. Third, the policy response: Kenya’s outsourcing push will only work if it treats AI as a force already inside the labor market rather than a future risk.
The Times piece is being read in policy circles as the first proper autopsy of an industry AI actually killed — not projected to kill, not threatened to kill. The lesson written in Nairobi is that when the task is turning generic text into other generic text, there is no floor under the price, and no loyalty to catch the workers on the way down.
Sources are listed in the frontmatter of this post. The New York Times article “Kenyans Made a Living Writing College Essays. Then A.I. Arrived.” was published September 5, 2026.