The AI jobs debate keeps circling around forecasts: will half of entry-level jobs vanish in five years, will agents come for junior coders, will the white-collar wave hit in 2027 or 2031. Kenya has stopped debating. The New York Times reported on 5 September that the essay-ghostwriting industry in Nairobi — estimated at up to 40,000 workers at its peak — has effectively vanished three years after ChatGPT launched. This isn’t a prediction about what AI might do to work. It’s the after-action report.
The numbers, via the NYT’s reporting as carried by SBS and Asia Business Daily:
- Teresios Bundi, 34, wrote more than 2,500 essays over 12 years, earning $40-70 per assignment — several times what his public health degree would have paid. “I never imagined AI could do this kind of work,” he told the NYT.
- Richard Esilaba, 38, had scaled to employing 100 ghostwriters. He closed the business and now does photo editing. AI-assisted, he says, but earning nothing like the old income.
- Alex Munyua, 30, made the pivot the entire “just reskill” argument depends on — from ghostwriting into data labeling and content moderation, the classic AI-economy landing pad. Those jobs are now disappearing too.
The Two-Step Collapse
What makes the Kenyan case a preview rather than an outlier is its structure. Step one: AI automates the core service. ChatGPT launched in late 2022; by 2026 the industry is gone. Step two, now under way: AI moves on the fallback work. Kenya became a global hub for data annotation — the human labour that trains and evaluates AI systems — precisely because it offers English fluency, high literacy, reliable power and a workforce familiar with Western culture, as one Nairobi AI company CEO told Rest of World. Kenya’s own draft AI policy, released in July by its Ministry of Information, Communications and the Digital Economy, put local AI data workers’ pay at $1.46 to $3.74 an hour. Those were the good jobs in the fallback sector.
The forward indicator is Scale AI’s own benchmark. In a joint test with the Center for AI Safety, the share of online freelance tasks that AI models could complete rose from 2.5 per cent in October 2025 to 16 per cent by July 2026. Nine months, six-fold. The base is small, and freelance task completion is not the same as employment effects — but the trajectory is the whole argument, and it’s coming from the company that brokers much of this work. When the marketplace itself publishes data showing its human supply shrinking in usefulness, that’s a different kind of signal than an AI lab’s marketing deck.
There’s an irony worth naming, carefully. The scale-up of AI evaluation work — people grading model outputs — is itself the mechanism teaching models to do the tasks being evaluated. Data annotation is the apprenticeship of its own replacement. That’s a structural feature of the industry, not an accusation against any firm: the work is real, paid and, by Kenyan standards, well-located. It is also, by design, temporary.
Why the Western Conversation Keeps Missing This
The AI-labour debate in English-speaking media is overwhelmingly about knowledge workers in the countries that build the models. The NYT’s Kenya story inverts that: the first fully-observed AI labour collapse happened in the global South, in work that was already digital, already remote, already priced by global competition. No unions, no severance, no retraining programme — workers describe returning to home villages or drifting between informal jobs. Kenya produces more than 100,000 university graduates a year into a labour market where roughly 80 per cent of jobs are informal and youth unemployment exceeds 25 per cent. The ghostwriting economy was, in effect, the formal sector’s overflow valve. AI closed it.
Bundi’s line to the NYT is the one that will age either badly or perfectly: “AI is now coming for bankers, accountants, engineers, and even architects. The job crisis is something that will eventually come for everyone.” Kenya just happened to be first because its work was easiest to hand to a model — fully remote, text-native, English-language, deliverable through a screen. The difficulty order for AI automation turns out to run roughly in reverse order of how much the work pays, and the people who found that out first were earning $1.46 to $3.74 an hour doing it.
NZ Angle: The Low-End Is the Leading Indicator
New Zealand keeps getting reassured that AI hits routine work first and augments the rest — which is true, and is exactly why the Kenya story deserves more attention here than it’s getting. This country’s own exposure to the same pattern is real and specific: Kiwi freelancers and agencies compete in the same global marketplaces where AI-native competition is now bid-setting; the employment data we do have shows the entry-level end thinning while experienced wages hold; and Anthropic’s own labour-market research found no systematic unemployment effect yet — a gap between task-level takeover and job-level outcomes we’ve tracked since July. Kenya is what the task-level takeover looks like when it completes in a market with no floor. The policy question for NZ isn’t whether a Nairobi-style collapse arrives in Wellington; it’s which local sectors share the same profile — remote, text-based, globally priced — and the honest answer includes parts of the writing, translation, design and basic dev services economy that this country sells to the world.
❓ FAQ
How big was Kenya’s essay-ghostwriting industry? The New York Times reported an estimated 40,000 people in Nairobi worked in the industry at its peak, writing university assignments primarily for overseas students, at rates of $40-70 per essay.
When did it collapse? Orders dried up and rates collapsed after ChatGPT’s late-2022 launch, according to the NYT’s reporting published 5 September 2026. Three years from peak industry to effective disappearance.
What is data labeling and why does it matter here? It’s the human work of classifying and verifying data used to train and evaluate AI systems. Nairobi became a global hub for it. Kenya’s draft AI policy puts local pay at $1.46-$3.74/hour — and those jobs are now also contracting, according to workers cited by the NYT.
What did Scale AI’s benchmark find? In a joint test with the Center for AI Safety, the share of online freelance tasks AI models could complete rose from 2.5 per cent in October 2025 to 16 per cent by July 2026. It measures task completion, not job losses — but it’s a six-fold increase in nine months from the industry’s own leading marketplace.
Does this mean mass AI unemployment generally? Not by itself. Kenya’s case is specific: fully remote, English-language, text-based work with no regulatory floor. It’s the clearest completed example of AI substituting for an entire labour category, which is why it functions as a preview for the highest-risk profile everywhere — not a forecast for all work.
🔍 THE BOTTOM LINE
Every AI-jobs forecast so far has been about the future, and every one has been arguable. The Kenya ghostwriting collapse is not arguable. An industry of tens of thousands, paying multiples of local professional wages, went from thriving to gone in three years, and the fallback sector — the data-labeling work that was supposed to absorb displaced digital workers — is itself thinning as models take over more of the task pipeline. The Scale AI benchmark’s 2.5-to-16 per cent jump in nine months is the same story told as a chart. When the first confirmed casualty arrives, the debate should update from “will it happen” to “which of our sectors has the Kenya profile.”
