The short answer
AI does not take "professions"; it takes tasks. What determines how fast a job erodes is not its title but three properties: whether its output is codifiable, whether it can be delivered remotely, and whether it is priced by volume. When all three hold, the job sits where the model competes directly.
This piece looks at three separate fronts — an industry that was erased, a bank that changed its hiring conditions, and a school system pushing AI out of the classroom — and draws out what they share.
Front one: an industry erased in Nairobi
In Kenya, ChatGPT wiped out an entire business model: writing academic papers for foreign students. Thousands of Kenyans wrote essays and term papers for university students in the US and UK in fields like medicine, computer science and engineering, sometimes using their clients' university logins.
According to researchers, at its peak early this decade at least 40,000 people worked in the industry in Nairobi alone. Teresios Bundi, 34, wrote more than 2,500 papers over twelve years and later charged $40 to $70 per text. After ChatGPT launched in 2022, both prices and orders collapsed.
The collapse did not stop at one job: other online work in the country also dried up, including transcription, data annotation and content moderation. Kenya's government had deliberately promoted online gig work from 2016 onward, and about 80 percent of jobs in the country are informal.
What remains is the most ironic part of the story: "humanizers" — people who rework AI text so it slips past plagiarism checks. The industry did not disappear; it moved one rung down and turned into cleaning up the model's output.
Which jobs go first?
Let us unpack those three criteria, because they determine a job's risk level more than its title does:
| Criterion | What it means | Its Nairobi equivalent |
|---|---|---|
| Codifiable output | The result can be delivered as a single file — text, spreadsheet, code | What was delivered was an essay file |
| Remote delivery | No physical link between where the work is done and where it lands | Written in Nairobi, read in London |
| Priced by volume | Pay is measured by units produced, not by hours or responsibility | $40 to $70 per text |
If all three hold, the job sits where the model competes directly. If two hold, the pressure lands on price but the work continues. If one holds, the effect usually shows up as acceleration rather than replacement.
There is one more distinction to watch: liability. Work whose output is signed, audited, or whose errors are charged to someone — accountancy, medicine, engineering sign-off — erodes more slowly even when it meets all three criteria. What is bought there is not the text but the person standing behind it.
Seen from Turkey
The Nairobi story has a direct analogue in Turkey: translators, copywriters, data annotators and image producers working for foreign clients through freelance platforms. All three meet the criteria above, and the price pressure comes from the same direction.
There is a difference, though: Turkish is not a language with as much data as English. Model performance in Turkish translation, localisation and editing is not as high as in English — a buffer in the short term. But assuming the buffer is permanent would be wrong; that gap narrows with every model release.
The second difference is on the demand side. The industry erased in Kenya depended on a single customer type — the student abroad. Any job tied to one customer type disappears at once when that customer moves to another tool. Diversity of customer base protects as much as skill does.
Front two: the condition changed at the bank
Swiss bank UBS wants AI skills from graduates and interns applying to start in 2027 in Global Banking and Markets. Candidates are expected to show how they use AI to improve outcomes and efficiency, and interviews will include questions on it.
The condition sits alongside classic criteria like a strong degree — UBS says AI skills complement academic and social abilities rather than replacing them. Spain's Santander is also looking for advanced AI users in some trainee programmes.
The number in the background: Morgan Stanley analysts expect more than 200,000 banking jobs in Europe to disappear within five years. Banks are increasingly automating routine junior work such as financial analysis, research and client presentations.
The entry-level paradox
The real problem here is which rung automation strikes. Routine junior work — analysis, compiling research, preparing presentations — is exactly what the model does best. But seniors are made from juniors. A banker learns how the sector works by building spreadsheets and decks for years.
JPMorgan's Europe head Conor Hillery warned precisely about this in December: juniors must not lose the fundamentals. But how the fundamentals of a job get learned after that job is automated is a question nobody has answered. UBS, meanwhile, is testing analyst avatars for client presentations.
Front three: the school pushes AI out
New York City has banned AI tools in public schools through eighth grade. From the new school year, roughly 600,000 students can no longer use them.
- Individual screens are off-limits entirely through third grade.
- Companion chatbots are banned at every grade level.
- For middle school, 45 minutes of daily screen time is recommended.
- Teachers may use AI to prepare lessons but not for grading.
Mayor Zohran Mamdani says the tech industry wants to frame AI education as inevitable, and the city is not going along. Parent groups like Parents for AI Caution, meanwhile, call the rules too weak and want a full two-year moratorium. Their worry: if children hand reasoning off to software, they may never develop independent thinking. British researchers recently named this "cognitive surrender."
A task force of teachers, politicians, experts and union representatives will produce a report by April 2027, which will underpin the city's future AI policy.
The lesson the three fronts share
Put the three stories side by side and a single tension appears: business is making the skill a requirement while education is delaying it. UBS wants AI competence from 2027 graduates, while New York keeps the children who will finish eighth grade in 2027 away from these tools.
Both positions are defensible. The bank's reason is today's productivity; the school's is that basic cognitive skills, once missed, do not come back. But without a bridge between them the result is this: instead of a generation that learns to use AI after learning to think, a generation that learns both at once and haphazardly.
What the numbers do not say
The figures in these three stories need careful reading. "40,000 people in Nairobi" is a researchers' estimate for the peak period; precise counting is impossible in an informal sector anyway. Morgan Stanley's "200,000 banking jobs in five years" is an analyst projection — an expectation from a model, not a loss that has happened.
The difference matters: the first is an approximate measure of something that occurred, the second a forecast of something assumed to occur. Using them in the same sentence lends realised loss and prediction the same certainty. That is the most common mistake in coverage like this.
New York's 600,000 is the soundest of the three, because it is an administrative measure of scope: the number of students in the affected grades. But it measures no effect either — only the size of the decision. What will measure the effect is the report due in April 2027.
What to do in practice
Concrete directions from all three fronts:
- Break your job into tasks. "Will my profession survive" has no answer; "which six hours of my week produce a codifiable output" does.
- Move away from work priced by volume. What was erased in Nairobi was work billed per text. What remains is work that carries responsibility.
- Learn the fundamental skill before the tool. Learning to use the tool takes weeks; learning to judge its output takes years.
- For hiring: requiring AI skills is easy, measuring them is hard. "How do you use it" separates the candidate who lists tool names from the one who audits output only if the second can give a concrete example of an error they caught.
- For parents: the third path between banning and permitting is an arrangement where use is discussed openly. New York allowing teachers to use AI for preparation but not for grading draws exactly that line.
In summary
Teresios Bundi's warning from Nairobi sums up all three fronts: "A.I. is coming for bankers, for accountants, it's coming for engineers, and A.I. will come for architects. It's coming for everybody."
Oxford professor Mark Graham expects similar upheavals worldwide. But the real lesson of the Nairobi case is not the speed, it is the order: first the volume work went, then the entry rung that fed it narrowed, and what finally remained was a job cleaning up the model's output. That order makes knowing which rung you stand on more important than ever.