Behind every AI model, tens of thousands of people label data, rate responses and filter content. In Kenya, these workers are reportedly paid between $1.46 and $3.74 an hour, compared with $21 to $27 for equivalent work in the United States. In late July 2026, the Kenyan government presented a draft policy introducing pay benchmarks, mandatory psychological support and written contracts. The issue is real, and it is also more nuanced than the word "exploitation" suggests.
We talk a lot about models, chips and billions. We talk little about the people without whom none of it would work. This article corrects that imbalance, trying to hold both ends: the findings, which are stark, and the complexity, which is real.
Who these workers are
An AI does not learn on its own. As we explained in our article on training, the third stage of building a model relies on humans ranking responses from best to worst. It is this stage that gives an AI its character, its politeness and its sense of what is acceptable.
These people do three distinct jobs. Data annotation involves labelling images, text or audio so a model can learn to recognise them. Response evaluation involves comparing model outputs and indicating which is better. Content moderation involves identifying and removing content that must never appear, which means having to look at it.
This work is massively outsourced. Research by the Dutch organisation SOMO notes that Amazon, Google, Meta, Microsoft and Nvidia together use at least thirty intermediary companies, from outsourcing firms to micro-task platforms. This architecture creates legal distance: the big companies do not employ these people directly, while still heavily shaping their conditions through pressure on prices and deadlines.
The figures, and what they say
The gap is the central fact. According to the materials accompanying the Kenyan draft, data workers there would earn between $1.46 and $3.74 an hour. In the United States, a moderator doing equivalent work earns between $21 and $27.
A ratio of ten to fifteen for the same task. This is not a productivity gap, since the work is identical. It is a location gap.
Added to this is a difficulty that does not affect everyone, and this should be stated clearly. The vast majority of data work involves labelling mundane images or comparing model responses, with no distressing exposure at all. But a much smaller fraction of these workers is assigned to handling the most violent content, so that models learn to refuse it. For them, the burden is of a completely different nature. More than 35 Kenyan workers have brought legal action to have this exposure recognised as an occupational risk, which no legal framework currently clearly provides for. We devote the second part of this series to that question.
This work is also, for many, a genuine opportunity. Some locally based employers offer monthly contracts worth roughly twice the Kenyan minimum wage, with health insurance and subsidised meals. For a young English-speaking graduate, it is often better than the available alternatives. The problem is therefore not that this work exists, nor even that it is offshored: it is the gap with the value produced, the lack of protections, and the fact that the psychological burden is neither recognised nor supported. Conflating these two criticisms weakens the second, which is the stronger one.
What Kenya is proposing
The draft policy, the result of an eleven-month consultation process begun in November 2025 with civil society organisations, universities and public agencies, sets out four main measures.
A reference pay framework, calibrated not to the local minimum wage but to international rates for the sector. This is the most innovative point: it amounts to saying that work sold at a global price should not be paid at a local price.
Mandatory psychosocial support for people exposed to difficult content. Written contracts and transparency on pay. And appeal mechanisms for contesting a decision or a non-payment.
The text accompanies an AI bill under discussion in Parliament, which would introduce a risk-tiered framework and a dedicated commissioner. Activists, however, argue that the bill falls short on protections that are strictly about wages.
A movement that goes beyond Kenya
The issue is neither African nor confined to low-income countries. In the Philippines, investigations have documented pay below the legal minimum wage among subcontractors working for major groups. In January 2026, employees of a provider in Dublin, which notably supplies AI training services, went on strike to secure union recognition and better conditions.
Kenyan workers have also founded a professional association, which gathered several hundred members within its first week. It is a sign of a shift: a workforce that was until now dispersed and invisible is beginning to organise itself as a counterpart.
Why it concerns us
There is a very simple reason. The ease with which an AI answers you politely, refuses a dangerous request and avoids shocking content is not a magical property of the model. It is the result of human labour, done by someone who looked at that content so you would not have to.
Last month we documented staggering sums: financial guarantees in the hundreds of billions around data centres, record quarterly results. Setting those figures against hourly pay of $2 is not demagoguery; it is a question of proportion.
That said, the solution is not obvious. Banning subcontracting would eliminate real jobs without improving anyone's lot. Imposing Western wages in economies where the cost of living is ten times lower would create other distortions. The Kenyan approach, which consists of indexing to the value of the work rather than to geography, is probably the most interesting avenue to watch.
What seems beyond dispute, however, is that the psychological burden of moderation should be treated as an occupational risk, with the protections that go with it. On this specific point, the debate does not really pit two sides against each other: it pits those who look against those who would rather not know.