This article describes the working conditions of AI annotators, including exposure to violent and traumatic content. No graphic descriptions are given, but the testimonies are harrowing. If you work in this sector or are confronted with it, resources are listed at the end of the article.
The story nobody tells you when you ask ChatGPT for a recipe
You type "write me a poem about autumn" into ChatGPT. In one second, a polite response appears, with no violent content, no racist remarks, no traumatic descriptions. This surface cleanliness does not come out of nowhere. It is the result of invisible work carried out by tens of thousands of people, in offices you will never see, in Nairobi (the capital of Kenya), in Manila (the capital of the Philippines), in Bangalore (India's tech capital).
These people are called data annotators or content moderators. Their job consists of labelling images, transcribing audio, checking responses generated by AI, identifying what is violent, racist, illegal, shocking. Without this human filtering work, no modern language model (LLM: Large Language Model, those large conversational AIs such as ChatGPT, Claude or Gemini) could function. And yet their existence is almost systematically obscured from public discourse on AI.
This article documents what this industry really covers in 2026, who these workers are, how they are paid, and what their exploitation says about the business model of contemporary AI.
The human supply chain that nobody shows
AI rests on a very specific supply chain. It always follows the same pattern. The American giants (OpenAI, Meta, Google, Microsoft, Anthropic) sign contracts with intermediaries (Sama, Scale AI, Majorel, Teleperformance, Remotasks). These intermediaries then recruit workers in countries with low wages and weak social protection.
The main players in this chain:
- Sama: a company based in San Francisco but with the bulk of its operations in Nairobi. For years it was the main moderation provider for OpenAI, Meta and several others. In 2023 it announced its withdrawal from content moderation to focus on image annotation, following several scandals.
- Scale AI: an American company valued at more than 14 billion dollars in 2025. Main clients: the Pentagon, OpenAI, Microsoft. Recruits massively in the Philippines via its subsidiary Remotasks.
- Majorel and Teleperformance: two French outsourcing giants, which have shifted massively towards AI contracts since 2023. Together they employ several hundred thousand people worldwide, with a growing fraction on AI annotation tasks.
Nairobi has become, within a few years, the world's leading hub for AI annotation. The Kenyan capital hosts operations for Meta, OpenAI, TikTok and many other companies. Why Kenya? Several reasons converge: a young and educated population, fluent English, low wage costs, and weak legal protection on questions of mental health at work.
According to a Time Magazine investigation published in January 2023, OpenAI subcontracted the labelling of toxic content for ChatGPT to Kenyan workers paid less than 2 dollars an hour, exposed daily to violent content, child sexual abuse, and descriptions of torture. The subcontracting company was Sama. The contract was terminated early by Sama after the public revelation, and several workers described lasting symptoms of post-traumatic stress. It was this investigation that turned the subject into an international case, but the system it describes is still running today under other names.
What do these workers actually do?
Four main categories of task structure the industry in 2026.
Image and video annotation. The worker looks at an image and labels what it contains: "red car", "pedestrian", "red light". These labels are used to train computer vision models (technology that allows AI to "understand" what it sees). Without these millions of manual labels, self-driving cars could not tell a stop sign from an advertisement. This task is generally the least traumatic, but also the worst paid: 1 to 3 cents per label, which works out at 0.50 to 2 dollars an hour depending on speed.
Content moderation. The worker looks at content reported by users (on Facebook, Instagram, TikTok) or generated by AI to check whether it contains prohibited material: graphic violence, sexual exploitation, harassment, extremist propaganda. This is the most traumatic task. Workers spend 8 hours a day looking at the worst of the Internet. Several clinical studies document PTSD (post-traumatic stress disorder, comparable to that of war veterans) rates reaching 30 to 50% among these workers after a few months.
Human reinforcement training. RLHF (Reinforcement Learning from Human Feedback, the main technique used to make ChatGPT polite and useful) requires humans to rank AI responses in order of quality. This work appears "light" but can expose evaluators to content produced by AI in "unfiltered" mode, sometimes extremely violent or pornographic.
The creation of synthetic dialogue. To train a chatbot to converse naturally, you need thousands of examples of well-written conversations. Workers write these conversations by hand, creating the examples that will teach the AI "how humans talk". These workers are the invisible scriptwriters of every chatbot you use. Several have reported that they were in fact "training" companion chatbots by sexting in real time with paying users, who believed they were talking to an AI. The exact term for this troubling practice is "human-in-the-loop deception".
Working conditions: what the investigations document
Several major journalistic investigations and academic studies converge on the real conditions in this industry. Here is what is factually documented.
On wages. In Kenya, annotators earn between $1.32 and $2 an hour for work done for OpenAI, according to Time Magazine. In the Philippines, Remotasks (a subsidiary of Scale AI) was the subject of a Washington Post investigation revealing allegations of non-payment of freelancers. In India, rates vary between $0.80 and $3 an hour depending on the task. In the United States, for the same tasks, the equivalent minimum wage would be $15 to $25 an hour.
On precarity. Almost none of these workers has a stable contract. Everything runs on the "gig economy" (workers are paid only for each task completed, with no sick leave, no paid holiday, no social security). In 2024, Remotasks withdrew abruptly from Kenya, leaving thousands of workers without income overnight, with nothing but a farewell email. As Kenyan Wall Street documented, this withdrawal was probably prompted by the emergence of local legal protections, which the company preferred to avoid by relocating.
On mental health. The psychological consequences are documented by several clinical studies and legal testimonies. Mophat Okinyi, a former Kenyan moderator, testified before the British Parliament in 2023 about the "images of sexual violence, torture, suicide" that he had to watch for 8 hours a day. He reported having developed severe depression, permanent nightmares, and the breakdown of a relationship as a result of his work. Several class actions are under way, notably in Kenya, against Meta and Sama, whose annotators are seeking compensation for psychological harm.
Several workers reported that they were not told at the point of recruitment what exactly the content they would be moderating would consist of. They were told "you'll be working on improving AI". They discovered on their first day that they would have to look at traumatic content. And that they were already signing a strict non-disclosure agreement that barred them from talking about it to their families. This is exactly what the AI ethics community has been condemning since 2022 under the name consent washing (the practice of obtaining formal consent without any real information about what the person is consenting to).
Why this structure persists: the brutal economic logic
One might think that as AI advances, the need for human labour would shrink. The opposite is happening. According to estimates from several analyst firms (Cognilytica, Gartner), the global data annotation market will grow from $2 billion in 2022 to more than $13 billion in 2028. The human workforce behind AI is not disappearing: it is growing.
Why? Three structural reasons.
First, the better AI gets, the more sophisticated the annotations it needs. At the start, a human was asked to label "dog" or "cat". Today, they are asked to assess whether an answer to a question of French criminal law is legally accurate and culturally appropriate. These tasks are infinitely more complex, therefore longer, therefore more costly, unless they are offshored to low-wage countries.
Second, use cases are diversifying. Every new application (self-driving cars, humanoid robots, conversational agents, medical AI) requires its own annotated datasets. Demand is growing exponentially.
Third, content moderation has become a major regulatory issue. With the European AI Act (the EU regulation on AI, which imposes transparency and safety obligations) and the American Take It Down Act (which we dissected in our article on deepfakes), platforms are legally obliged to moderate faster, and therefore to hire more moderators. These well-intentioned legal obligations paradoxically create growing demand for this exploited workforce.
The moral question we avoid
Here is the point the tech community finds very hard to face. The business model of today's generative AI rests largely on the exploitation of workers in the Global South. The Global South (a geopolitical term for the developing countries of Africa, Latin America, South-East Asia and Oceania, as opposed to the wealthy Global North) supplies the invisible raw material that allows an American or a European to generate an image in 3 seconds.
This structure is not new in the history of technology. The textile industry, consumer electronics, coffee, chocolate all rely on similar supply chains. What is new is that AI was presented as a post-industrial revolution, freed from physical constraints. Yet it remains deeply rooted in a highly conventional geography of globalised labour.
As we explored in our article on the environmental cost of AI, AI has two hidden footprints that nobody likes to look at: the physical footprint (water, electricity, land) and the human footprint (underpaid, traumatised, precarious labour). These two footprints form the invisible foundation on which the entire positive narrative of contemporary AI rests.
What could change (and what probably will not)
Several developments are under way, more or less credible.
The emergence of unions. In 2023, several Kenyan annotators formed the African Content Moderators Union, the first union dedicated to this profession. In 2025, similar efforts emerged in the Philippines and Colombia. These unions are still embryonic, but they are beginning to carry weight in contractual negotiations.
Multiplying legal challenges. In addition to the lawsuits against Sama and Meta in Kenya, several actions are under way in the United States (class actions over the psychological harm suffered by moderators), in the United Kingdom (duty of care owed by commissioning companies) and in Germany (joint liability for subcontracting conditions).
Gradual European regulation. The European Corporate Sustainability Due Diligence Directive (CSDDD, adopted in 2024) requires large companies to monitor working conditions in their supply chains, including for AI. Implementation is being phased in until 2028.
The partial automation of moderation. Paradoxically, AI could eventually automate part of content moderation, reducing the need for humans. But as several studies have shown, AI alone generates too many false positives and false negatives to fully replace humans. Human oversight will probably remain necessary for the coming decade.
Honestly, none of these developments will rapidly transform the situation. The business model works too well for the dominant players. And regulations, as we analysed in our article on the impossibility of fully regulating AI, have significant structural limits.
What you can do in practice
You are not responsible for OpenAI's supply chain. But a few actions remain within your reach if the subject matters to you.
Find out about your suppliers' practices. Anthropic, Mistral and some European labs document their annotation practices more openly than OpenAI. If you are a business choosing an AI supplier, you can build this criterion into your tenders.
Support investigative journalism. Time Magazine, the Washington Post, MIT Technology Review and Rest of World do crucial investigative work on these subjects. Subscribing to these outlets is a concrete action.
Talk about it. The silence around these workers is one of the main mechanisms that allow this industry to thrive. Raising it in professional conversations, mentioning it in presentations about AI, bringing it into public debate: this is probably the most useful action an individual can take.
Stay nuanced. For many workers in Kenya or the Philippines, these jobs, even badly paid, are among the best economic opportunities available locally. Naively boycotting these services could paradoxically harm them. The solution lies in improving conditions, not in the disappearance of this industry.
The mirror this industry holds up to us
Contemporary AI is often presented as an unprecedented technological rupture. In many respects, it is. But on the question of labour, it extends structures that are 200 years old. During the industrial revolution, European textiles relied on cotton picked by American slaves. During the digital revolution, consumer electronics relied on underpaid Chinese workers. During the AI revolution, frontier models rely on Kenyan, Filipino and Indian annotators exposed to the worst content on the web.
AI in 2026 has made the political choice, conscious or not, to reproduce these models of exploitation rather than challenge them. As we explored in our article on the disruption of graphic designers and in our analysis of the hidden economy of AI companions, the question of labour in the age of AI deserves far deeper societal reflection than is currently taking place.
These workers exist. There are hundreds of thousands of them. They make possible every "polite" answer from ChatGPT, every "clean" image from Midjourney, every "reliable" recommendation from Claude. The least we can do is name their existence and look at the human bill for our technological comfort. It solves nothing immediately. But it is the mandatory starting point for any future solution.