In this article
Today we look at the case that most cleanly proves generative artificial intelligence is neither ethereal nor dematerialized. Behind every kind reply from ChatGPT are real workers who, for months, read the worst of the internet —child sexual abuse, torture, self-harm, murder— to teach the system what to block. They were paid between 1.32 and 2 dollars an hour. They were in Nairobi, employed by a U.S. subcontractor called Sama. TIME's investigation in January 2023 documented the conditions. Almost nobody in Spain read it in full, and the few who did filed it as a moral anecdote instead of a structural piece. My take is biased because I've read the transcripts of the testimony and they're hard. Form yours calmly.
I've kept TIME's report in a drawer for two years. I take it out when someone tells me AI is "magic" or that "it trains itself." Neither claim is true. Contemporary AI rests on massive human labor and, in its safety layer, on traumatic human labor. Let's go through it slowly because the figures and the names matter.
What exactly the Kenyan workers did
The original report was written by Billy Perrigo and published by TIME on January 18, 2023, under the title OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic. It's a long, solid text, documented with leaked contracts, internal emails and direct testimony from workers. What follows is the summary of the central facts.
OpenAI needed, during 2021 and 2022, to build a safety filter to detect and block toxic content in ChatGPT's responses. To train that filter it needed a dataset of labeled examples: thousands of text fragments pulled from the internet, each classified by humans according to the type and severity of its toxicity. That human classification is what then allowed the machine-learning model to learn to discriminate.
OpenAI subcontracted the task to Sama, a U.S. company founded in 2008 by Leila Janah, headquartered in San Francisco with labeling operations in Kenya, Uganda, India. Sama describes itself as a social-mission company that "makes the digital economy ethical." It also works with Meta, Google, Microsoft, Amazon, Apple and other corporate clients.
Sama signed three contracts with OpenAI between November 2021 and March 2022 for a total of roughly 200,000 dollars. The contracts specified the supply of staff to label text fragments with descriptions of child sexual abuse, murder, suicide, torture, self-harm, incest and graphic violence. The Kenyan workers hired, mostly young men and women with university studies and no better labor options in the local market, read thousands of these fragments in 9-hour shifts, classified them in a web interface, and reported their work to intermediate supervisors.
The wage was between 1.32 and 2 dollars an hour, depending on level of experience and type of task. A full-time worker earned between 170 and 200 dollars a month. Sama terminated the contracts ahead of the planned schedule in March 2022, after a group of workers filed a complaint about the working conditions and the content they were exposed to without adequate psychological support.
The human consequences
The testimony gathered by Perrigo and the later reporting document serious consequences.
Several workers described symptoms consistent with post-traumatic stress disorder: recurring nightmares, panic attacks, difficulty concentrating, emotional distance from their families, loss of appetite. Some sought psychological care on their own initiative, paid out of their own wages. Others simply endured it. The mental-health services Sama promised contractually were, according to testimony, brief sessions with counselors lacking clinical training in vicarious trauma.
One of the workers, Mophat Okinyi, told the Wall Street Journal and later TIME that his marriage broke down as a direct consequence of the work, and that he suffered serious episodes of social isolation after leaving the company. Others recounted similar experiences with varying degrees of severity.
In January 2023, the Kenyan workers filed a formal petition to the Kenyan Parliament requesting specific regulation of content-moderation work. In July 2023, a group of former Sama employees filed a class-action suit against Meta and against Sama over working conditions in earlier content-moderation projects for Facebook. The case continues in Kenyan courts. In November 2024, other workers also filed a specific suit against OpenAI and Sama over the conditions of the labeling work for ChatGPT, also ongoing.
The corporate justification, and what it omits
OpenAI and Sama issued statements after the TIME report was published. The positions summarize roughly as follows.
OpenAI held that the labeling task was necessary to build a safe product, that it trusted Sama to manage the workers' wellbeing in accordance with contractual standards, and that after the investigation it tightened its criteria for selecting subcontractors. Sama declared that it paid above the minimum wage and above Kenya's average wage (which is factually true: the minimum wage in Kenya is roughly 0.45 dollars an hour), that the work was voluntary, and that it offered psychological support during the contract.
The workers' testimony contradicts the second and third points. Voluntary in the sense that no one signed under coercion, yes; voluntary in the real economic sense —that is, with a realistic possibility of refusing the offer without falling into a severe worsening of the family situation—, no. The psychological support existed formally but wasn't clinically adequate for vicarious trauma, according to the descriptions gathered.
The first point —that the wage exceeded Kenya's average wage— is true but structurally irrelevant. An objectively traumatic job doesn't become ethical because it pays twice the average wage of the country where it's performed. If the task requires continuous exposure to the most violent content on the internet during 9-hour shifts, the standards of labor and psychological protection must be equivalent to those applied in other sectors that handle similar material —police forensics, prosecution, war journalism, social work with abuse victims—. In those sectors, the protocols include mandatory rotation, daily exposure limits, continuous clinical supervision, the right to refuse specific tasks without penalty, structured psychological support for vicarious trauma. None of those standards was applied in Sama's work.
The structural, not exceptional, pattern
The OpenAI/Sama case isn't an exception in the field. It's a paradigmatic example of an industrial model the tech sector has used for two decades.
Meta —formerly Facebook— has used similar subcontractors for content moderation on Facebook and Instagram. Its operations in Kenya, also with Sama, gave rise to the 2023 class action mentioned above. Similar operations in the Philippines and Morocco have been documented by Time, The Verge, The New Yorker, MIT Technology Review and 404 Media over the past decade.
TikTok maintains content-moderation operations in India, the Philippines and Morocco with a comparable industrial model. The Wall Street Journal's 2023 investigations documented similar working conditions for the moderators who filter content uploaded to the platform.
Microsoft, Google, Amazon, Apple subcontract part of their annotation and moderation work through platforms like Mechanical Turk, Appen, Lionbridge and other intermediaries. The value chain is opaque: the end client rarely knows the operational detail of the subcontractor.
The pattern is uniform. The value chain of contemporary generative AI rests on a layer of massive human labor offshored to Global South countries with lower labor costs and lower regulatory protection. That work is invisible to the product's end user, doesn't enter press releases, doesn't appear in corporate sustainability reports. It exists, nonetheless, and without it the products wouldn't exist.
Mary L. Gray and Siddharth Suri documented this thoroughly in Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (Houghton Mifflin Harcourt, 2019). The book precedes the generative-AI boom but describes precisely the structure that later spread to the frontier field. The term "ghost work" captures the invisible character of the human effort that sustains seemingly automated services.
What happens when a user connects
Here comes the operational connection the case illuminates. Every time a user of ChatGPT, Claude, Gemini or any other conversational assistant asks something the system decides to refuse —"I can't help you with that," "that request isn't appropriate," "please rephrase your question"—, that refusal is a function of safety filters trained on manually labeled datasets. Datasets like the one the Kenyan Sama workers produced for OpenAI.
The "safety" of the model, in operational terms, isn't a property of the model. It's a property of the filtering layer added to it. And that filtering layer exists thanks to human workers who exposed themselves to traumatic content in precarious working conditions to produce the labels the model learned.
This isn't so the user feels guilt every time they converse with a chatbot. Individual guilt resolves nothing. It's so the user knows that the product that looks dematerialized has concrete human costs in concrete places. That awareness should translate into political and regulatory demand, not into personal distress.
The political question
This is personal opinion, but the documents of the case support it. The offshoring of grueling labor tasks to countries with lower regulatory protection isn't exclusive to the AI sector. It has been a practice of global capitalism for centuries. What the case of the Kenyan annotators adds is an extra layer: the offshoring of the psychologically traumatic work associated with the invisible underpinning of mass-consumer digital products.
The European public conversation about regulating AI —around the AI Act, around platform taxation, around data protection— has focused almost exclusively on protecting European citizens as users or as data subjects. It has barely considered how to regulate the labor practices of the value chains that sustain the products those citizens use. That's a visible asymmetry.
What I'd ask for —and here with some realism because there's an available international framework— is that the European Union require a mandatory audit of the labor-subcontracting chains of AI companies operating in the European market, with public disclosure of the labeling providers, the countries of operation, the real wages by task level and the mental-health protocols applied. This isn't revolutionary; it exists for textiles and for cocoa, with imperfect but measurable effects. Applying it to AI is a natural extension.
In the meantime, the individual user can at least know that when a chatbot refuses them a raw question, someone in Nairobi, Manila, Bogotá or Karachi exposed themselves to worse during a nine-hour shift so that refusal sentence could exist. It's operational information, not a pretext for drama.
The concrete figure to close on. According to Billy Perrigo's investigation in TIME (January 18, 2023), confirmed by leaked contractual documentation and by testimony from at least four workers with protected identities, the Kenyan Sama employees who worked on the OpenAI project reviewed between 150 and 250 text fragments per 9-hour shift over a period of roughly four months, earning between 1.32 and 2 dollars an hour depending on level. The concrete figure —between 1.32 and 2 dollars per hour of continuous exposure to the worst textual content the internet produces— is the operational translation of the real human cost of the product every ChatGPT user consumes for free or for twenty euros a month. That imbalance is exactly the pattern the field needs to correct if it intends to call itself responsible.
Definitions
Data annotation: the human task of classifying, labeling or annotating fragments of text, image, audio or video so that a machine-learning model can be trained on them. It's massive work and, for sensitive data, psychologically demanding.
Vicarious trauma: a psychological disorder caused by continuous exposure to accounts or images of violence suffered by third parties. It's documented in healthcare staff, war journalists, social workers and, more recently, digital content moderators.
Ghost work: invisible human labor that sustains seemingly automated digital services. A concept developed by Mary L. Gray and Siddharth Suri in their eponymous 2019 book.
Subcontracting chain: a structure by which a client company subcontracts tasks to an intermediary, which in turn subcontracts to a second intermediary, hiding from the end client the real conditions in which the work is performed.
References
Billy Perrigo, OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic (TIME, January 18, 2023). The original investigation that documented the case with leaked contracts and testimony.
Billy Perrigo, Inside Facebook's African Sweatshop (TIME, February 2022). The earlier investigation into Sama's operations for Meta in Nairobi, a direct precedent of the OpenAI case.
Mary L. Gray & Siddharth Suri, Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (Houghton Mifflin Harcourt, 2019). A theoretical and empirical framework on the invisible labor that sustains the digital economy.
Sarah T. Roberts, Behind the Screen: Content Moderation in the Shadows of Social Media (Yale University Press, 2019). Academic research on the conditions of content moderators on digital platforms.
Karen Hao, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI (Penguin Press, 2025). In-depth reporting on OpenAI with specific chapters on the labeler chain.
International Labour Organization (ILO), World Employment and Social Outlook: The Role of Digital Labour Platforms in Transforming the World of Work (International Labour Organization, 2021). An institutional framework on offshored digital labor.
Kate Crawford, Atlas of AI (Yale University Press, 2021). Chapters on the labor and geographic materiality of artificial intelligence.
Further reading
Sarah T. Roberts, Behind the Screen (Yale University Press, 2019). The canonical academic text on content moderation as invisible labor.
404 Media (404media.co). Continuous coverage of the working conditions in the AI value chain, with recurring field investigations.
Foxglove Legal (foxglove.org.uk). A British NGO that legally represents content moderators in various countries; public documentation on the ongoing cases.
Wired (wired.com). Recurring coverage since 2022 of the annotator and moderator chain in the Global South, with specific pieces on Kenya, the Philippines, Venezuela and Colombia.
You might also like
- Human feedback as training
- Why did Pegasus turn mass spying into routine?
- The energy cost of a conversation
Comments0
No comments yet.
Leave a comment