The Invisible Builders of AI: Displaced by the Machines They Trained

In a low-rise office in an industrial corner of Nairobi, a man in his thirties sat in front of a screen for nine hours a day, five days a week, reading the worst things human beings write about one another. The descriptions arrived as fragments of text, stripped of context, scrolling past in their thousands: child sexual abuse, bestiality, torture, suicide, incest, murder rendered in clinical detail. His job was to read each one, decide what category of horror it belonged to, and label it. Not to remove it from the world, but to teach a machine to recognise it, so that the machine would learn never to produce anything like it.

His name is Richard Mathenge, and he was a team leader at an outsourcing firm called Sama, working on a contract for a then-obscure San Francisco company called OpenAI. The model he was helping to train would, within months, become ChatGPT, the fastest-growing consumer application in history and the centrepiece of a technology now valued in the hundreds of billions of dollars. For this work, Mathenge and his colleagues took home somewhere between $1.32 and $2 an hour. When a Time magazine investigation surfaced the arrangement in January 2023, the company's spokespeople described the labour as essential, even noble. The data labellers themselves described recurring nightmares, fractured relationships, and a trauma that followed them home.

Mathenge would later be named to Time's first list of the hundred most influential people in artificial intelligence, sharing the page with Sam Altman, the chief executive of the company whose product his labour helped to make safe. The juxtaposition is almost too neat. The man at the top of the value chain and the man at the bottom of it, recognised in the same breath, separated by an ocean and an income gap measured in orders of magnitude. One is celebrated as a founder of the future. The other was, until very recently, not acknowledged as a builder at all.

This is the story the AI industry does not tell about itself. The large language models and image generators now reshaping white-collar work across the wealthy world did not emerge fully formed from clusters of graphics cards. They were taught, painstakingly and at enormous human cost, by hundreds of thousands of people in Kenya, the Philippines, India, and Venezuela, working under gig contracts that offered no benefits, no security, and no claim whatsoever to the products their labour made possible. And now, in one of the more brutal ironies of the present moment, the industry has begun to automate the very tasks those workers performed, displacing the people who trained the machines with the machines they trained. They were never called builders. They are being made redundant before the public has even registered that they exist.

The Hidden Production Line

To understand what these workers actually do, it helps to abandon the prevailing mythology of artificial intelligence as something that learns on its own. Modern AI systems are statistical engines trained on vast quantities of data, and that data, to be useful, must be organised, labelled, filtered, and rated by human beings. The work falls into roughly three categories, each invisible in a different way.

The first is data annotation: the labelling of raw information so that a machine can learn to interpret it. A worker might draw boxes around pedestrians and traffic lights in thousands of street images to train a self-driving system, or transcribe audio, or tag the objects in a photograph, or mark which sentences in a paragraph express a particular sentiment. It is repetitive, exacting, and mind-numbingly granular. The anthropologist Mary L. Gray and the computer scientist Siddharth Suri, in their 2019 book Ghost Work, gave this hidden labour its enduring name. Ghost work, they wrote, is work performed by a human but experienced by the customer as the seamless output of an automated process. The whole commercial proposition of AI depends on the human hand remaining unseen.

The second category is reinforcement learning from human feedback, or RLHF, the technique that turned raw language models into the fluent, agreeable conversational systems the public now uses. The mechanics are deceptively simple. A model generates several possible responses to a prompt; a human worker reads them and rates which is best, which is more accurate, which is less likely to cause offence. Multiplied across millions of comparisons, this human judgement is what teaches a model to sound helpful, to refuse harmful requests, to behave. Every well-mannered chatbot is, in this sense, a monument to the aggregated preferences of low-paid raters, most of whom will never be named.

The third category is content moderation, the work Mathenge and his colleagues performed, and it is the most psychologically corrosive of the three. To build a filter that keeps an AI system from generating descriptions of child abuse or graphic violence, the system must first be shown what those things look like, tagged by humans who absorb the material so the machine does not have to. The labour is a kind of sacrificial buffer. The worker reads the unreadable so that the user never will.

The geography of this work is not accidental. The industry gravitated towards places where wages are low, education is high, English is widely spoken, and economic desperation makes a dollar an hour seem, if not generous, at least survivable. Kenya became a hub for content moderation and labelling through firms such as Sama. India and the Philippines supplied enormous volumes of annotation and moderation labour. And Venezuela, after the collapse of its oil-dependent economy, became one of the largest sources of data workers on the planet.

The Venezuelan case is the starkest illustration of how the supply chain feeds on crisis. As MIT Technology Review documented in 2022, when Venezuela's economy imploded after the 2014 oil price crash, with hyperinflation reducing monthly salaries to a few dollars, a population of educated, internet-connected people suddenly found themselves available for piecework at almost any price. Platforms such as Appen and Remotasks, which acted as intermediaries for Silicon Valley giants, rushed in. Researchers including Julian Posada have detailed how Venezuela came to host one of the largest data-work populations in the world, with workers labelling images for self-driving cars and rating AI outputs from their homes, competing in real time against labellers across the globe for tasks that might pay a few cents each. Rest of World reported that even older Venezuelans, pushed out of formal employment by age discrimination and inflation, turned to annotation work to survive. The catastrophe did not merely coincide with the AI boom. It was, in a precise economic sense, an input to it.

The Cost of Reading the Unreadable

The phrase “content moderation” sounds bureaucratic, almost gentle, which is part of how the harm stays hidden. The reality is documented now in clinical language, in court filings, in psychiatric diagnoses.

In December 2024, lawyers acting for former Facebook content moderators in Kenya filed medical evidence with the Nairobi employment and labour relations court. Of 144 moderators who underwent psychological assessment, out of 185 involved in the legal claim, 81 per cent were classified as suffering from severe post-traumatic stress disorder, according to reporting by CNN. The diagnoses were made by Dr Ian Kanyanya, head of mental health services at Kenyatta National Hospital. These were people who had spent their working days viewing what the filings describe as necrophilia, child sexual abuse, terrorism, and graphic violence, hour after hour, with productivity targets that left little room to look away.

The harm does not end when the contract does. The Sama work for OpenAI was terminated in February 2022, some eight months ahead of schedule, when the traumatic nature of the material became untenable. But the workers who had performed it carried the consequences forward. Speaking to Slate in 2023, Mathenge described continuing to suffer from the explicit content he had been required to read long after the project closed, the images lodged in memory with no mechanism for their removal. The Time investigation found that the wellness counselling offered to workers was thin and that the demands of the job left little room to make use of even what was provided. The trauma was, in effect, an externality: a cost generated by the work, borne entirely by the worker, and absorbed nowhere on the balance sheet of the companies that benefited from it.

The Kenyan moderators became, almost by accident, the front line of a global reckoning. In 2022, a former Sama employee named Daniel Motaung launched what his lawyers at the British legal non-profit Foxglove called a world-first case against Meta and Sama. Motaung alleged that the work had inflicted serious psychological harm, that the pay was inadequate, and that when he attempted to organise his colleagues into a trade union, he was fired for it. Meta argued that it could not be sued in Kenya, since the moderators were formally employed by Sama, its outsourcing contractor. The Kenyan courts disagreed. In 2023 and again in 2024, the country's courts ruled that the cases could proceed to trial, rejecting Meta's appeals. It was a significant crack in the legal architecture that the technology industry had relied upon: the assumption that outsourcing the labour also outsourced the liability.

Out of this struggle came something the industry had not anticipated. In May 2023, around 150 workers from several outsourcing firms voted to form the African Content Moderators Union, the first organisation of its kind on the continent, a cross-company effort to win better conditions for the people performing this work. Mathenge, the OpenAI team leader, was among its organisers. Having been recognised by Time as one of the most influential figures in AI, he used the platform not to celebrate the technology but to demand that its human foundations be treated with dignity. He and other former ChatGPT moderators went on to petition the Kenyan parliament, calling for an investigation into OpenAI and Sama and for regulation of the outsourcing industry.

The psychological toll is not a side effect of poorly managed work. It is structural. The journalist Karen Hao, whose 2025 book Empire of AI traces these supply chains in detail, found that the trauma rippled outward from the workers into their families and communities, a slow contamination that no severance package was ever designed to address, because severance packages were rarely offered at all. Hao's central argument is that the AI industry functions as a kind of empire, extracting value, labour, and raw material from the Global South to enrich a small constellation of companies in the Global North, much as the colonial empires of the past extracted rubber, minerals, and bodies. The content moderators of Nairobi are, in this framing, not an unfortunate footnote to the AI revolution. They are its colonial subjects.

The Architecture of Invisibility

Why is this labour absent from the public conversation about AI? The answer is not simple neglect. The invisibility is engineered, sustained by a set of structural mechanisms that operate together to keep the human contribution out of view, off the balance sheet, and outside the conversation about who owns what.

The first mechanism is the outsourcing chain itself. A worker in Nairobi is rarely employed by OpenAI or Meta. They are employed by Sama, or by a labour broker contracted by Sama, who is in turn contracted by the technology company. Each link in the chain creates legal and moral distance. The company at the top can truthfully say it does not employ the worker at the bottom; the worker at the bottom has no contractual relationship with the company whose product they are building. When Meta argued in the Kenyan courts that it could not be held responsible because the moderators worked for Sama, it was deploying this distance as a legal shield. The whole structure is designed so that responsibility evaporates somewhere in the middle.

The second mechanism is the gig contract. These are not jobs in the traditional sense. They are tasks, parcelled out through platforms, paid by the piece, with no guaranteed hours and no benefits. Gray and Suri described in Ghost Work how this arrangement strips workers of nearly every protection associated with employment: no health cover, no job security, dismissal at any moment for any reason or none. Work appears unpredictably and vanishes the same way. When a project ends, or when workers organise, the tasks simply stop arriving. There is no factory gate to picket, no employer to confront, only a platform that has gone quiet.

The third and most consequential mechanism concerns intellectual property and value. The workers who train a model acquire no stake in it. Their labour is embedded in the model's weights as surely as an engineer's design is embedded in a bridge, but the legal framework treats their contribution as a service rendered and paid for, terminating all further claim. The model becomes the property of the company. Its value, which may run into hundreds of billions of dollars, accrues entirely upward. This is the asymmetry that the researcher Milagros Miceli, who leads the Data Workers' Inquiry at the Distributed AI Research Institute founded by Timnit Gebru, has spent years documenting. The Data Workers' Inquiry is a participatory research project spanning multiple countries, built on the principle that the people who do this labour are the foremost experts on it and should be the ones to describe it. Its findings are consistent across continents: the workers understand precisely how essential they are, and precisely how little of the resulting wealth flows back to them.

The pattern repeats wherever investigators look. A guide published by the Global Investigative Journalism Network in January 2026, drawing together the reporting of numerous outlets into labour abuses across the data annotation industry, maps the same recurring features from one country to the next: opaque subcontracting, wages pinned at or below local minimums, abrupt terminations, and a near-total absence of mechanisms by which workers might contest any of it. What the reporting it gathers makes plain is that these conditions are not the failings of a few bad operators but the standard operating logic of a global supply chain, reproduced across jurisdictions precisely because the model is profitable and the workers are interchangeable. The abuse is not a bug in the system. It is the system functioning as designed.

There is a deeper conceptual sleight of hand at work, too. The dominant narrative of artificial intelligence is one of machines that learn by themselves, that improve through training as though training were a natural process rather than a human one. This framing is not merely inaccurate; it is convenient. If the machine learns on its own, then no one is owed anything for teaching it. The erasure of the human labeller is the precondition for the myth of the autonomous machine. Every time a technology executive describes a model as having “learned” some capability, the millions of human judgements that produced that capability are quietly written out of the sentence.

The Industry Eats Its Own Foundations

This brings us to the cruellest turn in the story, and the one that gives the present moment its peculiar moral weight. The industry that was built on this hidden human labour has begun, methodically, to automate it away.

The logic is internal to the technology. Once a language model is sufficiently capable, it can be turned upon the very tasks that humans once performed to create it. A model can be used as a judge to rate the outputs of another model, the so-called LLM-as-judge technique, replacing the human raters whose comparisons taught the first generation of systems. A model can generate synthetic training data, fabricating examples that once had to be labelled by hand. Industry accounts of the field describe a layered workflow that has rapidly become standard: automated annotators handle the first pass of labelling, synthetic generators fill the rare-case gaps, and human beings are pushed to the narrow margin of borderline and policy-sensitive cases. The vast middle of the labelling economy, the repetitive volume work that sustained tens of thousands of livelihoods, is being absorbed by the machines themselves.

A paper presented to the arXiv preprint server in May 2026 by the researchers Robert Wolfe and Aayushi Dangol, catalogued as arXiv:2605.03295, mapped the ideology of this transition. Examining the public communications of five data annotation companies and their executives, the authors found that the industry increasingly frames human expertise as economically inefficient, something to be liberated, reformed, or replaced by cheaper AI-driven approaches that promise a superior return on investment. The market for what the paper calls “cheap expertise” is being reorganised around the premise that the human contribution, having served its purpose, is now a cost to be minimised. The brief that prompted this article frames the same finding in blunter terms: the industry is automating its own lowest-tier annotation tasks, which means the workers who trained the models are now being displaced by them.

The timing of this transition cannot be separated from the wave of capital that flooded the sector in 2025. The defining event was Meta's deal with Scale AI, the data-labelling company that had built its business by paying people across the world to annotate training data. In June 2025, Meta agreed to invest $14.3 billion for a 49 per cent stake in Scale, a transaction that valued the firm at over $29 billion and brought its 28-year-old founder, Alexandr Wang, into Meta as its chief AI officer. The deal was a vivid statement of what the market values. The infrastructure of human labelling had become worth tens of billions of dollars, and that value was crystallised in the person of a founder and a corporate balance sheet. None of it flowed to the labellers. The same companies that profited from arranging the work now have every commercial incentive to automate it, and the capital to do so.

The displacement this section describes is no longer a hypothetical inference from workflow diagrams. In April 2026, Sama, the very outsourcing firm that had run the OpenAI moderation work in Nairobi, issued formal redundancy notices to 1,108 workers at its Nairobi delivery centre. The layoffs followed Meta's decision to terminate its content and data-annotation contract with Sama, and they came after the firm had already pivoted away from content moderation to refocus its business on AI data labelling. It is a compact illustration of the whole argument. The same company named at the top of this article, the one that supplied the labour that made ChatGPT usable, has now shed more than a thousand workers as a client relationship collapsed and the work reorganised itself around the automated labelling that the industry now prefers. The loop closes on the people who were at the centre of it.

Consider, then, the full shape of what has happened. A worker in Nairobi or Caracas or Manila spends years teaching a model to recognise harmful content, to rate good answers above bad ones, to behave like something worth trusting. The model improves to the point where it can perform those judgements itself. The worker is no longer needed. There is no severance, because there was never employment. There is no retraining, because there was never an obligation. There is no acknowledgement that their labour is embedded in a product worth a fortune, because the entire legal and narrative architecture was constructed to deny precisely that. The people who built artificial intelligence are among the first to be made redundant by it, and they were never recognised as having built anything at all.

What the Conversation Leaves Out

Step back, and a question emerges that is larger than any single supply chain. The public debate about artificial intelligence has settled into a familiar frame: benefits versus harms. On one side, the promise of productivity, scientific discovery, abundance. On the other, the fears of job displacement, misinformation, existential risk. It is a debate conducted almost entirely in the future tense, and almost entirely about the wealthy world.

The story of the data workers reveals what this frame leaves out. The displacement is not a future risk; it has already happened, to the workers least equipped to absorb it and least likely to be counted in the statistics. When commentators warn that AI will displace workers, they generally mean knowledge workers in the Global North: lawyers, coders, copywriters, illustrators. They rarely mean the Kenyan moderator or the Venezuelan labeller, whose displacement is already underway and who possessed none of the cushions, the savings, the safety nets, the alternative prospects that make displacement survivable. The benefits-versus-harms debate is conducted on behalf of people who stand to lose something they currently have. It has remarkably little to say about people who built the entire edifice and were never permitted to own a brick of it.

The frame also obscures the question of distribution. To ask whether AI is, on balance, good or bad is to assume a single ledger on which gains and losses can be netted against one another. But the gains and the losses fall on different people. The productivity accrues to shareholders and consumers in wealthy economies. The trauma, the precarity, and now the redundancy accrue to data workers in the Global South. A technology can be enormously beneficial in aggregate and profoundly unjust in its distribution, and the aggregate framing is designed, whether by intent or by habit, to make the injustice disappear into an average. Karen Hao's image of empire is useful precisely because empires, too, generated immense aggregate wealth while distributing it with savage inequality.

And the frame leaves out the matter of acknowledgement itself, which is not merely sentimental. To recognise these workers as builders would be to admit that they have a claim, and claims have consequences. They might imply ownership, or royalties, or a share of the value, or at the very least the protections owed to employees. The refusal to acknowledge is therefore not an oversight but a defence. The conversation about AI's benefits and harms leaves out its builders because including them would change the answer to the question of who owes what to whom. It is easier to debate whether the machine is good or bad than to ask who paid for it with their mental health and their labour, and what they are owed in return.

There is something the workers themselves understood before any of the commentators did. They knew they were building something. The petitions to the Kenyan parliament, the formation of the African Content Moderators Union, the lawsuits, the testimony gathered by the Data Workers' Inquiry: all of it is the work of people insisting on a fact the industry preferred to suppress, that they are not bystanders to artificial intelligence but its authors. The conversation that leaves them out is not merely incomplete. It is, in a precise sense, a conversation about a product without any mention of the people who made it.

Toward Recognition

If the diagnosis is structural invisibility, the responses now emerging are attempts to make the labour visible, and to attach consequences to that visibility. None is a complete remedy. Together they sketch the outline of what a fairer arrangement might look like.

The first and most direct is organisation. The African Content Moderators Union demonstrated that workers separated by employer, geography, and the architecture of outsourcing can nonetheless find one another and act collectively. Unionisation cuts against the central design feature of ghost work, its atomisation, by reconstituting the workforce that the gig model was built to dissolve. The Kenyan lawsuits, supported by Foxglove, have pursued the same goal through the courts, establishing the precedent that a technology company cannot necessarily hide behind its outsourcing contractors. If that principle holds, the legal distance on which the whole structure depends begins to collapse.

The second response is the establishment of standards and the public ranking of companies against them. The Fairwork project, based at the Oxford Internet Institute and the Berlin Social Science Centre, assesses platforms against five principles: fair pay, fair conditions, fair contracts, fair management, and fair representation. Its investigations of cloudwork and microwork platforms, including Scale AI and Amazon Mechanical Turk, found that in most cases researchers could find no evidence that companies ensured workers earned even their local minimum wage, and that a substantial share of workers reported completing tasks for which they were never paid. By scoring companies publicly, Fairwork attempts to convert invisibility into reputational pressure, to make poor labour standards a liability rather than a hidden efficiency. From this work flows the proposal for something like fair-trade certification for data, a label that would allow buyers of AI systems, and eventually their users, to distinguish products built on decently paid labour from those built on exploitation.

The third response is regulatory, and it is the least developed. The Kenyan parliament's consideration of the moderators' petition, the slow construction of legal precedent through the courts, and the broader push to extend labour protections along the AI supply chain all point towards a future in which states assert that work performed within their borders is subject to their laws, regardless of how many corporate layers separate the worker from the ultimate beneficiary. The challenge is jurisdictional: the work is global, the value flows across borders, and no single regulator commands the whole chain. But the principle that liability cannot be outsourced, established in the Kenyan courts, is the seed of a more ambitious regulatory architecture.

The fourth response is conceptual, and the most radical. It is the argument, advanced by the technologist Jaron Lanier and the economist Glen Weyl under the banner of data dignity, that data should be understood as labour. If the information people produce, and by extension the judgements that workers contribute to training a model, are forms of labour rather than free raw material, then their extraction without fair compensation is a species of exploitation, and those who provide them are owed a share of the value they create. Lanier and Weyl proposed intermediary institutions, data trusts of a kind, that might negotiate on behalf of data providers against the concentrated power of the technology companies. Applied to the data workers of Nairobi and Caracas, the data-as-labour argument arrives at a conclusion the industry has spent years avoiding: that the people who trained the models have a legitimate claim on the wealth those models generate, not as charity, but as a matter of right.

None of these responses, on its own, dismantles the asymmetry. A union can win better conditions but cannot easily reach across borders to the company at the top of the chain. A certification scheme depends on buyers who care. Regulation runs into the limits of jurisdiction. The data-as-labour argument remains, for now, largely theoretical. But each represents a refusal to accept the central premise on which the invisibility was built, the premise that this labour does not count, that the people who perform it are not builders, and that the value they create belongs to someone else by natural law.

The image of Richard Mathenge on the Time 100 list, beside the founders and the executives, is worth returning to. It was meant, perhaps, as a gesture of inclusion, a recognition that the people at the bottom of the chain belong in the story of artificial intelligence. But it also exposes the gap it was meant to bridge. Mathenge earned that recognition not by building the technology, which he did, but by fighting to be acknowledged for it, which the technology's beneficiaries never required of themselves. He had to organise a union, file a petition, and testify to his own trauma in order to be seen. The founders were simply seen.

The machines that are now displacing workers across the wealthy world learned to do so from people who were displaced first, who taught them, and who were never thanked. If the conversation about artificial intelligence continues to take place as though those people did not exist, it will not merely be incomplete. It will be repeating, in the register of public discourse, the same erasure that the outsourcing contracts and the gig platforms and the myth of the autonomous machine were built to perform. The first task of an honest reckoning with AI is to name its builders. They have already named themselves. The question is whether the rest of the conversation is willing to listen.

References

  1. Perrigo, Billy. “Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic.” Time, 18 January 2023. https://time.com/6247678/openai-chatgpt-kenya-workers/
  2. Gray, Mary L., and Siddharth Suri. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt, 2019. https://ghostwork.info/
  3. Hao, Karen. Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI. Penguin Press, 2025. https://restofworld.org/2025/karen-hao-empire-of-ai-book/
  4. Williams, Adrienne, Milagros Miceli, and Timnit Gebru. “The Exploited Labor Behind Artificial Intelligence.” Noema Magazine, 13 October 2022. https://www.noemamag.com/the-exploited-labor-behind-artificial-intelligence/
  5. “How the AI Industry Profits from Catastrophe.” MIT Technology Review, 20 April 2022. https://www.technologyreview.com/2022/04/20/1050392/ai-industry-appen-scale-data-labels/
  6. “Amid Economic Collapse, Older Venezuelans Turn to Gig Work.” Rest of World, 2024. https://restofworld.org/2024/venezuela-online-gig-work/
  7. “Facebook Inflicted 'Lifelong Trauma' on Content Moderators in Kenya, Campaigners Say, as More Than 140 Are Diagnosed with PTSD.” CNN Business, 22 December 2024. https://edition.cnn.com/2024/12/22/business/facebook-content-moderators-kenya-ptsd-intl
  8. Foxglove. “Daniel Motaung Launches World-First Case to Force Facebook to Finally Put Content Moderators' Health and Wellbeing Ahead of Cash.” 10 May 2022. https://www.foxglove.org.uk/2022/05/10/motaung-facebook-world-first/
  9. Foxglove. “Big Win in Kenya! 185 Former Facebook Content Moderators to Take Their Case Against Mass Firing to Trial After Courts Slap Down Meta Appeal.” 23 September 2024. https://www.foxglove.org.uk/2024/09/23/facebook-content-moderators-kenya-meta-appeal/
  10. “150 African Workers for AI Companies Vote to Unionize.” Time, 1 May 2023. https://time.com/6275995/chatgpt-facebook-african-workers-union/
  11. “Richard Mathenge: The 100 Most Influential People in AI 2023.” Time, 7 September 2023. https://time.com/collection/time100-ai/6308980/richard-mathenge/
  12. “A ChatGPT Contractor Who Trained the A.I. in Kenya Says He's Still Traumatized.” Slate, May 2023. https://slate.com/technology/2023/05/openai-chatgpt-training-kenya-traumatic.html
  13. Distributed AI Research Institute. “Data Workers' Inquiry.” DAIR, led by Milagros Miceli. https://dair-institute.org/projects/data-workers-inquiry/
  14. Fairwork / Oxford Internet Institute. “Fairwork Cloudwork Ratings: Labour Standards in the Cloudwork Economy.” University of Oxford and WZB Berlin Social Science Centre. https://www.oii.ox.ac.uk/research/projects/fairwork-for-ai/
  15. Wolfe, Robert, and Aayushi Dangol. “Cheap Expertise: Mapping and Challenging Industry Perspectives in the Expert Data Gig Economy.” arXiv:2605.03295, May 2026 (forthcoming, CHIWORK 2026). https://arxiv.org/abs/2605.03295
  16. “Scale AI Founder Wang Announces Exit for Meta as Part of $14.3 Billion Deal.” CNBC, 12 June 2025. https://www.cnbc.com/2025/06/12/scale-ai-founder-wang-announces-exit-for-meta-part-of-14-billion-deal.html
  17. Lanier, Jaron, and E. Glen Weyl. “Should We Treat Data as Labor? Moving Beyond 'Free'.” American Economic Association Papers and Proceedings, 2018. https://www.aeaweb.org/articles?id=10.1257/pandp.20181003
  18. “Tips to Investigate AI Labor Abuses in the Global South.” Global Investigative Journalism Network, 15 January 2026. https://gijn.org/stories/investigating-ai-labor-abuses-global-south/
  19. Posada, Julian. “Embedded Reproduction in Platform Data Work.” Information, Communication & Society, 2022. https://posada.website/publication/embedded-reproduction-in-platform-data-work/
  20. “Former Meta Contractor Sama to Lay Off More Than 1,000 Workers in Kenya.” TechCabal / Associated Press, 16 April 2026. https://techcabal.com/2026/04/16/sama-lay-off-kenyan-workers-meta/

Tim Green

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Discuss...