Labour Rights

The Invisible Workforce: Data Labelling, Psychological Harm, and the Fight for AI Labour Rights

The AI industry's most sophisticated models are trained by some of the world's lowest-paid workers — and in late 2025, those workers began organising.

P J Laszkowicz

On October 30, 2025, the Collaboration on International ICT Policy for East and Southern Africa convened DataFest Africa in Kampala, Uganda.1 Among the sessions on AI governance and digital rights was a discussion that most technology conferences never hold: the labour conditions of the people who make artificial intelligence work.

The workers in question are data labellers and content moderators — the humans who classify images, transcribe audio, evaluate model outputs, and review the most disturbing content the internet produces so that AI systems can learn what is acceptable and what is not. They work in Nairobi, Lagos, Bogotá, Manila, and Accra, employed by outsourcing firms whose names rarely appear in the product announcements of the companies they serve. They are paid between one and three dollars an hour.2 They are essential to every frontier AI model in production today. And in late 2025, after years of exploitation documented by journalists, researchers, and the workers themselves, they began to organise.

This article examines the labour supply chain that underpins the AI industry, the documented harms it produces, the first organised worker responses, and the regulatory frameworks — particularly the European Union's Platform Work Directive — that may begin to change the terms of the relationship. It is not an article about whether AI should exist. It is an article about what it costs, and who pays.

The supply chain beneath the model

Every large language model, every image classifier, every content moderation system, and every autonomous driving algorithm depends on human labour to function. The labour takes several forms: labelling images with bounding boxes and categories, classifying text by sentiment or toxicity, evaluating model outputs for accuracy and safety, transcribing and translating audio, and — most controversially — reviewing and categorising the violent, sexual, and extremist content that AI systems must learn to identify and filter.3

This work is overwhelmingly outsourced. The major AI companies — OpenAI, Meta, Google, Anthropic, and others — contract with intermediary firms that recruit, manage, and pay the workers. The largest of these intermediaries include Scale AI (which operates the Remotasks platform), Sama (formerly Samasource), Appen, and CloudFactory.4 The intermediaries, in turn, recruit workers primarily in the Global South: Kenya, the Philippines, India, Venezuela, Colombia, Ghana, and Nigeria.5

The economics are stark. When OpenAI contracted with Sama to provide content moderation and data labelling for ChatGPT, it paid the firm approximately $12.50 per hour per worker.6 The workers themselves received roughly $2 per hour.7 This is not an outlier. Across the industry, hourly rates for data labellers in Kenya range from $1.50 to $2. In Venezuela, the range is $0.90 to $2. In the United States, comparable work pays $10 to $25 per hour.8

The global data labelling market is projected to reach $1.89 billion in 2025, growing at a compound annual rate of 23.6 percent.9 The AI companies whose products depend on this labour are collectively valued in the trillions. The workers who perform it earn less than the cost of a cup of coffee in the cities where those companies are headquartered.

Scroll. Click. Suffer.

On May 28, 2025, the human rights organisation Equidem published "Scroll. Click. Suffer," the most comprehensive investigation to date of the psychological harm inflicted on data labellers and content moderators working for major AI and social media platforms.10

The report documented interviews with 113 workers across Colombia, Ghana, Kenya, and the Philippines — workers employed through outsourcing firms to perform data labelling and content moderation for platforms including Meta, TikTok, and ChatGPT.11 A subset of 76 workers from Colombia, Ghana, and Kenya reported 60 independent incidents of psychological harm, including post-traumatic stress disorder, paranoia, depression, anxiety, insomnia, and sexual dysfunction.12

The nature of the work explains the harm. Content moderators review material that includes graphic violence, child sexual abuse, beheadings, torture, and sexual exploitation. Workers reported processing 700 to 1,000 cases in a standard shift, with an average of 7 to 12 seconds to evaluate each item.13 The volume and pace leave no time for psychological processing. The content is not occasional; it is the entirety of the job.

Equidem's research documented violations of International Labour Organisation standards, including protections on fair wages, the right to organise, and safeguards against forced labour.14 Workers reported that mental health support, where it existed, was inadequate — brief counselling sessions offered by the same employer whose work caused the trauma, with no independent clinical oversight and no long-term follow-up.

The irony is structural. These workers are performing what the AI industry calls "safety work" — the human labour that makes AI systems safe for everyone else. The process of making AI safe requires exposing humans to precisely the content that AI is being trained to detect and suppress. The industry's safety depends on their harm.

The TIME investigation and its aftermath

The current crisis did not begin in 2025. In January 2023, TIME published an investigation revealing that OpenAI had contracted with Sama to employ Kenyan workers who reviewed graphic and disturbing content to train ChatGPT's safety filters.15 Workers were paid less than $2 per hour. They were regularly exposed to depictions of sexual violence, child abuse, and murder. When they developed severe psychological symptoms — PTSD, paranoia, depression, anxiety, insomnia, sexual dysfunction — the support offered was insufficient.16

OpenAI's contract with Sama had begun in 2021.17 The company acknowledged that the work was "challenging" and described the workers' contributions as "immensely valuable."18 When the contract was terminated, workers reported being sent home without receiving pending compensation or continued access to medical care.19

In 2023, 184 former Sama content moderators filed a lawsuit against the company, alleging unfair termination and exploitative working conditions.20 Separately, a group of Kenyan workers petitioned parliament to investigate the conditions under which they had laboured, requesting that the Ministry of Labour examine the role of outsourcing firms and the international technology companies they served.21

The Sama case was not the only one. In May 2022, Daniel Motaung, a South African who had been recruited by Sama to moderate Zulu-language content for Facebook in Nairobi, filed a lawsuit against both Sama and Meta.22 Motaung had been paid approximately $2.20 per hour to review content including beheadings and child abuse material. When he attempted to organise a union to negotiate better conditions, he was dismissed.23

Motaung's case became a landmark. Meta's lawyers argued that the Kenyan court had no jurisdiction over the company because it was "not resident, trading or domiciled in Kenya."24 The court disagreed. In September 2024, the Kenyan Court of Appeal ruled that Meta could be sued in Kenya, marking what Amnesty International described as the first time Meta had been "significantly subjected to a court of law in the global south."25

The case continues. Its significance extends beyond the individual claim: it established that the companies at the top of the outsourcing chain — not just the intermediaries — can be held legally accountable for conditions at the bottom.

Remotasks: a case study in abandonment

On March 8, 2024, workers in Kenya who logged into Remotasks — the data labelling platform operated by Scale AI, valued at $14 billion — found that their access had been blocked without warning.26

Scale AI had cut access to its platform for workers in Kenya, Nigeria, and Pakistan simultaneously.27 Workers in some countries received three weeks' notice. Workers in Kenya did not, due to what Scale AI later described as an "operational error."28 Thousands of people who depended on the platform for their primary income were left with no work, no notice, and no recourse.

The situation was compounded by a payment crisis. In December 2023, AirTM — the platform through which Remotasks workers received payment — had announced that it was investigating funds earned through the platform. Workers' accounts were frozen, and AirTM subsequently stated that the funds had been "improperly obtained" and returned them to Remotasks.29 Workers who had performed the labour and earned the payment found their savings confiscated by the payment intermediary, at the direction of the labour intermediary, without any adjudication or appeals process.

Kellion Mrego, a worker who had been on the platform since 2018, described sticking with Remotasks through declining wages and work shortages — only to be removed without explanation or compensation.30 The pattern is characteristic: a workforce with no employment contract, no labour protections, no union representation, and no access to legal systems in the jurisdictions where their employers are incorporated.

The Data Labelers Association

At the end of 2023, a group of Kenyan data workers formed a collective to advocate for their rights.31 They initially sought to establish a trade union, but Kenya's labour framework made formal unionisation difficult for workers classified as independent contractors on international platforms.32 In early 2025, they chose an alternative path: on February 13, under the theme "Empowering the People Powering AI," they formally launched the Data Labelers Association.33

Three hundred and thirty-nine workers joined in the first week. Within months, the membership had grown to over 800.34

The DLA's demands are not radical. They are asking for fair wages, mental health support, transparent working conditions, and recognition of the work they perform.35 They are working with the African Content Moderators Union, the worker advocacy platform Turkopticon, and the Distributed AI Research (DAIR) Institute — the organisation founded by Timnit Gebru after her departure from Google.36

The Association is already engaged with Kenyan policymakers. It is communicating with the Ministry of Information, Communications, and Technology to help legislators understand the nature of platform-mediated data work and to advocate for legal protections for workers whose employers are incorporated in foreign jurisdictions.37

The DLA represents the first organised labour response to data labelling conditions in East Africa. It is small, under-resourced, and operating in a legal environment not designed for the work its members perform. But its existence marks a structural shift: the workers at the bottom of the AI supply chain are no longer waiting for the companies at the top to reform themselves.

From Nairobi to Brussels

On the other side of the AI labour supply chain, a different kind of response is taking shape. On October 23, 2024, the European Parliament adopted Directive (EU) 2024/2831 — the Platform Work Directive — the most significant piece of labour rights legislation directly addressing AI-mediated work.38

The Directive's scope is broader than its name suggests. It applies not only to ride-hailing and food delivery platforms but to any digital platform that organises work through algorithmic management — including, explicitly, data annotation and content moderation services.39 Member states have until December 2, 2026 to transpose it into national law.40

The Directive's provisions address the core mechanisms through which data labelling workers are exploited.

First, it creates a legal presumption of employment for platform workers who meet specified criteria, shifting the burden from the worker to prove they are an employee to the platform to prove they are not.41 This directly challenges the independent contractor classification that allows outsourcing firms to deny workers employment protections, benefits, and social security.

Second, it bans the algorithmic processing of certain categories of personal data about workers, including their emotional or psychological state, personal beliefs, and biometric data beyond authentication.42 For content moderators who review graphic material for eight hours a day, the prohibition on monitoring emotional state is not abstract — it protects them from surveillance systems that could penalise workers who show signs of psychological distress.

Third, it requires human oversight of significant decisions affecting workers. A person performing platform work cannot be dismissed based solely on a decision taken by an algorithm or an automated decision-making system.43 Given the documented pattern of workers being removed from platforms without explanation — as in the Remotasks case — this provision establishes a minimum standard of procedural fairness.

Fourth, it guarantees workers the right to information about how algorithmic systems are used in their management and evaluation, and the right to contest decisions made by those systems.44

The Directive is, in principle, the regulatory framework that the DLA is demanding through organising. The question is whether it will reach the workers who need it most.

The jurisdictional gap

The Platform Work Directive applies to workers performing platform work in the European Union. It does not apply to workers in Kenya, the Philippines, Colombia, or Ghana — the countries where the majority of data labelling and content moderation work is performed.

This is the structural limitation of even the most progressive labour regulation: it protects workers within its jurisdiction while the work migrates to jurisdictions where protection is absent. If the Directive makes it more expensive or legally complex to employ data labellers within the EU, companies may simply shift more of the work to countries where the Directive does not apply.

There are countervailing forces. The Motaung ruling established that courts in Kenya can exercise jurisdiction over companies like Meta, even when those companies are not incorporated there. The African Union's Continental AI Strategy, now in Phase I of implementation, includes governance frameworks and capacity building for AI labour.45 Kenya's National AI Strategy, with a $1.14 billion allocation, addresses the country's role in the global AI economy.46 Nigeria's Data Protection Commission introduced enforcement rules in September 2025 that require impact assessments for automated decision-making.47

But these frameworks are nascent. They lack the enforcement capacity, the institutional infrastructure, and — in many cases — the political will to challenge the outsourcing model that delivers cheap labour to the world's richest companies. The gap between the EU's progressive legislation and the conditions in Kenyan data labelling centres is not a failure of regulation in any single jurisdiction. It is a structural feature of a global supply chain designed to exploit regulatory arbitrage.

The ILO's decent work framework, the OECD's AI principles on inclusive growth, and the UN Global Digital Compact all articulate standards that, if applied, would transform the conditions under which data labelling work is performed.48 None of these instruments is binding. None has an enforcement mechanism. They represent aspirations, not obligations.

What fair AI labour looks like

The question is not whether the AI industry depends on human labour. It does. The question is whether that dependence will be acknowledged, compensated, and governed — or whether it will continue to be outsourced to countries where the cameras and the regulators do not reach.

Fair AI labour would require, at minimum, several structural changes. Wages should reflect the actual difficulty, skill, and psychological toll of the work, not the poverty of the labour market in which it is performed. Workers should have employment status, not independent contractor classification, with the protections that employment brings: social security, healthcare, paid leave, and protection from arbitrary dismissal. Mental health support should be independent, clinical, and ongoing — not a brief counselling session offered by the employer whose work caused the harm. Workers should have the right to organise and bargain collectively, without the threat of dismissal that Daniel Motaung and others have faced. And the companies whose products depend on this labour should bear direct responsibility for the conditions under which it is performed, not hide behind layers of outsourcing intermediaries.

The EU's Platform Work Directive addresses several of these requirements — within the EU. The DLA is pushing for them in Kenya through organising. The Motaung litigation is testing whether they can be enforced through courts.

None of these mechanisms, individually, is sufficient. Together, they represent the earliest stages of a response to a problem that the AI industry created, benefits from, and has shown little inclination to solve on its own.

The cost of intelligence

Every time a user interacts with a frontier AI model — asking a question, generating an image, receiving a recommendation — the interaction rests on a foundation of human labour. Someone classified the training data. Someone evaluated the model's outputs. Someone reviewed the graphic content that the model was taught to recognise and refuse. Someone, somewhere, was paid less than two dollars an hour for that work, and may have developed post-traumatic stress disorder in the process.

The AI industry's total investment in 2025 is measured in hundreds of billions of dollars. The annual cost of fairly compensating its data labelling workforce would be measured in the low single-digit billions — a fraction of a percentage of the capital being deployed. The disparity is not a technical limitation. It is a choice.

The formation of the Data Labelers Association, the Equidem investigation, the Platform Work Directive, and the Motaung litigation do not, individually, change that calculus. But they establish something that did not exist before: a record, a set of demands, a legal framework, and a precedent. The AI industry can no longer claim ignorance of the conditions under which its most essential labour is performed. The question is whether it will respond to the knowledge with structural change — or whether it will continue to outsource the cost of its intelligence to the people who can least afford to bear it.

Footnotes

  1. CIPESA, "CIPESA at DataFest Africa 2025: Advancing Ethical AI and Digital Rights in Africa," November 2025. ↩

  2. Wage ranges documented by TIME, Equidem, CBS News, and the LSE Media Blog across investigations from 2023 to 2025. ↩

  3. The taxonomy of data labelling work is described in detail in the OECD AI Policy Observatory's reporting on AI labour and in academic literature on the "ghost work" supporting AI systems. ↩

  4. Scale AI operates Remotasks as its primary crowdsourcing platform for data labelling. Sama (formerly Samasource) has been a primary contractor for OpenAI and Meta. Appen and CloudFactory are additional major intermediaries. ↩

  5. LSE Media Blog, "The perilous future of AI work in the Global South," 14 November 2025. ↩

  6. WeeTracker, "OpenAI Paid Sama $12 An Hour Per Worker — Kenyans Only Got $2," 25 November 2024. ↩

  7. TIME, "OpenAI Used Kenyan Workers on Less Than $2 Per Hour," January 2023. Workers reported earnings of $1.46 to $3.74 per hour, with the majority earning approximately $2 per hour. ↩

  8. The Conversation, "AI is a multi-billion dollar industry. It's underpinned by an invisible and exploited workforce," 2024. Venezuelan rates from the same source. ↩

  9. Market projection from industry analysis cited in o-mega, "Data Labeling Industry Guide," 2025. ↩

  10. Equidem, "Scroll. Click. Suffer: The Hidden Human Cost of Content Moderation and Data Labelling," 28 May 2025. ↩

  11. The 113 workers were employed through outsourcing firms serving Meta, TikTok, and OpenAI, among other platforms. ↩

  12. Equidem reported 60 independent incidents of psychological harm across 76 workers interviewed from Colombia, Ghana, and Kenya. ↩

  13. Equidem, "Scroll. Click. Suffer," 2025. Workers described processing 700 to 1,000 items per standard shift. ↩

  14. Equidem documented violations of ILO protections on fair wages, the right to organise, and safeguards against forced labour. ↩

  15. TIME, "OpenAI Used Kenyan Workers on Less Than $2 Per Hour: Exclusive," January 2023. The investigation by Billy Perrigo documented the conditions under which Sama employees labelled content for OpenAI. ↩

  16. Workers reported developing PTSD, paranoia, depression, anxiety, insomnia, and sexual dysfunction. TIME, January 2023. ↩

  17. OpenAI's contract with Sama for content moderation and data labelling began in 2021. ↩

  18. OpenAI statement quoted in TIME, January 2023. ↩

  19. Workers reported being sent home without pending compensation or continued medical care when the contract ended. Kenyan workers petitioned parliament over these conditions in 2023. ↩

  20. 184 former Sama moderators filed a lawsuit alleging unfair termination and poor working conditions. MedioNama, "Kenyan Workers Expose Disturbing Work Conditions in AI Data Labelling for OpenAI," July 2023. ↩

  21. TechCrunch, "Workers that made ChatGPT less harmful ask lawmakers to stem alleged exploitation by Big Tech," 14 July 2023. ↩

  22. openDemocracy, "Daniel Motaung: Facebook slammed for trying to dodge Kenyan whistleblower," 2022. ↩

  23. Motaung was paid approximately $2.20 per hour and was dismissed after attempting to organise a union. openDemocracy and Africanews, 2022. ↩

  24. Meta's jurisdictional argument cited in TIME, "Meta's Attempt to Dodge Major Kenya Court Case Fails," 2023. ↩

  25. Kenya Court of Appeal ruling, September 2024. Meta Platforms, Inc & 2 others v Motaung & 186 others, Civil Appeal E232 & E445 of 2023 (Consolidated), KECA 1262 (KLR), 20 September 2024. ↩

  26. Kenyan Wall Street, "Online Gig Workers Left Stranded as Remotasks Exits Kenya," March 2024. ↩

  27. Rest of World, "Scale AI's Remotasks is booting workers with no explanation," March 2024. ↩

  28. Scale AI apologised for the notification failure, describing it as an "operational error." Computerworld, "Report: Scale cuts off subsidiary's remote workers in several countries," 2024. ↩

  29. AirTM announced in December 2023 that it was freezing worker accounts and returning funds to Remotasks on the grounds that the funds were "improperly obtained." Workers who had performed the labour lost access to their earnings without adjudication or appeal. ↩

  30. Mophat Okinyi, "Impact of Remotasks' Closure on Kenyan Workers," Data Workers' Inquiry, 2024. ↩

  31. Computer Weekly, "Kenyan AI workers form Data Labelers Association," February 2025. ↩

  32. Kenya's labour framework presents challenges for workers classified as independent contractors on international platforms, making formal trade union registration difficult. ↩

  33. The Data Labelers Association was officially launched on February 13, 2025. Computer Weekly, February 2025; Martijn Arets, "Data Labelers Association speaks up for invisible workers," 2025. ↩

  34. 339 members joined in the first week. Membership exceeded 800 within several months. Martijn Arets, 2025. ↩

  35. DLA's core demands include fair wages, mental health support, transparent working conditions, and recognition of data labelling as skilled work. ↩

  36. The DLA works with the African Content Moderators Union, Turkopticon, and the Distributed AI Research (DAIR) Institute. Computer Weekly, 2025. ↩

  37. The DLA is in communication with Kenya's Ministry of Information, Communications, and Technology to inform legislative understanding of platform-mediated data work. Computer Weekly, 2025; BMZ Digital.Global interview with Joan Kinyua. ↩

  38. Directive (EU) 2024/2831 of the European Parliament and of the Council on improving working conditions in platform work, adopted 23 October 2024. ↩

  39. The Directive's scope covers any digital platform that organises work through algorithmic management. Analysis by AI Policy Lab Sweden confirmed explicit coverage of data annotators and content moderators. ↩

  40. Member states must transpose the Directive into national law by 2 December 2026. European Parliament, "Parliament adopts Platform Work Directive," 2024. ↩

  41. The legal presumption of employment shifts the burden of proof from the worker to the platform. Crowell & Moring, "New EU Directive Impacting Digital Platforms and Individuals Working for Them," 2024. ↩

  42. Article 7 of the Directive prohibits the processing of personal data on the emotional or psychological state of the person performing platform work, data related to private conversations, data to predict trade union activity, and biometric data beyond authentication. ETUI analysis, 2024. ↩

  43. The Directive requires human oversight of important decisions directly affecting workers, including dismissal. Fisher Phillips, "New EU Platform Work Directive Impacts Freelancers and Gig Economy," 2024. ↩

  44. Workers are guaranteed the right to information about algorithmic management systems and the right to contest automated decisions. Mason Hayes Curran, "The Platform Work Directive," 2024. ↩

  45. The African Union Continental AI Strategy, Phase I (2025–2026), focuses on governance frameworks and capacity building. ISHR reporting on the NGO Forum, 2025. ↩

  46. Kenya's National AI Strategy (2025–2030) includes a $1.14 billion allocation addressing the country's role in the global AI economy. ↩

  47. Nigeria's Data Protection Commission introduced the General Application and Implementation Directive in September 2025, requiring data protection impact assessments for automated decision-making. Tech in Africa, 2025. ↩

  48. The ILO Decent Work Agenda, the OECD AI Principles (particularly Principle 1.2 on inclusive growth), and the UN Global Digital Compact all articulate standards relevant to AI labour conditions. ↩