Who Actually Governs AI: Part 2 - Labor

In 2021, a worker in Nairobi opened a queue of text on his screen. His job was to read it, all of it, and label the parts describing child sexual abuse, bestiality, and murder so an AI model could learn what those things looked like and refuse to produce them. He was paid, after taxes, about $1.32 an hour.

The company paying for that labor was OpenAI. The company doing the paying, and pocketing the difference, was Sama, a San Francisco outsourcing firm that billed OpenAI roughly six times what it passed on to its Kenyan staff. Four years later, a similar company built on similar labor, Scale AI, was valued at $29 billion.

Every AI model that refuses to help you build a weapon or describe a child learned that refusal from somewhere. It learned it from workers like this one, paid a few dollars a day to sit with the worst material on the internet so a chatbot in San Francisco never has to.

The Work Nobody Puts in the Keynote

Picture the AI supply chain as a factory floor nobody photographs. At one end: researchers publishing papers about alignment and safety. At the other: a worker in Nairobi, Manila, or Bogota, staring at a queue of graphic text or video for eight hours, deciding whether it counts as "extreme" or merely "disturbing."

That second job has a name, human feedback or content annotation, and it is unglamorous by design. Before a chatbot can be trusted to refuse harmful requests, someone has to show it thousands of examples of exactly what harmful looks like. Someone has to draw the box around the weapon in the video, or write the label on the paragraph describing torture.

In late 2021, TIME reported that OpenAI had contracted Sama to do exactly this: label toxic text, including detailed descriptions of child sexual abuse, bestiality, murder, suicide, and self-harm, to build the safety filter behind ChatGPT.

Sama's Kenyan employees, reviewing this material for up to nine hours a day, were paid a take-home wage of $1.32 to $2 an hour depending on seniority. One worker told TIME he began having recurring visions after reading a single especially graphic passage. Sama canceled the OpenAI contract early, in February 2022, citing the toll on its staff.

Did You Know?

The people who label AI training data rarely work for the company whose product they're improving. OpenAI's contract ran through Sama; Meta's ran through Sama and, later, Majorel; Scale AI's runs through a platform called Remotasks. The outsourcing layer isn't incidental, it's the business model: it lets a Silicon Valley company buy labeled data without ever appearing on a labor complaint filed in Nairobi or Manila.

Ten Cents an Hour, Several Time Zones Away

Scale AI built a different version of the same pipeline, and it scaled faster. Through its platform Remotasks, the company recruited at least 10,000 workers in the Philippines to label images and video for its Silicon Valley clients.

The pay did not hold up. The Washington Post reported that some Filipino freelancers who once earned around $10 per labeling task saw per-task rates crater to under a penny after Remotasks expanded into cheaper labor markets in India and Venezuela.

One worker, a 23-year-old named Charisse, was paid thirty cents for four hours of work. Another, Jackie, expected $50 for three days of labeling and received $12.

Scale AI kept expanding into lower-cost countries anyway, opening operations across Africa and in Venezuela specifically because labor there was cheaper than in the Philippines, according to reporting from labor researchers covering the sector. A 2025 survey by the labor rights group Equidem, which spoke with more than a hundred workers spread across four countries in Latin America, West Africa, East Africa, and Southeast Asia, found Kenyan workers on micro-task platforms like Remotasks earning $1.50 to $2 an hour for the same labeling and moderation work that pays $15 or more in the United States.

A Lawsuit That Still Hasn't Closed

The clearest evidence of what this work does to people is sitting in a Nairobi courtroom. In 2022, a former Facebook content moderator sued Meta and Sama for unlawful termination after organizing protests over working conditions. The case grew into a class action joined by more than 180 moderators.

In December 2024, medical assessments filed with the court became public: of 144 moderators who volunteered for psychological evaluation, more than 140 were diagnosed with conditions including severe PTSD, generalized anxiety, and major depressive disorder. The assessments described moderators encountering gruesome murders, sexual violence, and child abuse content daily, with some describing recurring nightmares and panic attacks years after leaving the job.

Kenya's Employment and Labour Relations Court has already ordered Meta and its contractors to provide ongoing psychiatric care to affected moderators. Mediation between the parties collapsed in 2024. As of February 2026, a final ruling in the underlying case was still pending, more than three years after the original complaint was filed.

Equidem's 2025 survey found that fear, not indifference, is the main reason so little of this gets reported firsthand. In Colombia, 75 of 105 workers approached for interviews declined to speak, citing non-disclosure agreements. In Kenya, 68 of 110 did the same. The NDA, researchers found, functions less as a trade-secrets protection and more as a gag order on describing working conditions at all.

The Investigation That Disappeared the Same Season the Money Arrived

Here is where labor and power intersect most plainly. In August 2024, the US Department of Labor opened an investigation into whether Scale AI's treatment of its contract labelers violated the Fair Labor Standards Act, the federal law covering unpaid wages and worker misclassification. TechCrunch first reported the investigation's existence in March 2025.

Two months later, in May 2025, the Department of Labor closed the investigation without explanation. Scale AI said it was "pleased" with the outcome.

One month after that, in June 2025, Meta announced a $14.3 billion investment for a 49% stake in Scale AI, valuing the company at roughly $29 billion. Scale's founder and CEO, Alexandr Wang, left to become Meta's chief AI officer, taking his large personal stake in the company with him. Nothing links the investigation's closure to the investment directly.

But the sequence, an open federal labor probe, then no probe, then one of the largest AI infrastructure deals of the year, is the plainest illustration yet of a pattern running through this whole series: leverage in AI governance follows money, and the workers whose labor makes these products safe to ship have almost none of either.

Workers Are Organizing Anyway

The response building in Nairobi isn't silence. In early 2025, Kenyan data workers launched the Data Labelers Association, drawing 339 members in its first week and growing to roughly 800 within months. Its goals: wages workers can live on, real access to counseling, and contracts written in plain terms, for both gig-platform freelancers and outsourced BPO staff alike.

The association has already taken action beyond Kenya's borders. Alongside the French union CGT FAPT and a Pan-African moderators' network, it filed a formal legal notice against Teleperformance, one of the largest data-work providers to AI companies globally, and it now coordinates with groups including Turkopticon and the Distributed AI Research Institute. It is a workforce with almost no leverage over the companies it serves, building the only kind of leverage available: each other.

Knowlegic Perspective

Most conversations about AI governance are about what the technology will do to the world. This one is about what building the technology already did to specific people, in specific cities, for specific hourly wages.

The gap here isn't subtle. A federal labor investigation into Scale AI opened, then closed without explanation, in the same season Meta closed a deal that valued the company at $29 billion. A Kenyan court has spent three years adjudicating what more than 140 documented PTSD diagnoses are worth, while the companies involved negotiate mediation and appeals. Neither process moves at the speed of the industry it's trying to hold accountable.

What's changing is who gets to describe the trade-off. A data labelers' union forming in Nairobi in 2025 is a small thing against a $29 billion valuation. But it's the first time the workers doing this labor have had an organized voice in a conversation that has, so far, been conducted entirely by the companies buying their labor and the courts trying to catch up to it.

The people whose judgment calls teach an AI model what it should refuse to say were paid $1.32 to $2 an hour to make that call. The company built on their labor is worth $29 billion. Between those two numbers is most of what "AI governance" actually means in practice.

Read Next: Who Actually Governs AI: Part 3 - Defense

Sources & References

• OpenAI Used Kenyan Workers on Less Than $2 Per Hour: Exclusive, TIME (2023)
• Facebook inflicted 'lifelong trauma' on content moderators in Kenya, campaigners say, as more than 140 are diagnosed with PTSD, CNN Business (2024)
• Over 140 Kenyan Facebook Moderators Diagnosed With Severe PTSD, Kenyans.co.ke (2024)
• The Department of Labor just dropped its investigation into Scale AI, TechCrunch (2025)
• Meta Finalizes $14.3 Billion Scale Investment, Hires Its CEO, Bloomberg (2025)
• Kenyan AI workers form Data Labelers Association, Computer Weekly (2025)