Ghost Workers

The human beings training AI for two dollars an hour

Land Promised, Purpose Betrayed | In Other News

The land fights get the zoning meetings and the petitions. The other half of AI's exploitation story happens thousands of miles from any data center, in cramped offices and home computers across Nairobi, Medellin, and Manila, where the actual training of these systems depends on a global workforce that most users of AI chatbots will never think about.

They are sometimes called data workers. Researchers who study the industry, including Mary Gray and Siddharth Suri in their influential book on the subject, call them ghost workers: people who label images, moderate violent content, and rank AI outputs, work that is essential to how these systems function and almost entirely invisible in the finished product.

What the work actually pays

Multiple investigations, spanning CBS News, the Wall Street Journal, and independent researchers, converge on a consistent range. Kenyan content moderators working through subcontractors for major AI and social media companies have reported hourly pay between roughly 1.10 and 3.74 dollars, depending on the contractor and the year of the report. A Venezuelan worker interviewed by Canadian Affairs, laboring for five different data labeling platforms from Colombia, described earning between 5 and 25 cents per completed task, work she called being treated like a ghost, unseen despite contributing directly to technological progress the rest of the world benefits from.

"We're like ghosts. No one knows that we exist, even though we are contributing to society's technological progress."

That is Oskarina Fuentes, describing her work to Canadian Affairs. It is a sentiment echoed across nearly every worker account collected on this subject, from Nairobi to Manila to Medellin.

The content itself is the harm

Unlike most low wage outsourced labor, this work carries a specific and well documented psychological cost. Workers assigned to content moderation and safety filtering are asked to review graphic violence, child exploitation material, and other extreme content for eight hour shifts, in order to teach AI systems what to reject and filter out for everyone else. More than 140 Facebook content moderators in Kenya have been diagnosed with post traumatic stress disorder, according to reporting compiled by researchers at Temple University's Institute for Law, Innovation and Technology. Workers describe insomnia, panic attacks, and lasting damage to personal relationships.

Mophat Okinyi, a former content moderator for a company contracted to review ChatGPT outputs, has described the psychological toll of the work in Kenyan court filings that led to a landmark 2023 ruling. A Kenyan judge found that Meta, not merely its contractor, was the "true employer" of workers like Okinyi, a decision later upheld on appeal that opened the door for content moderators to sue major tech companies directly in Kenyan labor courts rather than being limited to claims against an intermediary staffing firm.

Precarity by design

The instability of the work compounds its psychological toll. Contracts are frequently structured week to week or month to month. Workers describe being locked out of accounts and denied final pay after raising concerns about conditions, with no clear channel to dispute the decision. One platform, Remotasks, abruptly shut down its Kenyan operations in 2024 after workers began speaking publicly about unpaid wages, a move researchers describe as a recognizable pattern: when workers in one country begin organizing effectively, companies quietly relocate the work rather than address the underlying conditions.

Kenyan workers have not stayed passive in the face of this. The Data Labelers Association formed specifically to push for a code of conduct covering fair pay, psychological support, and the right to organize. In 2024, nearly 100 Kenyan data workers wrote an open letter to the sitting United States president, describing their working conditions as amounting to modern day slavery and asking that American companies be held accountable in American courts for labor practices carried out abroad through subcontractors.

Not every company has responded the same way. Anthropic has stated it requires subcontractors handling sensitive content to follow wellbeing guidelines and to pay data labelers at least 16 dollars an hour. OpenAI has said it maintains its own standards for subcontractors covering workplace safety and fair compensation, and confirmed it ended its relationship with Scale AI, the company that operated Remotasks in Kenya. Whether standards like these are consistently enforced across the long, opaque chain of subcontractors that typically separates a worker in Nairobi from the company whose model they are training remains an open and largely unverifiable question for anyone outside the companies themselves.

The uncomfortable arithmetic

The wage gap is the plainest way to see the imbalance. Comparable domestic outsourcing labor in the United States has been reported at roughly 12.50 to 16 dollars an hour. The same category of work performed in Kenya has been documented at 1.10 to 3.74 dollars an hour, a gap researchers describe as geographic arbitrage: locating the most psychologically demanding tasks in the labor markets with the least bargaining power and the weakest legal protection.

AI's public face is a chat window that feels instant, frictionless, almost magical. Behind that window sits a global supply chain of human judgment, much of it performed by people paid a small fraction of what the same task would cost in the country where the product is sold. A technology this profitable can afford to pay the people who make it safe. The industry's own stated wage floors, when they exist at all, prove that. The question is why they remain the exception instead of the rule.

Sources: CBS News/60 Minutes, Canadian Affairs, Brookings Institution, Temple University Institute for Law, Innovation and Technology, Qhala/Medium, techimpactonworkers.iccr.org, Weizenbaum Institute research cited across multiple outlets.