The Ghost Workers Powering ChatGPT: Inside AI's Hidden Labor Economy

The pitch from Silicon Valley has been consistent for years: AI is coming for your job. The language shifts, "augmentation," "transformation," "displacement", but the underlying message is the same. Humans are increasingly unnecessary. The machines are learning. Step aside.

There is a detail, buried just beneath the surface of that narrative, that the companies building these systems would prefer you not examine too closely. The machines are not learning on their own. Every large language model, every image classifier, every content moderation system running at scale, they require vast quantities of human-annotated data to function. Someone has to tell the model that this text is helpful and that text is harmful. Someone has to label the image. Someone has to review the output that cannot be shown to the public. That someone is a data worker. There are millions of them. And the investigative documentary produced by A More Perfect Union, which has now accumulated nearly a million YouTube views, went and found some of them.

What the investigation uncovered is a labor economy built on paradox: the companies claiming to make human work obsolete are, right now, entirely dependent on a hidden underclass of human workers to power their products. The workers are underpaid, routinely exposed to psychological harm, and nearly invisible, by design. The founders building on top of their labor are, in several cases, becoming the youngest self-made billionaires in history.

The People Behind the Prompt

The investigation's most important contribution is not statistical, it is personal. The workers who agreed to speak, many under conditions of anonymity out of fear of tech company retaliation, are not a caricature. They are educated, thoughtful, and acutely aware of the irony of their situation.

There is "Jen," an Ivy League PhD graduate who took data annotation work because it was available, flexible, and she needed income. The documentary captures what this work actually felt like for her. Jen called her mother crying, trying to explain a financial situation she could not quite articulate. "I can't see where this money is coming from," she told her mother. An Ivy League doctorate. Data annotation. And still, at the end of the month, the math didn't work.

Then there is Ozzy, a philosophy graduate from Oregon, whose story centers on a project with a name that should give anyone pause: Project Arsenic. Ozzy was assigned to review AI-generated content as part of a quality and safety pipeline, work that falls under the broad category of content moderation. What that meant in practice: animals being killed. Celebrities depicted in gore scenes. Imagery the documentary describes as "furniture constructed from humans." Ozzy had nightmares for weeks. He shared documentation of what he had been asked to review. Surge AI, the platform through which the work was contracted, publicly denied having graphic content of that nature. The documentation Ozzy shared directly contradicted that denial.

Krystal Kauffman's story is different in texture but similar in structure. A longtime data worker who came to platform labor after illness made traditional employment difficult, Kauffman eventually crossed into organizing. She became one of the people behind Turkopticon, a tool built to let workers on Amazon's Mechanical Turk platform rate and review the requesters who post tasks, giving workers some small amount of collective visibility in a marketplace designed to render them invisible. Kauffman's trajectory, from isolated gig worker to researcher and organizer, is the story of what it looks like when someone refuses to accept the terms of the system unexamined.

The Economic Reality, in Numbers

The human stories are load-bearing. But so are the numbers. Journalist Karen Hao, who conducted the core investigation, did not rely only on individual accounts. The data that sits behind those stories is structural and consistent.

86% of data workers struggle to meet their financial responsibilities. This is not a marginal cohort, it is the overwhelming majority of the workforce that makes AI training possible. One in four rely on Medicaid or food stamps. More than one in five have experienced homelessness. The median earnings figure lands below $23,000 per year. For context: the federal poverty level for a single person in the United States is approximately $15,000. These workers are not in poverty, but they are close enough that a single medical bill, a single missed contract, a single platform policy change can push them there.

And then, on the other end of the same supply chain: Alexandr Wang. The founder of Scale AI, one of the largest data labeling companies in the world, the intermediary layer between AI giants and the workers who do the actual annotation, became the world's youngest self-made billionaire in 2021. He was 25. That record did not stand long. Three 22-year-olds who founded Mercor, an AI hiring platform, displaced him from the title.

The arithmetic deserves to be stated plainly. Workers sometimes earn $40 for a 13-hour day, under $3.10 an hour for skilled cognitive labor. The founder of the company that brokers that labor became a billionaire before he turned 26. This is not a bug in the system. The gap between those two figures is how the economics work.

Project Arsenic: The Psychological Toll Nobody Talks About

Content moderation is not new. Platforms have employed human reviewers to screen harmful material since the early days of social media. What is new is the specific character of the work being done now, and the fact that it is being done, at scale, by contractors with no employment protections, no mental health support structures, and no legal recourse when a company denies they were exposed to what they were demonstrably exposed to.

Project Arsenic is the clearest example in the investigation of what this looks like in practice. Ozzy, the philosophy graduate from Oregon who took the work, was asked to evaluate AI-generated content for quality and safety purposes. The content he saw was not incidental or occasional. It was systematic: animals being killed, celebrities placed in gore sequences, and what the investigation describes as "furniture constructed from humans." These are not edge cases that slipped past a filter. They are the category of content that requires human review precisely because no automated system can be trusted to evaluate it consistently.

Ozzy had nightmares for weeks. He documented what he was asked to review. And when he did, Surge AI, the platform that had contracted him, publicly denied that its pipeline contained graphic content of that nature. That denial was, on the documentary's evidence, false. The documentation Ozzy shared contradicted it directly.

The specific details of Project Arsenic matter less than the structural condition they reveal. When workers review harmful content, they bear the psychological cost. When they speak about it, the company denies it occurred. When they seek documentation, they are often told the work is confidential, covered by non-disclosure agreements, or simply not subject to the kind of accountability that would apply to an employee. The contractor classification, the same classification that allows platforms to avoid minimum wage requirements, benefits, and labor protections, also allows them to deny responsibility for what they ask workers to do.

The US Department of Labor opened an investigation into Scale AI's practices, examining potential violations of fair labor standards, specifically around contractor classification and wage compliance. That investigation reflects a regulatory awareness that the contractor model, as deployed in AI data work, may be exceeding even the permissive limits currently set for gig labor.

The Invisible Supply Chain: Why This Stays Hidden

The invisibility of this workforce is not accidental. It is a product of the supply chain structure that the industry built, deliberately or not, to maximize cost efficiency while distributing accountability to the point of diffusion.

The chain works like this: AI giants, OpenAI, Google, Meta, the hyperscalers, need training data and content evaluation. They do not hire the workers who produce it. They contract with data work companies, Scale AI, Surge AI, Appen, Lionbridge, who compete aggressively for those contracts on price. Those companies, in turn, deploy their work through platforms populated by contractor workers: Amazon Mechanical Turk, Remotasks, Clickworker, and dozens of smaller equivalents. Each link in the chain creates distance between the end beneficiary and the worker. Each link also represents a layer at which wages can be compressed and accountability can be deflected.

The workers themselves understood the architecture. Karen Hao, who led the investigation, reports that the workers she spoke with were afraid of tech company retaliation, not retaliation from the small intermediary platforms, but from the AI giants at the top of the chain. Workers who speak publicly risk being blacklisted from the platforms where they find work. The platforms depend on the AI giants for contracts. The alignment of incentives is total: no one in the chain benefits from workers speaking publicly, and the workers have no institutional protection when they do.

Tim's academic study, referenced in the investigation, captured the working conditions in terms that reveal a workforce under constant precarity: workers described scrambling to complete tasks before roles ended, working midnight shifts to capture tasks before other contractors in different time zones claimed them, and living under the persistent anxiety of a labor market in which work could disappear without notice or explanation. There is no severance. There is no unemployment insurance. There is only the next task queue.

The Ideology of Automation: "Most Humans Are Unnecessary"

The economic conditions of data workers do not exist in a vacuum. They are downstream of a set of beliefs, held and articulated by some of the most powerful people in the technology industry, about what human labor is worth and what the future should look like. Those beliefs are not merely descriptive. They function as policy choices, and as Daron Acemoglu, MIT economist and Nobel laureate, makes clear in the investigation, they are choices, not inevitabilities.

Acemoglu does not mince words about the ideological dimension of what is happening. There is a strand of thinking in Silicon Valley, he calls it an "elitist attitude", that the automation of human work is not merely economically efficient but morally neutral, perhaps even desirable. Some in the industry have gone further, suggesting explicitly that most humans are unnecessary in a future shaped by AI. The sentiment surfaces in different forms: in claims about AI's productivity multipliers, in projections about the percentage of jobs that will be automated, in the language of "augmentation" that implies human workers are inputs to be optimized rather than people with interests that deserve consideration.

Acemoglu's framing is important because it cuts through the technological determinism that often shapes how these conversations go. The outcome he describes, a small number of corporations controlling most of the labor economy, with a large fraction of workers sidelined, is not the inevitable result of AI's capabilities. It is the result of policy choices about how AI is deployed, who captures the gains, and what protections workers are afforded. "That is a choice," he says. The emphasis matters: it could be a different choice.

Mary Gray, whose 2019 warning about the "Uber-ization of all knowledge work" now reads as prescient, framed the same dynamic in structural terms. The logic that Silicon Valley applied to car rides, disaggregate the work, classify the workers as independent contractors, capture the margin, and scale, was never going to stay in transportation. It was always going to be applied to every form of work that could be broken into discrete, measurable tasks and assigned through a digital platform. Data annotation was an early and obvious target. It will not be the last.

The Loop: AI Takes Jobs, Displaced Workers Train More AI

The most structurally disturbing finding in the investigation is not any single statistic. It is the feedback mechanism that the current AI economy is building toward, a loop that, once established, becomes self-reinforcing and increasingly difficult to break out of.

The mechanism works in stages. AI systems, trained and improved by data workers, are deployed in ways that automate or displace jobs across the economy. Those displaced workers, needing income, turn to whatever flexible, platform-mediated work is available. Data annotation and AI training tasks are available. The displaced workers take them. Their labor makes the AI systems better. Better AI systems displace more workers. Those newly displaced workers join the annotation platforms. Repeat.

This is not a theoretical future scenario. It is, in embryonic form, already the structure of the labor market. The workers in the investigation did not choose data annotation as a career path. Many of them arrived at it after illness, job loss, or the collapse of a previous source of income. Krystal Kauffman's story is emblematic: she came to platform labor after illness made other work difficult. What she found there was not a safety net. It was another version of precarity, wearing the costume of flexibility.

The loop has a particular cruelty to it that goes beyond the economic. The workers who train AI are not doing unskilled work. They are applying judgment, domain knowledge, and ethical reasoning to tasks that the AI cannot yet reliably do itself. Jen's Ivy League doctorate is not irrelevant to her annotation work, it is directly applicable. The models she helps improve will eventually be capable enough to perform tasks that currently require that level of education. When they are, she will be displaced from annotation work too, by the very systems she helped train.

Organizing Back: The Leverage Data Workers Actually Have

The investigation does not end at diagnosis. It traces the early, fragile, but real attempts by data workers to build collective power in a labor market designed to prevent it, and it identifies the structural leverage that workers in this economy actually possess, even if they have rarely been in a position to use it.

Turkopticon, the tool built by Krystal Kauffman and collaborators for Amazon Mechanical Turk workers, is the oldest and most legible example. In a marketplace where requesters are anonymous and workers are interchangeable, Turkopticon created a ratings layer, a mechanism for workers to warn each other about bad-faith requesters, to share information about which tasks paid fairly and which were designed to waste time, and to introduce a small amount of collective accountability into a platform that otherwise had none. It did not transform the economics of Mechanical Turk. But it established the principle that workers in disaggregated digital labor markets can build tools for mutual aid and information sharing.

The Department of Labor investigation into Scale AI is a more direct form of institutional pressure. If the investigation concludes that Scale AI misclassified workers as independent contractors to avoid minimum wage and benefits requirements, the precedent could force a reclassification across the industry, potentially requiring platforms like Scale AI, Surge AI, and their competitors to treat annotation workers as employees, with the protections and costs that entails. The industry has every reason to fight such a reclassification vigorously. But the investigation exists, which means the legal vulnerability exists.

The garment industry coalition model offers a longer historical analogy. The garment workers of the early twentieth century were, in structural terms, not unlike AI data workers today: disaggregated, paid by the piece, dependent on intermediary contractors who competed on price, and isolated from each other in ways that made collective action difficult. The Triangle Shirtwaist Factory fire of 1911, the most visible moment in that history, produced regulatory and organizing momentum that eventually transformed the industry's labor conditions. The transformation was not fast, not complete, and not without backlash. But it happened.

The leverage that data workers have, and that most have not yet been in a position to exercise, is this: the AI systems they train break without them. Not immediately, not in ways that are obvious to the end user, but the feedback loops that keep these models aligned, helpful, and safe depend on continuous human evaluation. A coordinated work stoppage by a significant fraction of the annotation workforce would not crash a model overnight. But it would, over time, degrade the systems that depend on human oversight in ways that the companies involved would be unable to conceal. That leverage has never been organized at scale. It is not clear it ever will be. But it exists.