Microsoft and OpenAI Workers Flag 'Largest Theft of Labor' in AI Training

A set of court documents made public this week pulled back the curtain on what workers and legal observers are calling a foundational problem in how the largest artificial intelligence companies build their systems. Inside Microsoft and OpenAI, according to the newly unsealed filings, there was documented concern about the scale and legality of using millions of news articles—published work created by journalists, editors, and newsrooms—to train the AI models that now generate billions in corporate value.

The documents do not reveal a casual oversight. They show internal awareness, at both companies, that the practice raised serious questions. News organizations and their employees had not consented to this use. They had not been compensated. The articles were taken wholesale from the internet and fed into machine learning systems designed to absorb patterns from human-written text so that the resulting AI could generate its own prose, answer questions, and perform tasks that increasingly compete with human labor.

What makes the unsealing significant is that it transforms what had been private corporate deliberation into public record. Lawyers, regulators, and journalists can now see that the companies themselves understood the stakes. The internal concern documented in these filings suggests the practice was not treated as routine or uncontroversial within the organizations that carried it out. If the companies had doubts, those doubts were not allowed to slow the work.

The case centers on a straightforward claim: that millions of published articles were incorporated into AI training datasets without authorization or payment. News organizations—which employ the writers, photographers, and editors whose work appears in those articles—have argued this constitutes both copyright infringement and a form of labor exploitation. The workers who created the content received nothing. The companies that published it received nothing. The AI companies that extracted value from it received everything.

This is not a small-scale dispute. The volume of material involved is enormous. Millions of articles means decades of accumulated journalism, the daily output of hundreds of newsrooms, all harvested and processed into training data. The economic model is stark: take freely, build commercially, pay no one.

The unsealing of these documents arrives at a moment when the AI industry faces mounting pressure on multiple fronts. Regulators in the United States and Europe are beginning to scrutinize how these systems are built. Other lawsuits are pending. And there is growing recognition, even among some technologists, that the current approach to training data—essentially, take what exists on the internet and use it—may not be sustainable legally or ethically.

What happens next will likely depend on how courts interpret the law and how much political will exists to enforce it. The documents themselves do not resolve the underlying questions. They simply show that the companies involved knew those questions existed and proceeded anyway. For workers in publishing and journalism, the unsealing represents a shift: their concerns are no longer just complaints from an industry losing relevance. They are now part of the official record of how the most powerful AI systems in the world were built.