OpenAI and Microsoft Insiders Called Training Data "Theft" in Internal Messages, Court Filings Show

Internal communications and deposition testimony from Microsoft and OpenAI employees are complicating the fair use argument the two companies have relied on to justify training AI models on copyrighted material. According to records surfaced in ongoing litigation, a Microsoft director described the industry's data-collection practices as the "largest theft of labor in human history," while OpenAI's head of ChatGPT wrote in an internal message that the company's products "are largely substitutive, period." The statements, first reported by The Decoder, were produced as evidence in copyright cases against the companies.

The fair use doctrine, codified in Section 107 of the U.S. Copyright Act, turns on four factors, including the purpose of the use and its effect on the market for the original work. Companies defending their training pipelines have argued that ingesting copyrighted text and images is transformative and does not substitute for the originals. The internal language cuts against that framing on both counts: "theft" speaks to the taking itself, and "substitutive" speaks directly to the fourth factor, which weighs market harm.

The Microsoft director's remark is not a stray aside from a junior employee. The quote, recorded in company communications, uses the language of labor appropriation, which is the theory plaintiffs have pressed in cases brought by authors, visual artists, and news organizations. If a jury or judge accepts that the company's own personnel viewed the practice as theft, it becomes harder for counsel to argue that the same conduct was a good-faith, transformative use of publicly available material.

The ChatGPT executive's assessment is arguably the more damaging of the two for the defense. Fair use analysis asks whether the new product serves a different purpose from the copyrighted work, and an internal admission that it substitutes for the original undercuts that argument from the inside. Plaintiffs have long argued that models trained on books, articles, and code can reproduce or replace the market for those works; the message suggests at least one senior OpenAI figure agreed with the substance of that claim.

Neither company has commented publicly on the specific quotes, and both are expected to argue that internal characterizations do not determine the legal question. Judges have so far split on fair use in AI training disputes, with some rulings favoring defendants on the transformative-use prong while leaving market-harm questions unresolved. The documents now give plaintiffs a concrete record to cite when courts weigh whether the training was fair or simply unauthorized.

What happens next depends on how the presiding judges treat the evidence. If the internal statements are admitted and credited, they could shift the balance in cases where the fair use defense has been the industry's primary shield, and they may push companies to strike licensing deals rather than litigate. The open question is whether a handful of candid internal messages can do what years of public argument have not: force the AI industry to pay for the material it trains on.