AI Copyright Lawsuits Mount After Microsoft 'Theft' Memo Unsealed

The Quote Now Driving the AI Copyright Fight

In mid-September, a judge in Manhattan unsealed filings in the newspaper industry's copyright case against OpenAI and Microsoft. One line from those filings spread quickly. In an internal document written shortly after the suit was filed, Microsoft's Director of Applied Science, Brent Hecht, warned that millions of people would soon see large models "hoovering up" their work as "an astonishing theft of unprecedented proportions," and possibly "the largest theft of labor in human history."2 He also wrote that almost nobody who made the content expected it to be used this way, and that they are not being paid for it.2

The Hecht line was not the only one. The plaintiffs also quote OpenAI cofounder and president Greg Brockman. According to the filing, he called generative AI an "existential threat" to publishers and described the products as "largely substitutive, period", adding that they would become more so as they improved.5 The motion also says Hecht wrote that a win for the defendants on fair use could "make a complete mockery of the idea of 'fair use.'"5 Three weeks later, USA Today Co. and its local papers sued OpenAI for more than $250 million.13

That timing has encouraged a story that one admission opened the floodgates. The record supports a narrower but more lasting conclusion. The floodgates were already open. What these admissions change is the kind of evidence plaintiffs now look for and use.

What the Unsealed Filings Allege

The material comes from summary judgment briefing in consolidated litigation brought by The New York Times, the New York Daily News, the Orange County Register, the Chicago Tribune and other papers.13 The Authors Guild and a group of best-selling writers are part of the same consolidated proceeding.2

The newspapers' numbers are large. The unredacted filing alleges that more than 3.9 million copies of Times content were used for training, along with 7.3 million copies from the Daily News, the Register and their sister papers.1 It says Microsoft measured drops of 83% to 93% in clicks on links to those publications.1 It cites survey data in which 36% of Times subscribers who use ChatGPT said they no longer needed the Times at all.1 It also quotes an OpenAI internal estimate from February 2026 that ChatGPT was receiving about 1 million prompts a week seeking local news.2

The plaintiffs say the internal papers also describe how the content was gathered. According to their account, it was obtained by quietly getting around paywalls, scraping at scale, and deliberately removing copyright notices from training data.6 The Daily News reported a further allegation: that the companies copied "billions of web pages" and traded them with each other in "horse trading"-style deals instead of licensing from publishers. Microsoft did not answer a question about that claim.2 The papers are also asking for sanctions against OpenAI over alleged destruction of evidence.3

Another Microsoft document cited in the filing may be more damaging over time than the "theft" quote. It warns that the company's AI content strategy had started a "doom loop" that would hurt both its models and the wider web. It also notes how unusual it is for a product to threaten "the economic foundations of its essential suppliers."3

Where the Coverage Splits

The newspapers' own reports, published by outlets that are plaintiffs in the case, present the documents as proof of knowing theft. Daily News attorney Steven Lieberman said the world can now see what the defendants "thought all along about the fairness of their own behavior."1 Tech outlets such as TechCrunch and Futurism took a similar line. They described the statements as cutting against the market-harm part of fair use and as plaintiffs "arguing their case for them."56

Microsoft's response is narrow. A spokeswoman said Hecht's comments reflect "one employee's individual perspective" and are not legal analysis. She said the "doom loop" document was about broad changes in how people find information, not the copyright questions before the court.1 In a declaration filed with its own summary judgment motion, another Microsoft executive described Hecht as someone paid to play the contrarian.1

One explainer urged caution. It stressed that the Hecht quote is an allegation in a party's filing, not a court finding, and that the internal messages describing the data practices as "sketchy" are the plaintiffs' characterization, not an admission by the companies.8 That point deserves weight. A research director's privately expressed worry is not the same as a corporate admission of liability, and judges decide fair use on the four statutory factors, not on what employees wrote in internal documents.

Still, Microsoft's defense addresses tone more than substance. The click-through declines, the subscriber surveys and Brockman's "substitutive" remark go to market substitution, which is the fourth fair use factor. Earlier rulings have treated that factor as decisive.

Why Substitution Evidence Is Now Decisive

The two major book-training rulings from 2025 show why this evidence matters. In Bartz v. Anthropic, the court found that training on books was fair use. It reasoned that the fourth factor guards against replacement, not competition, and noted that the authors had not alleged Claude reproduced their text. Storing pirated copies, however, was not protected.14 That case settled for $1.5 billion, roughly $3,000 per work.14 Anthropic still says publicly that training on books is fair use.20

In Kadrey v. Meta, the judge ruled for Meta on training. He made clear this did not mean Meta's use was lawful, only that the plaintiffs had made the wrong arguments and had not built a record on market dilution.14

Taken together, those opinions told future plaintiffs what to prove: show that the product actually replaces the original. The September filings try to do exactly that, with the defendants' own documents. Seen this way, the Hecht quote gets the attention, but the click and subscription data are what could decide the case.

The Third Circuit's September 29 decision in Thomson Reuters v. Ross Intelligence adds to the pressure. It upheld findings that Westlaw headnotes are copyrightable and that Ross's use of them to train its legal search tool was not fair use.15 It has been described as the first federal appellate ruling on fair use in AI training.1516 Its reach is limited. Ross built a non-generative tool, copied the material before generative AI was widespread, and was building a direct competitor.15 Ross's remaining principals have said they will ask the Supreme Court to review it.16

For legal tech in particular, the ruling is a warning. Vendors building research tools on proprietary editorial content now have appellate precedent against them. Generative AI companies will argue the case does not apply to them, but plaintiffs now have an appeals court decision to cite that they did not have before.15

The Wave Was Already Under Way

The idea that one admission caused the surge does not fit the dates. One public tracker counts 248 lawsuits against 76 AI companies. Of those, 145 are copyright cases, and filings in 2026 are already on pace with 2025's total of 86.15 OpenAI is named in 55 cases, nearly one in four.15 Before USA Today sued, OpenAI already faced claims from The Intercept, Ziff Davis, CBC/Radio-Canada, Encyclopaedia Britannica, Merriam-Webster, The Seattle Times and a coalition of almost 400 local papers.13 wikiHow filed its own suit in August.7 Shareholder derivative suits have also been filed against Apple's and Microsoft's leadership over training on unlicensed material.7

The USA Today complaint does show how plaintiffs are adjusting. It identifies more than 160,000 entries from its papers' websites in OpenAI's WebText dataset. Like the Times filings, it relies on internal communications that it says show OpenAI deliberately targeted news content.11 It also asks the court, under Section 503(b) of the Copyright Act, to order the destruction of GPT models and training datasets containing its journalism.11 That is a remedy aimed at the models themselves, not only at damages.

Not every court has sided with rights holders. The Ninth Circuit recently upheld dismissal of DMCA claims against GitHub Copilot. It held that AI outputs are new works, not copies with copyright information removed, and it refused to "transform run-of-the-mill copyright infringement claims into DMCA claims."9 One commentator said a different outcome would have opened its own floodgates.9 Since USA Today is also bringing DMCA claims over stripped bylines and titles, that precedent could limit part of its case.11

Washington Sides With AI Developers

The federal government's position runs against the publishers. The Justice Department, which is not a party, filed a statement of interest in the OpenAI multidistrict litigation supporting fair use for AI training.9 The newspapers' coverage described it as urging courts to back the tech companies so the US can win the global AI race.1 One analysis noted that the statement addresses the news-publisher claims, not the separate factual allegations in the authors' filings.8

The result is a split. The executive branch is backing a broad fair use reading while discovery keeps producing evidence that cuts against it. Cross-motions for summary judgment in the Times case were filed September 4.15 Other rulings are close: Concord Music v. Anthropic is set for an October 21 hearing, and the Mosaic LLM litigation for October 30.18

The Bottom Line

The internal documents did not create the AI copyright fight, and they do not decide it. What they show is that the outcome probably depends on whether plaintiffs can prove substitution with the defendants' own data, rather than on the abstract question of whether training is "transformative." Both Microsoft and OpenAI apparently measured that harm internally. Settlements like Anthropic's suggest defendants would prefer to pay than risk a precedent.15 If the Times case reaches a ruling, the click-through figures may matter more to the judge than the theft quote.