On September 17, 2026, a stack of newly unredacted court papers turned the most important AI copyright case in America into front-page news all over again. In a summary-judgment brief filed in The New York Times Company v. Microsoft Corporation, lawyers for The New York Times quoted a Microsoft research director who had privately described AI training on other people's work as "an astonishing theft of unprecedented proportions" and "perhaps the largest theft of labor in human history." Within a day the line was in Reuters, the Washington Post, the Financial Times, TechCrunch, Engadget, and dozens of other outlets. Nearly three years after it was filed, the case is now heading toward the ruling that may decide whether training AI on copyrighted work without permission is legal.
The case, No. 1:23-cv-11195, was filed December 27, 2023 in the U.S. District Court for the Southern District of New York. It is now the lead case in a consolidated group of lawsuits before Judge Sidney H. Stein that includes other news publishers, among them the Daily News, as well as book authors. The defendants are Microsoft and OpenAI. The claim is copyright infringement under 17 U.S.C. § 501, plus a related claim that the companies stripped copyright-management information out of the Times' articles.
What the lawsuit says
The Times' theory has stayed the same since 2023. OpenAI and Microsoft copied millions of its articles in full, without permission, to build commercial products: ChatGPT and Microsoft's Copilot. Those products, the Times says, now compete with the journalism they were built from. The newspaper is asking for liability on four kinds of copying: scraping its articles (most of them behind a paywall), training models built to predict that expressive text, using its content in real-time chatbot answers, and the two companies sharing the copied material with each other. Press reports put the publishers' count at roughly 10.8 million articles taken.
On September 4, 2026, both sides moved for summary judgment, which means each asked Judge Stein to decide key questions without a trial. OpenAI asked the court to rule that pretraining and its "Browse" feature are fair use. It also argued that copies Browse made before publishers updated their robots.txt files were impliedly licensed, and that the copyright-management claim fails as a matter of law. Microsoft asked for judgment on every claim against it. The defendants rely heavily on two 2025 decisions from federal courts in California, Kadrey v. Meta and Bartz v. Anthropic, in which judges found that training AI models on books was transformative fair use. OpenAI also reportedly told the court it found only 24 verbatim reproductions of Times content across 20 million ChatGPT logs. The U.S. Justice Department filed a brief backing OpenAI, arguing that training on millions of works produces a "transformative public benefit" that outweighs harm to publishers.
The September 17 unsealing gave the Times' side of the story. According to the partly unredacted brief, as reported by TechCrunch, Bloomberg Law, 404 Media, and others:
- A January 2023 internal Microsoft memo by Director of Applied Science Brent Hecht warned that people would "soon consider large models 'hoovering up' all their work to be an astonishing theft." Hecht also wrote that a fair-use win would "make a complete mockery of the idea of 'fair use.'"
- A later Microsoft document described a "doom loop": "It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business." Microsoft's own data reportedly showed Copilot cutting click-through to nytimes.com by as much as 93% compared with ordinary Bing search.
- Nick Turley, OpenAI's head of ChatGPT, wrote that its products are "largely substitutive, period" and an "existential threat" to publishers.
- Microsoft CEO Satya Nadella testified that "anything that is paywalled should be licensed," and that if he had known OpenAI trained on paywalled material he would have required it to retrain its models.
- When a researcher flagged "a hack to get around nytimes paywall," OpenAI president Greg Brockman allegedly replied, "ah nice."
Microsoft says Hecht's comments "reflect one employee's individual perspective, are not a legal analysis, and do not represent the company's views," and that Copilot is not a substitute for journalism. TechCrunch noted that most of the quotes come from the Times' own brief, while the underlying exhibits remain sealed, so they are shown without their full context. Nothing has been decided. These are allegations and advocacy on both sides.
The hard part: fair use turns on substitution, not copying
The quotes got the headlines, but the legal fight is narrower, and more useful to understand. Nobody seriously disputes that OpenAI copied the Times' articles. The question is whether that copying is fair use, which courts decide by weighing four factors: the purpose and character of the use, the nature of the work, how much was taken, and the effect on the market for the original.
The California rulings the defendants rely on leaned on the first factor. Training, those judges said, is "transformative" because a model learns patterns instead of republishing the books. But the Supreme Court's 2023 decision in Andy Warhol Foundation v. Goldsmith made clear that being transformative is not enough when the new use serves the same purpose as the original and competes with it. That is why the Times has built its case around the fourth factor, market substitution, and why lines like "largely substitutive, period" and "users won't click" matter so much. The Times argues that the California authors never showed AI outputs replacing their books. A newspaper whose traffic collapses when a chatbot answers the question instead may be able to show exactly that.
The second issue is just as relevant for creators: how the copies were obtained. Getting around a paywall and stripping copyright notices are not just bad optics. Removing copyright-management information can create separate liability under the DMCA (17 U.S.C. § 1202), and courts weighing fair use have started looking hard at whether the source material was taken from pirate libraries or from behind access controls. In Bartz v. Anthropic, the same judge who called training fair use let the claims over pirated books go forward. That exposure led to a reported $1.5 billion settlement.
Why it matters if you make films, not newspapers
It is easy to treat this as a fight between giants. But how Judge Stein rules on substitution will shape what every creator can do when an AI system is trained on their screenplay, their footage, their score, or their poster art. Here is what to take from it now:
Register your copyrights, and register early. The Times can seek statutory damages of up to $150,000 per work for willful infringement because its works are registered. For a film, that means the screenplay, the finished picture, and original music. Unregistered works usually cannot get statutory damages or attorney's fees, which often makes a lawsuit uneconomical for an independent creator.
Keep your copyright information attached to your work. Copyright notices, credits, watermarks, and file metadata are "copyright-management information." Stripping them knowingly is a separate violation under § 1202. Put them on screeners, EPKs, trailers, and stills, and keep them in your files. You can't enforce a right that has been quietly erased from your work.
Control access and write down your terms. The fight over paywalls and robots.txt is a fight over intent. Password-protected screeners, watermarked links, and clear terms of use all show that you never offered your work for free use. OpenAI's implied-license argument depends on publishers having left the door open.
Put AI language in your contracts. Distribution, sales-agent, and licensing agreements increasingly address whether your film can be used for AI training, and who gets paid if it is. If your deal doesn't mention it, ask. A licensing market for training data, which the Times itself joined when it licensed content to Amazon in 2025, is evidence that unlicensed use causes real market harm. It also gives you a potential new revenue stream.
Your chain of title is your evidence. Each of these arguments depends on being able to prove what you own, when you created it, and what you have and haven't licensed. Clean, documented rights are what let a distributor, an insurer, a collection account manager, or a court treat your film as a protected asset.
Opposition briefs on the cross-motions are reportedly due in late October, with replies in November. Judge Stein will then decide whether all or part of the case goes to trial, which could happen in 2027. Whatever he decides will likely be the most-cited AI copyright ruling to date, and it will reach well beyond newsrooms. We'll track the briefing and any ruling as it lands. Follow the filings, parties, and coverage on the case page.
This post is editorial commentary on public court filings and news coverage, not legal advice. The allegations described are unproven and the defendants dispute them. Quoted internal statements come from a party's brief and may lack context, and being named as a defendant is not evidence of liability. Details come from press reports on the docket and may be refined as the case develops.