OpenAI, Microsoft and News Publishers Ask a Judge to Decide AI Fair Use Before Trial

Both sides in the highest-stakes copyright fight over AI training asked a federal judge on 4 September to resolve the central question of the case without a trial, setting up a ruling that could determine how American law treats the ingestion of published work into large language models.

Related reading: British Columbia Sues OpenAI and Sam Altman Over the Tumbler Ridge School Shooting · OpenAI, Anthropic, Google and Meta Testify Under Oath as New York Weighs Ten AI Bills.

OpenAI, Microsoft and a group of news publishers each filed motions for summary judgment before Judge Sidney H. Stein in the Southern District of New York, according to court filings and reporting by Reuters. The dispute, consolidated as a multidistrict litigation, pits the companies against The New York Times, the Daily News Plaintiffs, Ziff Davis, the Center for Investigative Reporting and The Intercept.

The filings describe two records that barely intersect. OpenAI says it copied almost nothing in a way that matters. The publishers say it took everything.

Two versions of the same pipeline

The plaintiffs identify more than 10.8 million asserted works, according to OpenAI's brief. The Times alone asserts 6,030,928 articles published between 1928 and 2024. The publishers' combined brief argues that the companies "repeatedly copied" entire articles without permission to build "substitutive commercial AI products," and it leans on internal Microsoft and OpenAI communications, portions of which are redacted.

The publishers frame the harm in market terms. AI assistants, they argue, now answer questions that readers once brought to news sites, capturing the audience and the advertising that paid for the reporting. Their brief invokes Andy Warhol Foundation v. Goldsmith for the proposition that substitution is "copyright's bête noire," and cites Google v. Oracle and Authors Guild v. Google as precedents that permit only narrow, functional kinds of copying. They also point to Kadrey v. Meta, where Judge Vince Chhabria wrote that news publishers "present even stronger arguments against fair use" than the book authors in that case.

OpenAI's response is that the copies serve a transformative purpose and that the market evidence does not support the publishers' theory. Its brief argues that pretraining is fair use because it learns statistical patterns and produces new text rather than distributing the underlying articles, and that the works at issue are factual news reports, which receive thinner protection than expressive fiction or song lyrics.

The regurgitation number

The most concrete dispute is about how often the models reproduce training text verbatim. OpenAI's expert searched a sample of 20 million ChatGPT conversation logs produced in discovery and found 24 instances of verbatim regurgitation, a rate of 0.00012 percent. The brief notes that the publishers' own experts produced figures in the same range: 0.00011 percent for the Times and Daily News, 0.00002 percent for Ziff Davis, 0.000008 percent for the Center for Investigative Reporting and 0.000002 percent for The Intercept. The longest passages identified ran to 29 and 43 words.

The publishers treat those numbers differently. Their brief argues that regurgitation is the wrong lens, because the injury comes from the products competing with the originals, not from occasional verbatim output. They also press a claim under the Digital Millennium Copyright Act, alleging that the removal of copyright management information concealed infringement. OpenAI counters that its expert reviewed all 68 alleged regurgitations in the parties' opening reports and found that the source would have been obvious to a user in 63 of them, and that a user who supplies the source text to prompt an output cannot claim attribution was concealed.

The browse question and the licensing window

A second front concerns Browse, ChatGPT's web-browsing feature. OpenAI argues that copies Browse makes to extract facts are fair use, and that any copies made before publishers updated their files to block the crawler were impliedly licensed. The Times waited nearly a year after ChatGPT's launch before blocking OpenAI's user agent and implemented a hard block for ChatGPT-User in April 2024. The Intercept blocked in January 2024. On that record, OpenAI argues that the early copies fell under the line of cases running from Field v. Google.

The publishers see the same timeline as evidence of disregard. Their brief describes the companies as bypassing paywalls and other protective systems, and cites internal messages, several of them redacted, to argue that employees understood the access restrictions and worked around them.

How we got here, and why the timing matters

The motions are the culmination of more than two years of litigation and discovery. The Times sued in December 2023, and other publishers joined. In August 2026, Judge Stein dismissed with prejudice the news plaintiffs' contributory infringement claims based on a "material contribution" theory, following the Supreme Court's decision in Cox Communications v. Sony Music Entertainment, which held that material contribution is not a valid basis for contributory liability. The judge also denied the Times and Daily News leave to amend their complaints to add inducement theories at a late stage, an order dated 6 August.

The commercial relationship at the center of the case has shifted during the litigation. Microsoft and OpenAI amended their agreement on 27 April 2026, dropping revenue-share payments and converting their intellectual property licence from exclusive to non-exclusive. Reports in ppc.land note that the amended deal gives each company more room to work with other partners, a detail that may matter to the damages phase even if it does not touch the fair-use analysis.

The wider legal picture is unsettled. Anthropic agreed to a $1.5 billion settlement in a separate authors' case in September 2025, the largest figure attached to a sector copyright claim. Meta won summary judgment on fair use in June 2025, though the judge described the record as thin. The Justice Department filed a statement on 3 September supporting OpenAI's fair-use defence, the day before the motions landed, a sign that the outcome is being watched well beyond the parties.

The four factors, and where the case turns

Fair use is a four-factor test, and the briefs argue over each. On the first factor, the purpose and character of the use, OpenAI says pretraining is transformative because it learns statistical relationships from text rather than republishing it. The publishers say the use is commercial and substitutive, and that a product which answers the same questions is not transformative at all. The Supreme Court's 2023 decision in Andy Warhol Foundation v. Goldsmith, which narrowed the definition of transformative use, is central to both arguments.

The second factor, the nature of the work, favours OpenAI more than it would in a case about novels or music. News articles are factual and published, and courts extend thinner protection to factual works. The publishers respond that individual articles are expressive, that their selection and arrangement is original, and that describing an entire corpus as facts understates the editorial work involved.

The third factor, the amount taken, is where the records diverge most sharply. OpenAI points to the minute regurgitation rates as evidence that the models do not reproduce the works. The publishers argue that copying articles in full for training is by definition taking the whole work, whatever the model later emits. OpenAI's brief concedes that entire works were ingested but argues that the scale does not decide the question if the use is transformative.

The fourth factor, market harm, is likely to decide the case. The publishers say their licensing market for AI training is real and growing, and that unlicensed training undermines it. OpenAI says the publishers have not shown that ChatGPT displaces subscriptions or advertising revenue, and that the deals publishers have signed with other AI companies show a market exists rather than that it was harmed by copying.

Damages, if the defence fails

If Judge Stein rejects fair use, the case does not end; it moves toward a trial on liability and remedies. The publishers are seeking statutory damages calculated per work, and with more than 10.8 million asserted works the arithmetic is daunting. Statutory damages for non-willful infringement can reach $30,000 per work and up to $150,000 for willful infringement, though courts have discretion and mass infringement cases rarely produce awards at the statutory maximum for every item. The publishers are also seeking a ruling that damages run per article rather than per issue, a distinction that matters enormously when a single newspaper edition contains dozens of articles.

What the ruling would change

A grant of summary judgment on fair use for either side would end the case's central theory before trial and would likely shape how every other pending AI copyright suit is litigated. If Judge Stein finds that training on copyrighted news is transformative fair use, the remaining claims would narrow to output-level and licensing questions. If he rejects the defence, the case moves toward a trial on liability and, potentially, statutory damages calculated per work, a structure the publishers are seeking because 10.8 million articles is an enormous multiplier.

Two open questions are likely to drive any ruling. The first is how the court weighs the fourth fair-use factor, the effect on the market for the work, in a world where the publishers have begun selling their own AI licensing deals. The second is how much internal company commentary about copying will be treated as evidence of commercial purpose. The redactions in both briefs show that the parties consider those documents decisive, even if the public cannot yet read them.

For now, the only certainty is that the question of whether AI training is fair use is no longer hypothetical. It is briefed, and it is waiting on one judge in Manhattan.