The New York Times v. OpenAI: The Fair-Use Question the Bartz Settlement Left Open

When three authors settled Bartz v. Anthropic for $1.5 billion, they closed the first great AI copyright case before a jury could rule on the merits: the court had held that training on books was fair use, and the settlement bought off only the separate sin of having pirated the copies. The question of whether the training itself is lawful was left standing. The case most likely to answer it is the one the Times filed in December 2023, now the lead action in a consolidated docket of a dozen newspaper and author suits before Judge Sidney H. Stein in Manhattan.⁶

Two things make this the case to watch. In January 2026 the court handed the plaintiffs an extraordinary discovery win, ordering OpenAI to produce twenty million de-identified ChatGPT conversation logs.¹ And in June 2026 the Times sharpened its theory, dropping a weaker claim and recasting Microsoft not as a bystander but as the party that built the supercomputer to make the copying possible.⁴ By the end of this piece you should be able to say precisely what a court has already decided here, what it has not, and why the fair-use verdict — when it comes — will be the one that governs the news.

What the court has already held

The defendants moved to dismiss in early 2024. On 26 March 2025 Judge Stein rejected most of the motion, letting the core copyright claims proceed to a jury; from the bench he did not release an opinion but promised one "expeditiously," and the Times said it welcomed the chance to "present a jury with the facts."²

The written opinion, filed days later, drew the lines that still shape the case. Direct and contributory infringement liability for materially contributing to someone else's infringement while knowing of it — here, the theory that ChatGPT itself emits copies of Times articles and that OpenAI had reason to know it would both survived. What the court trimmed was the Digital Millennium Copyright Act theory: Stein dismissed the §1202(b) "removal of copyright-management information" claim against OpenAI in the Times action, reasoning that treating a paraphrased excerpt as a distributed "copy" would risk "boundless DMCA liability," while letting the same claim stand in the parallel Daily News and Center for Investigative Reporting suits.³ The surviving spine is the ordinary one: reproduction during training, and reproduction in output.

Why discovery became the battlefield

A fair-use defence is decided on facts, and the facts here are usage data. In July 2025 the news plaintiffs moved to compel a sample of 120 million consumer ChatGPT logs; OpenAI countered that a 20-million sample — roughly half a percent of the tens of billions of logs it retains — was "surely more than enough," and the plaintiffs accepted that number.¹ The fight then moved to a subtler ground. In October 2025 OpenAI said it would not hand over the whole 20-million sample after all, proposing instead to run search terms across it and produce only the conversations that mentioned the plaintiffs' works.

Magistrate Judge Ona Wang refused, and on 5 January 2026 Judge Stein affirmed her, in an order worth reading for its reasoning rather than its result. The court held that even "output logs that do not contain reproductions of News Plaintiffs' works may still be relevant to OpenAI's fair use defense" — because the frequency with which ordinary prompts do not reproduce the news bears directly on whether the model is a market substitute for it.¹ Users' privacy interests were real, the court found, but "one factor in the proportionality analysis," adequately protected by three safeguards: cutting the set from tens of billions to twenty million, OpenAI's de-identification tool, and the existing protective order. OpenAI's lead authority, a Second Circuit wiretap case, was distinguished on the ground that ChatGPT users "voluntarily disclosed" their conversations, unlike the targets of a wiretap.¹

OpenAI has cast the demand as an assault on user privacy, arguing that the Times wants it to retain even deleted chats indefinitely and that it will store the historical data locked down, audited, and handed to no one "at this time."⁸ That framing matters for public opinion; it did not move the court, which treated the logs as ordinary business records held in the normal course.

The turn toward Microsoft

The consolidation itself changed the stakes. What began as three newspaper suits is now a single multidistrict proceeding gathering the Times, the Daily News and other Alden titles, the Center for Investigative Reporting, the Authors Guild class, and further publishers under one judge and one discovery schedule.⁶ A finding on fair use in this forum will not be confined to one plaintiff.

On 25 June 2026 the Times moved for leave to file a third amended complaint that narrows and sharpens. It dropped a claim that OpenAI "secondarily" infringed by failing to stop users from generating protected text, and expanded its contributory-infringement case against Microsoft, alleging that Microsoft did not merely benefit from the training but actively encouraged it — building bespoke supercomputing infrastructure and providing "services" tailored to download copyrighted works at scale.⁴ "Microsoft actively encouraged OpenAI to steal our copyrighted works," a Times spokesman said; "our core claims remain the same from the day we filed."⁵ The legal logic is deliberate: inducement liability reaches the party that supplies the means with knowledge of the infringing end, and Microsoft is both the deeper pocket and the owner of the Azure hardware on which the models were trained.

The fair-use crux, steelmanned

Strip away the discovery skirmishes and one question remains, the same one Bartz settled around. OpenAI's strongest case is that training is a "transformative, non-expressive" use: the model learns statistical structure, not prose, and reproduces articles only when a user works to extract them, which is a misuse rather than the product. It builds, in its words, on "publicly available data, in a manner grounded in fair use."²

The Times's strongest case is the fourth fair-use factor: market effect. Its complaint documents ChatGPT reproducing article passages near-verbatim and, more damaging, substituting for the paper — a chatbot that answers "what did the Times report" is competing with the subscription, and the paper has a live licensing market that Amazon and others already pay into.⁹ The DMCA ruling already gives a preview of how Stein reasons about the line between learning and copying, and the logs order tells you the jury will see how often, in real use, the model actually reproduces the news.³ That is why this case, not the settled ones, is the one to watch: it is on a path to a verdict, and the verdict will be about the news.

Go deeper

Read Judge Stein's 5 January 2026 order affirming production of the twenty-million-log sample — four pages, and the clearest window into how the court weighs relevance against user privacy.¹ Then read the March 2025 motion-to-dismiss opinion for the "boundless DMCA liability" passage and the contours of the surviving claims.³ Track the live docket on CourtListener, where the third-amended-complaint motion and OpenAI's response will land, to watch whether the Microsoft-inducement theory survives.⁷ The number to watch for an engineer: what fraction of the twenty million logs, once analysed, actually reproduces protected text — that ratio is the market-substitution argument in empirical form.