New York Times v OpenAI: Case Status and Key Rulings

The New York Times sued OpenAI and Microsoft on 27 December 2023 in the Southern District of New York. More than two years later the case has produced significant rulings on pleading and on discovery, and none at all on the question everyone is waiting for. Keeping those apart is the point of this page.

What was filed

The complaint alleges that OpenAI and Microsoft copied millions of Times articles to train large language models, that the models can output verbatim or near-verbatim passages from those articles, and that the resulting products substitute for the Times’s own offerings. The causes of action include direct copyright infringement, contributory infringement, vicarious liability, removal of copyright management information under the DMCA, trade mark dilution and unfair competition by misappropriation. The complaint is notable for attaching exhibits of alleged verbatim regurgitation, which converts a general argument about training into a specific evidentiary claim about outputs.

Dockets and filings in the case are accessible through CourtListener’s RECAP archive. Read the orders rather than coverage of them; the gap between what an order says and what it is reported to say has been wide in this case.

Consolidation into an MDL

In April 2025 the Judicial Panel on Multidistrict Litigation centralised the copyright actions against OpenAI, which had been filed in several districts, into a multidistrict litigation in the Southern District of New York before Judge Sidney Stein. The Times’s case, along with actions brought by other news organisations and by authors, proceeds within that structure.

Consolidation is a case management device and decides nothing on the merits. It matters practically for two reasons: coordinated discovery means a ruling on a discovery dispute affects all the constituent cases, and a single judge’s view of the common legal questions now carries much further than it otherwise would.

The motion to dismiss ruling

Judge Stein ruled on the defendants’ motions to dismiss in late March 2025, with an opinion following in early April 2025. He denied the motions in substantial part, allowing the core direct infringement claims and the contributory infringement claims to proceed. Some claims did not survive: the DMCA copyright-management-information claims were dismissed, as were certain state-law and unfair competition theories on preemption grounds. Arguments that the claims were time-barred were addressed rather than accepted wholesale.

What a denial of a motion to dismiss means is narrow and constantly misreported. At that stage the court assumes the complaint’s factual allegations are true and asks only whether they state a legal claim. It is not a finding that anything alleged happened, not a finding of infringement, and not a rejection of fair use — fair use is an affirmative defence that is generally resolved on a developed record, at summary judgment or trial, not on the pleadings. The ruling establishes that the theory is legally cognisable and that the case continues.

The contributory infringement holding is the most consequential piece. Allowing a claim that the defendants are liable for infringing outputs produced by users keeps in play the argument that a model provider bears responsibility for what its users can get the system to emit, which is a different and broader exposure than the training claim.

The output log preservation order

In May 2025 a magistrate judge ordered OpenAI to preserve output log data that would otherwise have been deleted, including data covered by user deletion and by consumer retention settings, on the basis that it was potentially relevant evidence about regurgitation. OpenAI objected publicly and procedurally, arguing that the order conflicted with user privacy expectations and with its own commitments, and the scope of the obligation was subsequently the subject of further litigation and modification.

The general mechanism is worth understanding independently of this case: a litigation hold obligation can override a product’s ordinary deletion behaviour, and it attaches to the party in litigation. For anyone building on a provider’s API, this is a live consideration — your data retention position is a function of your provider’s legal exposure as well as of its policy, and it interacts directly with the storage limitation and deletion commitments discussed in the retention limits page.

What remains open

Everything that matters, which is the honest summary.

• Fair use. Not decided. The defendants’ principal defence has not been ruled on, and the analysis will turn on a record that did not exist at the pleading stage.
• Whether any output infringed. Not decided. The complaint’s regurgitation exhibits are allegations; the defendants dispute how they were produced and whether they are representative.
• Whether training copies were made as alleged. A matter for discovery and proof.
• Remedies. Nothing has been awarded, and the remedies sought include relief that would be extraordinary if granted.

Two nearby data points frame how little can be inferred. District courts elsewhere have reached fair-use conclusions in both directions on different records in 2025, discussed in the Copyright Office Part 3 page, and the one American case to have produced a merits fair-use holding against an AI defendant did so on facts far from generative AI — see Thomson Reuters v Ross. Anyone offering a prediction of this case’s outcome is offering an opinion, and the value of following the docket is that you do not have to rely on one.