The New York Times v. OpenAI: Why a Manhattan Judge Let the Output-Infringement Claims Stand

On March 26, 2025, Judge Sidney H. Stein of the Southern District of New York issued the ruling that the entire generative-AI industry had been waiting on: a decision on the motions to dismiss in the consolidated newspaper litigation led by The New York Times Company v. Microsoft Corp., No. 1:23-cv-11195 (SHS) (S.D.N.Y.). The court denied the heart of the defendants’ challenge, allowing the publishers’ direct and contributory copyright claims to proceed. For an industry that had grown accustomed to favorable pleading-stage rulings in California, the decision was a jolt — and for Los Angeles’s media, entertainment, and creator economy, it is the clearest signal yet that output-based infringement theories are viable.

At a glance

• Case: The New York Times Company v. Microsoft Corp., No. 1:23-cv-11195 (SHS) (S.D.N.Y.), consolidated with the Daily News and Center for Investigative Reporting actions
• Decided: March 26, 2025 (opinion and order on motions to dismiss), Judge Sidney H. Stein
• Survived: direct copyright infringement (including conduct outside the three-year window); contributory infringement; trademark dilution (Daily News); certain DMCA § 1202(b)(1) claims against OpenAI
• Dismissed: common-law unfair competition by misappropriation (with prejudice); most DMCA § 1202(b)(1) and (b)(3) claims (with leave to amend); a related “abridgment” theory
• Status: Pending — the case proceeds into discovery; fair use was not decided on the motion to dismiss

Outputs, users, and contributory liability

The most important survival is the contributory-infringement claim. The publishers allege not only that their articles were copied during training, but that ChatGPT can be prompted to regurgitate their content verbatim or near-verbatim, and that OpenAI and Microsoft are secondarily liable when users elicit and use those outputs. Judge Stein held those allegations sufficient to proceed, finding it plausible that the defendants knew of and materially contributed to user infringement.

This is precisely the theory that several California judges had treated skeptically, dismissing output-based claims where plaintiffs could not point to specific infringing generations. By letting contributory liability advance, the Southern District of New York opened a doctrinal divergence between the coasts. The split is not merely academic: it shapes forum strategy, settlement leverage, and the eventual circuit-level questions that will define how the Copyright Act applies to model outputs.

The statute-of-limitations ruling

A quieter but significant holding concerns timing. The defendants argued that claims based on conduct more than three years before the complaints were filed — including 2019–2020 training activity — were time-barred. Judge Stein rejected that defense at the pleading stage, allowing the older conduct to remain in the case. For training that occurred years before any lawsuit, this ruling preserves the publishers’ ability to reach the foundational data-collection conduct rather than confining the dispute to recent activity. Given how much model training predates the litigation wave, the limitations holding has outsized practical importance.

What the court dismissed — and what it did not decide

The order was not a clean win for the publishers. The common-law unfair-competition-by-misappropriation claim was dismissed with prejudice, and most of the DMCA § 1202 claims — alleging removal of copyright-management information from training copies — were dismissed, though largely with leave to amend and with certain § 1202(b)(1) claims against OpenAI surviving in the Daily News and Center for Investigative Reporting actions. The CMI claims foundered on the same difficulty that has dogged them elsewhere: the challenge of tying CMI removal to specific, identifiable outputs.

Crucially, the court did not decide fair use. A motion to dismiss tests the sufficiency of the complaint, not the merits of an affirmative defense that turns on a fact-intensive four-factor analysis. OpenAI’s central defense — that training and the resulting model are transformative — remains fully open and will be litigated on a developed record. Nothing in the March 2025 order forecloses it; the order simply holds that the publishers have stated claims worth testing.

Open questions

The decision leaves the war to be fought. First and foremost is fair use: how the four factors apply to training on news articles, and whether the capacity for verbatim regurgitation and the substitution of AI answers for clicks weigh against transformativeness and harm the market for the originals. Second is causation and the realistic frequency of infringing outputs — how often, and under what prompting, ChatGPT actually reproduces protected text, and whether such outputs are representative or adversarially induced. Third is the coast-to-coast divergence on contributory liability, which can only be resolved by the appellate courts. Fourth is the role of the emerging text-case fair-use rulings as persuasive (not binding) authority once this case reaches summary judgment.

Implications

• For news publishers and content owners: Output-based and contributory theories are viable in the Southern District of New York, and the limitations ruling preserves access to years-old training conduct.
• For AI developers: The motion-to-dismiss stage is no longer a reliable off-ramp for output claims; the defense burden shifts to building a fair-use record and to demonstrating that infringing outputs are rare and not the product of design.
• For litigators weighing forum: A real East Coast/West Coast divergence on contributory liability now exists; venue and consolidation decisions carry doctrinal consequences.
• For licensing markets: A surviving infringement case strengthens publishers’ hand in licensing negotiations, the practical resolution toward which much of this litigation is pushing.

Frequently asked questions

Did the court rule that OpenAI infringed The New York Times’s copyrights? No. It denied most of the motion to dismiss, meaning the Times’s claims are plausible enough to proceed. Whether infringement occurred, and whether fair use excuses it, remains to be decided.

Why is this ruling considered different from the California AI cases? Several California judges dismissed output-based and contributory theories at the pleading stage. Judge Stein allowed the contributory-infringement claim to proceed, creating a divergence that may ultimately require appellate resolution.

Was fair use decided? No. Fair use is an affirmative defense unsuited to resolution on a motion to dismiss; it will be litigated on a full factual record at summary judgment or trial.

Authorities and sources

• The New York Times Company v. Microsoft Corp., No. 1:23-cv-11195 (SHS) (S.D.N.Y.): opinion and order on motions to dismiss, filed April 4, 2025 (S.D.N.Y.); docket entry (Justia).
• Coverage and analysis: NPR, “Judge allows ‘New York Times’ copyright case against OpenAI to go forward”; Axios, “NYT case against OpenAI and Microsoft can advance”.