Judge Orders OpenAI to Turn Over 20 Million ChatGPT Logs

A U.S. magistrate judge in Manhattan on December 3 directed OpenAI to hand over about 20 million de identified ChatGPT conversation logs to a coalition of news organizations led by The New York Times, as part of a long running copyright lawsuit alleging that the company’s models reproduced protected content. The order, issued in the consolidated case, represents one of the most significant judicial decisions to date on how discovery obligations interact with privacy concerns in artificial intelligence litigation.

Magistrate Judge Ona Wang concluded that the conversation logs were relevant to the plaintiffs’ claims and that the proposed de identification measures would sufficiently mitigate the privacy risks to users, Reuters reported. The news organizations say the logs will help them evaluate whether and how the model used their copyrighted articles in training or output. Plaintiffs have long argued that examples of verbatim or near verbatim reproduction are central to proving their claims.

OpenAI has appealed the order and warned that forcing production could erode user privacy and trust. The company argued that many of the chats requested are irrelevant to the litigation and that broad disclosure would expose sensitive user information even when redaction is applied. By filing an appeal, OpenAI signaled it will seek higher court review of both the scope of discovery and the adequacy of the de identification protocols imposed by the magistrate judge.

Legal experts say the dispute highlights competing priorities that courts must now navigate: the need for meaningful discovery to adjudicate novel copyright claims, and the imperative to protect user privacy and commercial confidentiality in an era when massive datasets underwrite powerful AI systems. The scale of the order, covering tens of millions of conversations, is unusual for civil litigation and raises practical questions about cost, review logistics, and the technical adequacy of de identification.

For news organizations and other plaintiffs pursuing AI related claims, the ruling could open a path to material evidence about how large language models were trained and whether specific works are reflected in the models’ outputs. For technology companies, the decision underscores a risk that the datasets used in development may be subject to court supervision and disclosure, even when firms assert proprietary or privacy based defenses.

The order may also influence how companies design data governance policies, particularly the handling and retention of user interactions that might later be sought in litigation. If courts require production of large scale interaction logs even after de identification, companies may face trade offs between data minimization practices and the utility of retained datasets for product improvement.

The case is likely to move up the appellate ladder, where judges will confront technical questions about de identification methods, the definition of relevance in AI training disputes, and the appropriate limits of discovery. Whatever the outcome, the December 3 decision has already become a touchstone in debates about transparency, accountability, and privacy in the age of generative artificial intelligence.