NYT, Daily News and others say the ChatGPT maker 'chose obstruction' over releasing the training data and logs that could show how its AI learned from journalism.
A coalition of news publishers filed court papers Thursday asking a Manhattan federal judge to sanction OpenAI for allegedly hiding the datasets and ChatGPT logs that could show how its AI models were trained on copyrighted journalism. The New York Times, the New York Daily News and seven sister MediaNews Group papers, the Chicago Tribune, Ziff Davis, and the nonprofit Center for Investigative Reporting are asking the court to punish the company for "discovery misconduct": court discipline, not a fine, that could include evidentiary preclusion or an "adverse inference" instruction telling a jury to assume the hidden evidence would have hurt OpenAI (AP News).
The filings, made in the consolidated case In re OpenAI, Inc. Copyright Infringement Litigation (MDL No. 1:25-md-03143-SHS-OTW) in the Southern District of New York, mark the first time the publishers' two-year discovery fight has been put in front of a federal judge as a standalone sanctions motion. The underlying copyright suit, filed in late 2023 by the Times against OpenAI and Microsoft, tests whether training AI chatbots on digitized news articles is protected as "fair use," the legal doctrine that lets limited copying occur without permission for purposes like commentary, criticism, or news reporting. Thursday's motion asks the judge to assume, in effect, that the materials OpenAI allegedly hid would have confirmed the publishers' case.
In plain language, the publishers accuse OpenAI of misrepresenting for two years its ability to search its own training data and ChatGPT logs for copyrighted content. The motion cites a recent deposition of an OpenAI employee that the publishers say contradicts the company's earlier claims about that search-and-retrieval capability, according to Bloomberg Law's reporting on the filing. Steven Lieberman, an attorney for the New York Daily News, said OpenAI has been "making misrepresentations" and that the motion asks the court to "punish OpenAI for hiding and destroying evidence showing how ChatGPT was trained on stolen journalism," language the Washington Post and Variety both attribute to plaintiffs' counsel. "Stolen journalism" is Lieberman's characterization, not a judicial finding; OpenAI has denied the underlying copyright allegations in earlier filings and the question of whether training on news content is fair use remains unresolved.
What the publishers are asking for, and what they are not, matters. An adverse-inference instruction would let a jury at trial assume that the withheld datasets and logs would have shown the AI was trained on the publishers' copyrighted articles, a powerful tool in a case where the central question is what went into the model. Evidentiary preclusion would block OpenAI from offering its own evidence on the affected topics. Both remedies are rare against a defendant of OpenAI's standing, and either would alter the trial's center of gravity without the judge deciding the underlying fair-use question. The publishers are not asking for a copyright ruling. They are asking for a procedural penalty that could force one.
The consolidated suit also embeds the publishers' broader grievance about AI's effect on the news business. AI chatbots and Google's AI-generated search summaries have begun siphoning web traffic from publishers, cutting off the ad dollars tied to clicks and raising the question of whether the chatbots compete unfairly as information sources without doing the journalism that gathering the news requires. That second-order effect, who pays for reporting once readers stop visiting the sites that produce it, is the part of the case that has drawn attention from editors and trade groups even as the legal fight has stayed inside the courtroom.
OpenAI did not immediately respond to a request for comment Thursday, the wire pickup confirmed by U.S. News & World Report noted. The judge's next move is to consider the motion on the papers. If she treats the allegations as serious enough to warrant an evidentiary hearing, the publishers would get a chance to develop the deposition record in open court. If she denies it, the case returns to the slower discovery track that has defined the past two years. Either way, the question of what OpenAI's training data actually contains, the question the motion says OpenAI has been fighting to keep the court from asking, now sits on a federal docket with a date.