News publishers seek sanctions against OpenAI in copyright case

News organizations led by The New York Times asked a federal court Thursday to impose severe sanctions on OpenAI, accusing the company of concealing searchable ChatGPT output logs in a copyright case that could shape how AI companies use news content. The publishers say the logs are central evidence for whether ChatGPT reproduced paywalled articles and harmed the market for their work, issues tied directly to OpenAI’s fair use defense.

In a sanctions motion, the news plaintiffs alleged that OpenAI misled them and the court for roughly two years about the feasibility and burden of searching large samples of ChatGPT logs. The filing says an April deposition of OpenAI privacy engineer Vincent Monaco revealed that OpenAI had already searched large anonymized datasets before the litigation, despite representing that such work was technically difficult, costly and intrusive to user privacy.

The plaintiffs say OpenAI had two de-identified log samples, one with 10 million logs and another with 78 million logs, that could have been made available earlier in discovery. According to the motion, OpenAI had also searched those datasets for Times content as part of work on a possible filter to block reproduction of copyrighted material.

OpenAI disputed the allegations. A company spokesperson said the motion was part of an effort by the Times to gain access to more user data, and called the allegations “blatantly false.” The spokesperson said OpenAI would continue to defend user privacy and fair use. The company also pointed to the Times dropping some claims as evidence that the publishers’ case had weakened.

The Times has rejected that framing. Graham James, a Times spokesperson, previously told Ars Technica that the case had been streamlined and strengthened by adding claims against Microsoft. He said the core allegation remained that Microsoft and OpenAI used millions of Times works without authorization to compete with the publisher’s products and enrich themselves.

The fight centers on discovery. The publishers originally sought a 120 million-log sample, but OpenAI produced a smaller 20 million-log sandbox. The plaintiffs say that sample was damaged by 19 billion AI-generated redactions, and that the court later described it as “unusable.” OpenAI later removed some redactions, according to the filing, but the publishers say fields including news domains and names remained obscured, limiting their searches.

The motion also alleges that OpenAI deleted parts of the 20 million-log sample and deleted or compressed billions of logs that should have been preserved under a court order requiring retention of chats. The publishers say Monaco testified that OpenAI considered compliance with the preservation order, decided it would be hard and took no steps to do so.

The publishers are asking the court to bar OpenAI from relying on the 20 million-log sample, find that withheld logs contained substantial reproduction of the publishers’ copyrighted material, prevent OpenAI from arguing otherwise and instruct the jury that OpenAI deleted billions of logs.

Ian Crosby, lead counsel for the Times, said in a statement to Ars Technica that OpenAI had claimed searches were infeasible while hiding that it had already performed them. OpenAI has not disclosed how much of the requested data, if any, it believes can be searched without compromising user privacy. The court’s decision on sanctions could materially affect OpenAI’s ability to argue that its systems did not substitute for the publishers’ own products.

This story draws on original reporting from Ars Technica.