The newspaper and other publishers allege that OpenAI hid evidence about how ChatGPT used copyrighted news content.

The New York Times and several other publishers asked a federal judge to sanction OpenAI, accusing the company of misleading the court about its ability to search training data and ChatGPT records while continuing to delete or compress billions of user conversations during the litigation.

The motion, filed in federal court in New York, is the latest development in the publishers’ copyright lawsuit against OpenAI and Microsoft. The plaintiffs argue that OpenAI obstructed discovery, failed to preserve evidence, and made inaccurate statements about what information it could retrieve from its systems.

Two questions shaped the case from the beginning: how much news content OpenAI used to train its models, and how often ChatGPT reproduces or relies on copyrighted material in its responses. The motion argues that records of ChatGPT conversations and internal analyses could help answer those questions.

The publishers allege that OpenAI repeatedly told both the court and the plaintiffs that it lacked practical tools to search its training datasets and ChatGPT output logs for copyrighted material. According to the filing, a second deposition of OpenAI executive Vincent Monaco later revealed that the company had already built tools capable of searching those datasets and conducted related analyses before the lawsuit was filed.

The motion also focuses on ChatGPT “output logs,” or records of user conversations, which the publishers say OpenAI continued to delete or compress even after the court ordered the company to preserve them. 

A separate dispute concerns a sample of 20 million ChatGPT conversations that the court ordered OpenAI to produce as part of the discovery process. The publishers contend that OpenAI deleted part of that sample, replaced other conversations, and applied billions of redactions, making the material difficult to use.

Further, the filing references an internal OpenAI effort, Project Giraffe. The publishers say the project included tests and datasets designed to measure when ChatGPT reproduced copyrighted material, contradicting OpenAI’s earlier claims that it lacked the tools to conduct such searches. The motion argues that OpenAI failed to disclose those capabilities during discovery.

The publishers are asking the court to:

• Bar OpenAI from using the disputed sample of 20 million ChatGPT conversations to support its defense that ChatGPT does not systematically reproduce copyrighted material;
• Assume that ChatGPT records either contained, or would have shown if properly preserved and produced, substantial amounts of copyrighted news content;
• Prevent OpenAI from arguing the opposite at trial;
• Instruct the jury that those conclusions should be treated as established and that OpenAI deleted billions of chat logs; and
• Require OpenAI to pay attorneys’ fees, expert costs, and other expenses related to the discovery fight.

The court has not ruled on the motion, and the allegations in the filing have not been proven.