OpenAI Accused of Withholding Evidence in ChatGPT Copyright Lawsuit, NYT Claims

The New York Times and the New York Daily News have accused OpenAI of concealing evidence about its ability to search ChatGPT conversation logs and AI training datasets for copyrighted material. The allegation marks the latest development in the publishers’ two-year copyright lawsuit, which alleges that OpenAI trained its generative AI models on their journalism without authorization and reproduced portions of that content in ChatGPT responses.

Throughout the legal proceedings, OpenAI has maintained that it was unable to search its training corpus for specific copyrighted works. The company also argued that searching or producing its vast collection of ChatGPT conversations would be technically difficult and raise privacy concerns, as user conversations would need to be retrieved, processed, and de-identified before being shared.

The publishers requested access to this information to determine whether their copyrighted articles were included in OpenAI’s training data and to assess how frequently ChatGPT generated responses that used or reproduced their journalism.

According to court filings, an April deposition of OpenAI data privacy engineer Vinnie Monaco revealed that the company had already conducted internal searches and evaluations of its training datasets to identify copyrighted journalistic material.

The deposition also allegedly disclosed that before The New York Times filed its lawsuit, OpenAI had compiled a database of approximately 78 million de-identified ChatGPT conversations. According to the plaintiffs, the company used this dataset internally to evaluate how often its AI models reproduced copyrighted works. They further claim that shortly after the lawsuit began, OpenAI introduced a “loom” filter into an internal system known as “Project Giraffe” to detect and record instances in which ChatGPT generated responses that closely reproduced source material.

The publishers argue these revelations are significant because they had originally requested a sample of 120 million ChatGPT conversation logs during discovery. OpenAI negotiated that request down to a sample of 20 million conversations, which was eventually submitted to the court last December. However, the plaintiffs allege that the sample contained extensive redactions, rendering it largely unusable. They also claim OpenAI deleted billions of ChatGPT outputs after the lawsuit was filed, despite a court preservation order, and replaced millions of requested records with different conversation logs.

According to the plaintiffs, these actions unnecessarily complicated efforts to obtain information that OpenAI had already collected and analyzed internally.

Ian B. Crosby, lead counsel representing the publishers, said that if OpenAI genuinely believed its use of the publishers’ journalism was lawful and protected under fair use, it would not have concealed evidence relating to those activities.

The New York Times and the New York Daily News are now asking the court to impose sanctions against OpenAI. Their requests include preventing the company from relying on the submitted 20 million chat log sample as evidence, accepting that ChatGPT conversation records would have demonstrated substantial reproduction of the publishers’ content, preventing OpenAI from arguing that its submitted data does not show significant regurgitation, and requiring the company to pay legal costs associated with obtaining the disputed evidence.

OpenAI has rejected the allegations. Company spokesperson Drew Pusateri said the publishers were attempting to gain access to private user conversations as their legal claims weaken.

In a statement, Pusateri said The New York Times was making false allegations while attempting to obtain access to conversations involving users who have no connection to the lawsuit. He added that OpenAI will continue to defend user privacy and its position that its AI training practices fall within long-established principles of fair use.