OpenAI accused of hiding ChatGPT logs in news lawsuit

OpenAI is facing calls for "serious sanctions" after news organizations led by The New York Times accused the company of concealing searchable ChatGPT logs that could prove alleged copyright regurgitation. The accusation came in a heavily redacted sanctions motion filed Thursday after a re-deposition in April exposed previously undisclosed log samples and searches.

What did the plaintiffs allege?

The news plaintiffs say OpenAI repeatedly lied about its ability and willingness to search ChatGPT outputs and deliberately withheld evidence. The complaint alleges that OpenAI told the court searching large anonymized samples was infeasible, burdensome, and invasive of user privacy, even though OpenAI had already run searches on de-identified datasets as part of internal research, the filing said.

Ian Crosby, lead counsel for The New York Times in the case, told Ars that OpenAI "lied to The Times, The Daily News Plaintiffs, the public, and the court," and that the company claimed searches were infeasible while having already performed them, the article states. The plaintiffs argue those misrepresentations prolonged discovery, inflated costs, and denied access to highly relevant evidence.

What evidence did OpenAI allegedly hide and why does it matter?

Monaco, identified in filings as OpenAI privacy engineer Vincent Monaco, testified that OpenAI had two large de-identified samples spanning 10 million and 78 million logs which were not disclosed to plaintiffs for more than two years. News plaintiffs allege OpenAI had searched those samples for The Times content while arguing publicly and in court that such searches were not possible without extensive work.

Instead of handing over those datasets, OpenAI provided a heavily redacted 20 million-log "sandbox" sample, a fraction of the roughly 120 million news logs plaintiffs originally requested, the filing says. The company used AI to perform some 19 billion redactions to that sample, prompting the court to find the sample "unusable." Even after some redactions were removed, plaintiffs say many fields, including domains and names, remained redacted and hampered their searches.

Plaintiffs also allege OpenAI randomly deleted parts of the 20 million sample and deleted or compressed billions of logs that should have been preserved, and that OpenAI "decided" not to comply fully with the court's Preservation Order, the filing states.

How has OpenAI responded?

An OpenAI spokesperson told Ars that the news plaintiffs' sanctions motion is a late litigation tactic to obtain more logs and invade user privacy. The spokesperson said that the Times' decision to drop some claims demonstrated the plaintiffs' weakening case, and that OpenAI will "continue defending our users' privacy and the long-established principles of fair use." The Times' spokesperson Graham James disputed that framing and said the suit was streamlined and strengthened by adding claims against Microsoft, and that core claims against Microsoft and OpenAI remain the same.

Why it matters

The log samples and any evidence of article regurgitation are central to whether training on copyrighted news content is fair use. The plaintiffs requested that the court find withheld logs showed "substantial" regurgitation and bar OpenAI from arguing otherwise. If the court finds OpenAI knowingly concealed evidence and applies "serious sanctions," the company's defensive posture on market harm and fair use could be substantially weakened, making it harder for OpenAI to show the training did not cause prohibited substitution.

What to watch

Watch for the court's ruling on the sanctions motion and whether the judge accepts plaintiffs' request to prohibit OpenAI from contesting that withheld logs showed substantial copying. Also monitor whether the court deems the 20 million-log sample admissible after the extensive redactions and whether plaintiffs gain access to the 78 million-log dataset Monaco testified existed.

Timeline of the disclosures

• April: A re-deposition of OpenAI privacy engineer Vincent Monaco revealed that OpenAI had previously searched large de-identified datasets.
• Thursday: News plaintiffs filed a heavily redacted sanctions motion alleging concealment of log samples and other misconduct.
• Discovery period details: Plaintiffs sought roughly 120 million news logs; OpenAI provided a 20 million heavily redacted sample and allegedly held 10 million and 78 million de-identified logs back from plaintiffs.