OpenAI Risks Billions as Court Weighs Privilege in Copyright Row

OpenAI Inc.'s failure to shield internal communications about pirated books from copyright plaintiffs threatens to expose the company to billions in damages and potentially debilitating sanctions.

Authors and publishers suing the artificial intelligence giant have secured access to some Slack messages and emails discussing OpenAI’s deletion of a dataset containing pirated books and are seeking additional attorney communications about the decision. If they succeed, the communications could demonstrate willful infringement, triggering enhanced damages of as much as $150,000 per work.

The stakes reach beyond a hefty damages award. If the court finds that OpenAI destroyed evidence anticipating litigation, sanctions could follow. The judge could issue monetary penalties, limit OpenAI’s defenses, or even issue a default judgment in plaintiffs’ favor.

The case will test the Manhattan court’s willingness to hand over hallowed attorney communications to a wave of copyright owners suing key players in the AI industry over how they train large language models. Authors and publishers sought similar communications from Anthropic PBC in their AI copyright class action, before that case settled in August for a record $1.5 billion.

“Finding out what attorneys said or what clients said to attorneys and back and forth probably gives us a lot of evidence regarding state of mind,” said David Schultz, a professor at Hamline University. The disclosure would be an “enormous” blow to OpenAI’s defense, he said.

The US District Court for the Southern District of New York last week ordered OpenAI to turn over most employee communications about the data deletion that the AI company argued were protected by attorney-client privilege. OpenAI may appeal the decision. A separate bid for OpenAI’s correspondence with in-house and outside attorneys remains pending.

Mounting Pressure

AI companies’ state of mind is front and center in dozens of copyright lawsuits, as a finding of willful infringement can increase damages to $150,000 per work, from $750. Tens of millions of books and news articles are at issue in the dozen copyright cases against OpenAI that were consolidated in April in the Manhattan court for pretrial activity.

Anthropic said its was pushed to settle with book authors in August due to “inordinate pressure” to avoid a trial that could have put the company on the hook for as much as $1 trillion in damages for downloading of millions of pirated books from Library Genesis, the pirate library OpenAI also downloaded, before deleting the data.

The OpenAI plaintiffs are similarly positioning themselves for a settlement, even though the OpenAI case isn’t as far along as the Anthropic lawsuit when the parties struck a deal, Saint Louis University law professor Yvette Joy Liebesman said. OpenAI’s internal communications could serve as a “bigger bargaining chip” if they reveal willfulness or spoliation, which could set a tangible estimate for potential damages.

“There may be a smoking gun, we don’t know,” she said. But the plaintiffs’ attorneys are “going to get as much information as possible to get as much money for plaintiffs as possible.”

Novel Strategy

Plaintiffs are pursuing a rare combination of arguments to pierce attorney-client privilege, a protection courts rarely breach.

They first contend OpenAI waived its privilege for attorney correspondence by selectively disclosing one reason for deleting the LibGen data, saying it was due to “non-use.”

“You can’t pick and choose what you’re going to introduce into the record,” said Schultz, noting that the US Court of Appeals for the Second Circuit, where the Manhattan court sits, doesn’t recognize selective waiver.

OpenAI has walked back its statement, claiming the “due to” phrasing was never intended to disclose privileged reasons for the removal and that its revision to remove to the wording wasn’t a “flip flop,” as plaintiffs described it.

But the damage may be done. The fact OpenAI put forth a reason is “pretty damning” even if the company now claims the wording was accidental, said New York Law School professor Rebecca Roiphe. “The plaintiffs’ point of ‘no take backs’ is right.”

She noted OpenAI offered “the most innocuous reason” for deletion. “If the attorney conversations confirm that, then why are they fighting it so hard?”

‘Nuclear Option’

The plaintiffs are also invoking the crime-fraud exception to privilege, arguing OpenAI’s communications may reveal criminal copyright infringement or intentional destruction of evidence.

If an attorney advised OpenAI to destroy the data rather than to simply stop using it, that lawyer could be considered a participant in the infringement, Liebesman said.

That argument is the “nuclear option” because courts rarely find privilege waived, to prevent parties from constantly seeking disclosure, said Snell & Wilmer LLP Partner Nathan Mammen. He’s never seen the crime-fraud exception asserted in a copyright case. It’s more common in patent cases when accused infringers claim the invention is invalid because the patent owner committed fraud on the US Patent and Trademark Office.

It’d be “remarkable” if OpenAI’s counsel, Latham & Watkins LLP and Morrison & Foerster LLP, were recommending destruction when that’s “clearly spoliation” of evidence, said Roiphe.

The firms didn’t immediately respond to requests for comment. OpenAI didn’t comment on the privilege issue.

Potential Punishment

Schultz, Roiphe, and Mammen all predicted Judge Ona T. Wang would order in camera review, where the court examines the documents in private before issuing a decision.

Her recent order mostly rejecting OpenAI’s privilege contentions over related documents could make her gravitate towards plaintiffs, Mammen said, since she’s already found the company improperly withheld materials.

“If a party is acting aggressive, that just predisposes the judge,” Roiphe added.

If the records reveal spoliation, the court could issue an adverse inference instruction against OpenAI, telling the juries in any later trials to assume the evidence would have been unfavorable to the company and destroying it gave it an unfair advantage. The court could even tell the jury to assume OpenAI’s conduct is willful.

“Juries hear that and it’s a very powerful stick,” Mammen said. “It doesn’t guarantee the outcome, but it’s a heavy thumb on the scales.”

The case is In Re: OpenAI, Inc. Copyright Infringement Litigation, S.D.N.Y., No. 25-md-03143.