OpenAI faked inability to search training data, hid billions of logs, NYT says - Ars Technica

Skip to content

AI

Biz & IT

Cars

Culture

Gaming

Health

Policy

Science

Security

Space

Tech

Forum

Subscribe

Story text

Size

Small<br>Standard<br>Large

Width

Standard<br>Wide

Links

Standard<br>Orange

* Subscribers only

Learn more

Pin to story

Theme

Search

Sign In

Sign in dialog...

Text<br>settings

Story text

Size

Small<br>Standard<br>Large

Width

Standard<br>Wide

Links

Standard<br>Orange

* Subscribers only

Learn more

Minimize to nav

OpenAI is facing calls for “serious sanctions” after fighting to keep news organizations from snooping through millions of logs to find evidence of users skirting their paywalls by prompting ChatGPT to regurgitate their articles.

This evidence is considered among the most important to both sides, potentially either dooming OpenAI as an infringer or exonerating its chatbot technology as a transformative fair use of news sites’ content.

In a sanctions motion Thursday, news organizations suing OpenAI—led by The New York Times—accused the AI firm of repeatedly lying for years to conceal evidence of infringement that could hobble OpenAI’s defense.

These alleged lies were exposed when the court compelled an “ill-prepared witness,” OpenAI privacy engineer Vincent Monaco, to be re-deposed. During the subsequent April deposition, he inadvertently revealed that OpenAI misled the court for two years about the cost and burdens of searching ChatGPT logs, NYT’s filing said.

Among the most shocking revelations, OpenAI allegedly pretended from the earliest stages of the case that it did not have the technical ability to search large anonymized samples of ChatGPT logs when it had actually already conducted such searches prior to the start of litigation, NYT alleged.

Sanctions are warranted because “OpenAI’s concealment of this fact withheld highly relevant evidence, prolonged discovery, inflated expenses, and burdened the Court,” news plaintiffs alleged.

Asked for comment, an OpenAI spokesperson suggested that NYT’s sanctions motion was a late litigation effort to access more logs and infringe more users’ privacy. The spokesperson claimed that when the NYT recently dropped some claims in the lawsuit, it was a sign that news plaintiffs’ case was crumbling, not OpenAI’s defense.

“As the Times’ case weakens and they’ve been forced to drop claims against us, they’re persisting with their efforts to invade the privacy of people who have nothing to do with this case, including by making these blatantly false allegations,” OpenAI’s spokesperson said. “We’ll continue defending our users’ privacy and the long-established principles of fair use.”

However, last month, NYT spokesperson Graham James disputed to Ars that news plaintiffs’ case was weakened by dropping claims. He suggested instead the suit was streamlined and strengthened by adding claims against Microsoft. “Our core claims remain the same from the day we filed this lawsuit—that Microsoft and OpenAI stole millions of The Times’s copyrighted works to compete with our products and illegally enrich themselves,” James said.

OpenAI allegedly hid 80M log sample

Although the sanctions motion is heavily redacted, it’s alleged that Monaco testified that OpenAI had two large samples—spanning 10 million and 78 million logs—which had already been de-identified and could have been made available to news plaintiffs early on to maximize the discovery period.

“Not once did OpenAI disclose the existence” of those samples over two years, news plaintiffs alleged.

Even more frustrating to plaintiffs, OpenAI had already searched those samples for NYT content as part of its research into “creating a filter that could be used to block the regurgitation of copyrighted content,” the court filing said.

“OpenAI was willing and able to search its output logs—when it benefitted OpenAI,” NYT alleged, accusing the ChatGPT maker of “making the discovery process as burdensome as possible.”

Court says OpenAI sample is “unusable”

In a statement to Ars, NYT’s lead counsel, Ian Crosby, suggested that OpenAI obstructed access to logs and distorted evidence to shield its fair use claims.

“For over two years, OpenAI lied to The Times, The Daily News Plaintiffs, the public, and the court,” Crosby said. “It claimed searching ChatGPT outputs for copies of The Times’ and the Daily News Plaintiffs’ content was infeasible, burdensome, and invasive of users’ privacy—while at the same time concealing that it had already done such searches. If OpenAI genuinely believed that copying our clients’ journalism was fair and legal, it wouldn’t have hid the truth about having done it.”

Instead of being transparent about the existing samples, OpenAI forced news plaintiffs to spend eight months searching in a “sandbox,” where they could only access a heavily redacted sample of 20 million logs. That sample was much smaller than the 120 million news logs plaintiffs originally...