17 Publishers Seek Sanctions Against OpenAI Over Hidden Plagiarism-Detection Tool

A coalition of seventeen media organizations, led by The New York Times, filed an extensive motion for sanctions against OpenAI in federal court in Manhattan on Thursday. The publishers claim the company ran the very same internal searches it had previously told the court were technically impossible, all while deleting billions of ChatGPT conversation logs that the court had ordered it to preserve.

The dispute between the publishers and OpenAI has been running for more than two years and centers on whether training ChatGPT on news articles without a license constitutes copyright infringement or falls within fair use. Until now, the case has mostly revolved around how much copyrighted material actually ended up in the model's training data and how much of it the model reproduces in its responses to users.

What Project Giraffe Reveals

At the heart of the new motion is the existence of an internal OpenAI tool code-named Project Giraffe. It reportedly used a Bloom filter, a probabilistic data structure used to quickly check whether a given element belongs to a set, to detect and log instances where ChatGPT's responses reproduced passages from copyrighted texts.

According to the publishers, the tool went live shortly after the media organizations filed their original lawsuit, making it an internal record of the very same infringement that OpenAI was simultaneously telling the court was impossible to measure. In other words, the company already had data showing the scale of the problem at the moment it was assuring the court that no such data could be obtained.

Allegations From Publishers' Lawyers

Steven Lieberman, an attorney representing the New York Daily News, accused OpenAI of misleading the court for two years about its ability to search its training sets and logs for copyrighted content. The publishers also claim OpenAI made billions of ChatGPT conversations unavailable, partly by deleting them and partly by stripping them of search functionality.

OpenAI has made false representations for two years regarding its ability to search copyrighted content in training sets and logs - Steven Lieberman, attorney for the New York Daily News

OpenAI's Response

OpenAI spokesperson Drew Pusateri rejected the allegations, claiming that as The New York Times's case weakens and the publishers withdraw one claim after another, they are persisting in trying to violate the privacy of people unconnected to the dispute. The company maintains that the restrictions on sharing ChatGPT logs are meant to protect user privacy, not to hide evidence.

As the Times's case weakens and they are forced to withdraw claims, they continue to pursue the privacy violation of individuals unrelated to this case - Drew Pusateri, OpenAI spokesperson

What's At Stake

The publishers are seeking financial penalties, reimbursement of legal costs, and a formal court finding that the ChatGPT chat logs prove OpenAI misused their copyrighted content. If the court grants the motion for sanctions, it could hand the publishers a significant procedural advantage without having to prove copyright infringement on the merits.

This is the most serious escalation yet in a dispute being closely watched by the entire media and tech industry. The outcome will affect not only this particular case but also how other publishers and creators assess the risk of taking legal action against companies building large language models on content scraped from the internet.

Sources: US News (usnews.com), Spokesman-Review (spokesman.com), TechTimes (techtimes.com)