OpenAI, embroiled in a lawsuit with The New York Times and Daily News over alleged unauthorized use of their content for AI training, has encountered a significant data mishap.
According to the publishers’ legal counsel, OpenAI engineers accidentally erased search data critical to the case from one of the virtual machines used to investigate the training datasets, Techcrunch reports.
The lawsuit, filed in the U.S. District Court for the Southern District of New York, accuses OpenAI of scraping copyrighted content to train its AI models without permission. Earlier this year, OpenAI provided virtual machines for the plaintiffs to identify traces of their work in the AI’s training data. However, on November 14, OpenAI engineers deleted the publishers’ search data from one of the virtual machines. Although much of the data was recovered, the loss of folder structures and file names rendered the information unusable for determining how the plaintiffs’ content was allegedly utilized in AI model training.
In a letter to the court, attorneys for the publishers emphasized the severe impact of the deletion. They reported that over 150 hours of investigative work since November 1 must now be redone, costing significant time and resources. While the plaintiffs acknowledged that the deletion appeared unintentional, they argued the incident highlights the need for OpenAI to directly search its datasets using internal tools.
OpenAI declined to comment on the issue. The company continues to assert that using publicly available data to train AI models, including news articles, falls under fair use. However, it has recently pursued licensing agreements with several major publishers, including the Associated Press and News Corp., signaling a potential shift in approach. The terms of these agreements remain undisclosed, though reports suggest some deals are worth millions annually.
The case underscores the growing legal and ethical tensions around AI training practices, particularly as companies like OpenAI monetize models developed with publicly accessible, and sometimes copyrighted, content.
