OpenAI says its engineers mistakenly deleted data that could have supported The New York Times and Daily News’ copyright infringement claims.
The data, which was stored on a virtual machine provided by OpenAI, contained search results of the publishers’ content within the company’s AI training datasets. This deletion has marred proof for the case.
The incident occurred on November 14, when OpenAI engineers accidentally erased the search data, which had been collected over 150 hours of work since November 1.
Although OpenAI attempted to recover the lost data, it was unable to restore the folder structure and file names, making the recovered data unusable for identifying where the alleged copyrighted content was incorporated into the training models.
As a result, The New York Times and Daily News’ legal teams now face the task of recreating their work from scratch, requiring further hours of labour and additional computer processing time.
The publishers’ attorneys have stressed that, while they do not believe the deletion was intentional, the event reveals OpenAI’s position as the entity most capable of conducting the necessary searches using its own systems.
The publishers argue that OpenAI’s use of their content without permission to train AI models like GPT-4 constitutes a violation of copyright law.
OpenAI, however, defends its actions, asserting that training models on publicly available data is covered under fair use, and that it is not required to pay for or license content used in this manner.
Nevertheless, OpenAI has recently entered into licensing agreements with several publishers, including the Associated Press and the Financial Times, though the terms of these deals remain undisclosed.
Reports disclose that some of these partnerships could be worth millions of dollars annually, with Dotdash Meredith, one of the partners, reportedly receiving at least $16 million a year.
OpenAI has yet to comment on the specific circumstances of the data deletion or confirm whether its AI systems were trained on the publishers’ content without consent.
