Unsealed court filings allege OpenAI and Microsoft knew AI training practices risked harming the web

Court documents unsealed in the New York Times' lawsuit against OpenAI and Microsoft include internal communications in which employees and executives warned that the companies' AI training practices risked creating a 'doom loop' that would damage the web.

Microsoft Director of Applied Science Brent Hecht characterized the companies' scraping of data to train models as the 'largest theft of labor in human history' and said it made a 'complete mockery of the idea of fair use.' Microsoft disputed Hecht's characterization, stating his views were not the company's official position.

An internal Microsoft document cited in the filing described the company's AI content strategy as having started a 'doom loop' that would simultaneously harm the performance of its models and the entire web by threatening the economic foundations of content suppliers.

Satya Nadella acknowledged internally that chatbots had replaced search for many users, reducing the need to visit original sources. OpenAI representatives admitted they were 'unaware' of efforts to detect or remove paywalled content from training data, despite Nadella stating that paywalled content should be licensed.

OpenAI employees acknowledged that models like GPT-4 were 'insanely good at regurgitation,' having memorized large amounts of data, and provided examples of ChatGPT outputting verbatim passages from news articles in response to queries.

Microsoft acknowledged in filings that 'almost no one intended for their content to be used' in the way AI companies were using it, and that creators were not compensated for such use.

OpenAI Policy Director Jack Clark warned that the technology was 'creating systems that substitute for the labor of the people that define the culture of society,' with internal documents describing ChatGPT as 'the modern newsstand.'

OpenAI's own analysis attributed a drop in referral traffic to publishers, including the New York Times, to AI-generated summaries, speculating that search referrals may have declined by as much as 60%.