Court documents in The New York Times’ lawsuit against OpenAI and Microsoft show internal Microsoft discussion over whether using news articles to train AI models amounted to a sweeping misuse of creative labor, while both companies continue to argue the training at issue is protected by fair use.

The newly surfaced material includes a 2023 memo from Brent Hecht, Microsoft’s Director of Applied Science, describing large AI models as “products that destroy their own supply chains” and warning that many people viewed the systems as an unprecedented appropriation of their work. Court filings also cite his characterization of copying millions of news articles without permission as “the largest theft of labor in human history.”

Microsoft said those memos reflected Hecht’s effort to provide “different and asymmetric perspectives” and did not represent the company’s official position. The distinction matters because the case centers not on a public employee campaign, but on internal communications that plaintiffs are using to argue the companies understood the legal and commercial sensitivity of training on publisher content.

According to the filings, Microsoft CEO Satya Nadella testified that content behind a paywall should be used only with authorization from rights holders. He also said that, if he had known beforehand that OpenAI was using paid content for training, he would have used Microsoft’s rights to require the model to be retrained. Separate court materials also cite an OpenAI employee referring to a “hacker method” for bypassing The New York Times paywall in a message to President Greg Brockman, who replied, “Nice.”

OpenAI and Microsoft both maintain that the relevant model training qualifies as fair use. The dispute began when The New York Times sued at the end of 2023, and the original report says 11 publishers have since joined the case. With limited verified background beyond the court disclosures, several underlying details remain unresolved, including the full scope of the datasets at issue and how the court will weigh authorization, licensing boundaries, and market harm.

Why It Matters

This case goes beyond reputational risk for two major AI companies. It cuts into one of the core legal questions facing generative AI: whether training on commercially valuable text can be treated as transformative use when the source material is expressive, licensed, and in some cases paywalled. News publishing is a particularly sensitive test because there is a real market for licensing this content and a clear concern that AI outputs could compete with the original product.

The disclosures also sharpen the compliance issue for AI infrastructure more broadly. If courts place greater weight on consent, source restrictions, or the reuse of data acquired under different terms, developers and their partners may face tighter controls over how training corpora are assembled, shared, and refreshed.