Internal Microsoft Documents: AI Training Called “the Largest Theft of Labor in History”

The court has lifted part of the seal on materials from the three-year-old lawsuit that The New York Times filed against OpenAI and Microsoft, accusing the companies of copyright infringement in training their generative models. The new documents show that, in internal correspondence, employees at both companies assessed what was happening far more harshly than they did in public.

Most of the quotes come from the newspaper's own motion rather than from the attached materials, which remain under seal — meaning the statements are cited without their original context. Even so, the wording is telling. In a January 2023 memo, Microsoft's director of applied science Brent Hecht called the data collection “an astonishing theft of unprecedented scale” and “the largest theft of labor in human history.”

The argument against “fair use”

The legal dispute hinges on the fair use doctrine, and so far the courts have more often leaned toward the AI companies; earlier this month the Trump administration, too, came out in OpenAI's defense. Yet one of the doctrine's key criteria is that the use must not supplant the original or harm its market. And this is precisely what the newly surfaced admissions contradict.

By Microsoft's own data, the “answer engine” Copilot cut click-throughs to the NYT website by as much as 93% compared with an ordinary Bing search. A January 2024 presentation by Hecht describes this as a “doom loop” that “will harm both our models and the entire web all at once.” The document states plainly: it is unusual for an end product to threaten the economic foundation of its own suppliers, yet that is exactly the situation the company created for its “content supply chain.”

ChatGPT chief Nick Turley wrote that publishers face an “existential threat,” and that products like the chatbot “substantially substitute” for the original and will do so ever more. OpenAI president Greg Brockman called the models “excellent at news.” Satya Nadella acknowledged under oath that talking to a chatbot replaces a visit to the source website; he also said that anything behind a paywall should be licensed, and that, had he known paid content was being collected, he would have demanded that OpenAI retrain the models.

Scale and methods

The numbers are striking: OpenAI's intermediate datasets alone contain more than 91,692 copies of material from the NYT, the Daily News and the Center for Investigative Reporting, while one set based on Common Crawl included more than 2 million documents from nytimes.com. Data from the joint Project Mango effort gathered at least 160,903 unique works by news publishers.

The documents also describe ways of circumventing protections. When OpenAI researcher Nick Ryder told Brockman about a “hack to get around the NYT paywall,” he replied, “oh nice.” Copyright notices were deliberately scrubbed from the pre-training data — so that the model would not surface them to users.

OpenAI and Microsoft did not respond to a request for comment.

Source: TechCrunch