Documents Reveal Microsoft Called AI Scraping Mass Theft

Unsealed documents from the copyright infringement lawsuit the New York Times has been pursuing against OpenAI and Microsoft since 2023 show executives from both companies admitting, in internal communications, that training their models on protected journalistic content was, in their own words, theft. The material was revealed by TechCrunch and is the latest escalation in a dispute that has already lasted three years.

A caveat is in order: much of the material comes from the brief filed by the Times itself, not from the original exhibits, which remain under seal. The quotes below appear without the full context in which they were said, but their substance is already enough to shift the tone of the debate over training AI on web data.

What the documents reveal

In an internal memo from January 2023, Brent Hecht, Microsoft's director of Applied Science, described the company's practice as "a shocking theft of unprecedented proportions" and "the largest theft of labor in human history." A year later, in January 2024, the same executive wrote an internal presentation warning that the decline in news site traffic caused by Microsoft's Copilot was creating a "doom loop" that would "harm the performance of our models and of the entire web at the same time."

On OpenAI's side, Nick Turley, head of ChatGPT, wrote that products like the chatbot pose an "existential threat" to publishers, because they are "broadly substitutive" and "will become increasingly substitutive as they improve." OpenAI president Greg Brockman went so far as to describe the models as "excellent at news," a phrase that, in the context of the lawsuit, is being used against the company: if a model is good enough to replace the original source, it no longer fits the fair use defense.

The scale of the copying

The documents contain previously unseen numbers on the volume of content used. OpenAI's mid-training datasets contain more than 91,692 copies of articles published by the NYT, the Daily News and the Center for Investigative Reporting. A dataset derived from Common Crawl gathered more than 2 million documents from the nytimes.com domain alone. Meanwhile, data collected by the internal initiative called Project Mango, a joint effort between Microsoft and OpenAI, amounts to 160,903 unique works from the outlets suing the companies.

The filing also details how this content was obtained: scraping Bing's own index, deliberately bypassing paywalls, and intentionally stripping copyright notices before the text reached the training model, to keep the model from "spitting out" those notices in its responses. In a message exchange cited in the lawsuit, researcher Nick Ryder told Brockman about a "hack to get around the nytimes paywall"; Brockman's reply was simply "ah nice."

Nadella under oath

Microsoft CEO Satya Nadella testified this year and stated that "anything behind a paywall should be licensed by whoever wants to use it [...] for grounding or training." He said that if he had known OpenAI had scraped and trained on paid content without a license, he would have invoked Microsoft's contractual right to demand the models be retrained. Under oath, Nadella also agreed that talking to chatbots has "replaced [...] getting the information straight from the site on the AI platform, instead of having to go to the original source."

This kind of admission works against the fair use defense, the doctrine in American law that allows the use of protected work without authorization in certain cases, such as parody, criticism or reporting. One of the central criteria of the fair use test is precisely whether the new use substitutes the market for the original work. Microsoft's own internal data shows this happening: the click-through rate to the NYT domain fell by up to 93% when searches went through Copilot, compared to traditional Bing search.

The dispute comes at a delicate moment for big AI tech companies: in early September, the Trump administration filed a brief defending OpenAI's unlicensed use of protected material to train its LLMs. American judges have, so far, been largely favorable to the argument that training constitutes fair use, but the new documents open up a specific crack: evidence that the companies themselves knew about the economic harm they were causing and pressed ahead anyway.

What changes for those building with AI in Brazil

For those developing products using third-party LLMs, whether via the OpenAI API, Azure OpenAI Service or Copilot, the lawsuit changes nothing in the short term in terms of access to the model. But it redraws the risk map for those building on top of these platforms.

The most concrete point is the origin of the training data. If an American court decides that a large part of a model's corpus came from deliberate, concealed scraping, that could trigger demands for retraining, licensing changes and even temporary unavailability of certain models or versions, something that has already happened on a smaller scale in other copyright disputes involving generative AI. Product teams that rely on a specific model as the core piece of a feature (semantic search, support assistant, content generation) would do well not to treat this as zero risk.

The second point is about building with RAG (retrieval-augmented generation) on top of paid or protected content. The logic used against Microsoft, that grounding on paywalled content requires licensing, is the same logic that applies to any Brazilian product that scrapes news portals, subscriber blogs or proprietary databases to feed its own chatbot or search engine. The 93% drop in NYT click-through via Copilot is a data point any search or growth squad should study: if your AI product answers the question without sending traffic to the source, you're in the same legal territory Microsoft is now trying to defend.

Brazil does not follow the American doctrine of fair use, but it has its own copyright limitation mechanisms under Law 9.610/98, Brazil's copyright statute, and regulatory discussions on AI (such as bill PL 2338/2023, a proposed AI regulation framework) still haven't settled how to treat data mining for training. Legal teams at Brazilian companies that train or fine-tune their own models on data scraped from the web would do well to document dataset provenance now, before American case law starts influencing local decisions, something that usually happens with a one- to two-year lag.

Finally, it's worth noting the statement from attorney Steven Lieberman, who represents the New York Daily News in the lawsuit: "the evidence revealed here for the first time shows that OpenAI and Microsoft knew that what they were doing was wrong." Neither OpenAI nor Microsoft responded to TechCrunch's request for comment. The lawsuit is ongoing, and more documents may be unsealed in future phases.