OpenAI and Microsoft privately called their AI a 'doom loop' for the web and shipped it anyway

A 92-page filing unsealed this week in The New York Times' copyright suit against OpenAI and Microsoft contains internal warnings that both companies' AI products would create a "doom loop" for the open web and amount to "the largest theft of labor in human history." A "doom loop," in the document's framing, is a self-reinforcing collapse of the open web in which chatbots stop sending readers to the publishers that trained them, those publishers lose the traffic that funds reporting, and the next generation of models has less to learn from. The plaintiffs' summary judgment brief is dated September 17, 2026; the docket is on CourtListener.

Microsoft Director of Applied Science Brent Hecht is the most quoted voice in the filing. He called ChatGPT and Copilot's harvesting of training data "the largest theft of labor in human history" and said Microsoft's legal defense made "a complete mockery of the idea of 'fair use.'" The same filing quotes Hecht and OpenAI's Head of ChatGPT describing the products as an "existential threat" to publishers. The Verge's report on the unsealing, TechCrunch's coverage, and Decrypt each run the "largest theft" line in their lead paragraphs.

That language came inside the companies. Microsoft's public posture is different. Spokesperson Alex Haurek told The Verge the comments reflect "one employee's individual perspective," are "not a legal analysis," and "do not represent the company's views." In a separate court filing, Microsoft AI's GM for Data Strategy and Ops Jordan Usdan said Hecht "holds divergent, academic, and forward-looking views" and is "not someone who speaks for Microsoft specifically as to his theoretical views on AI's potential effect on content creators." Microsoft is contesting characterization, not the existence of the documents.

The filing also includes a remark from Microsoft CEO Satya Nadella that lands harder than the executive probably intended. Asked about the relationship between chatbots and the open web, Nadella conceded that AI assistants have "basically replaced search" and removed the need to "go straight to the source" for information. Read against the Hecht quotes, that concession describes the mechanism the brief calls a "doom loop."

That pattern has a beat name, "Google Zero," and it has moved from theory to measurable fact: the point at which AI assistants stop sending readers to the news sites they learned from, and publishers see a measurable drop in referral traffic. Publishers including The New York Times now report that a meaningful share of their audience no longer arrives through a search result at all; the question is asked of a chatbot instead, and the answer is generated without a click. The court filing turns that trend from a complaint into a documented internal prediction.

The plaintiffs' legal theory is that the training itself, not just the output, is the harm. Hecht's "mockery of fair use" line is doing legal work in the brief, not rhetorical work. Fair use has historically required a showing that a secondary use does not substitute for the original or harm its market. Internal statements that the companies knew their products would substitute for, and starve, the publishers they trained on complicate that defense.

The unsealing lands at a moment when publisher licensing is moving from ad-hoc deals to industry rule. OpenAI and Microsoft have both signed content agreements with major publishers in the last two years, often as part of the same litigation. The court filing is unlikely to be the only record that matters. But it is the first time the contemporaneous internal view of senior staff is on the public page, and it changes the leverage for the next round of deals: a publisher can now point to a Microsoft director's own characterization of the practice, not just to an outside critic's.

The remedies the filing points toward are not new. They include publisher licensing deals that pay for the data the models were trained on, attribution standards that surface the source of an answer rather than absorbing it into a chatbot's voice, and traffic-restitution terms that route a measurable share of chatbot answers back to the originating site. None are silver bullets. Each is harder to argue against when the same companies' employees have already conceded, in writing and under oath, that the alternative was a "doom loop" for the web they learned from.

The next hearing in the case is scheduled for later this fall.