'Doom Loop': OpenAI and Microsoft Admits LLMs Are Destroying the Web and Built on Theft

In progressJason Koebler (404 Media)article2026-09-17read planted ai llmsbusiness

Synopsis — AI-drafted from Dan's notes

A 404 Media piece by Jason Koebler on the New York Times’ summary-judgment brief in its copyright suit against OpenAI and Microsoft, unsealed on 17 September with depositions and internal documents that had been redacted. His argument is that the companies privately describe LLM training the way their critics do, and that this undercuts the fair-use and transformation defence they make in court.

The material he pulls out runs in two directions. One is theft: a Microsoft document predicting that millions of people will see their work taken to train models, and calling the scraping a theft on a scale never seen before. The other is the “doom loop”: a Microsoft document saying its AI content strategy hurts its own models and the web at once, because answer engines starve the publishers whose work the models learn from, which makes the LLM a product that destroys its own supply chain. Around those sit the smaller items: Satya Nadella’s testimony that clicks to news sites fell by more than 90 percent on Bing, an OpenAI workaround for the Times paywall that Greg Brockman greeted with approval, an engineer testifying that users won’t click links however prominent they are, Jack Clark writing about systems that substitute for the labour of the people who make culture, and OpenAI calling itself an existential threat to publishers.

The quotes hold up in other reporting on the same filing (TechCrunch, the Times itself), with three corrections. The theft lines and the doom-loop deck are the work of one person, whom 404 does not name: Brent Hecht, Microsoft’s director of Applied Science, in a memo of January 2023 and a presentation of January 2024. Microsoft says his comments reflect one employee’s view and not the company’s, so the headline’s “Microsoft admits” is stronger than the documents. TechCrunch sources the click figure to Hecht’s presentation, not to Nadella: Copilot’s answer engine cut click-through to the Times’ domain by as much as 93% against traditional Bing. And the Clark line is old. He left OpenAI at the end of 2020 to co-found Anthropic, and 404 gives it no date.

The brief is the plaintiffs’ advocacy, with quotes chosen by the Times’ lawyers to win a motion. Judge Sidney Stein is weighing the summary-judgment motions, and the suit now has eleven other publishers as plaintiffs. Neither company commented to TechCrunch.