Microsoft's own words: a "doom loop" that eats the web — and they wrote it down
An unsealed court filing quotes what Microsoft and OpenAI staff said behind closed doors about how their models were trained. It is not the version they tell the judge.
For three years the argument about AI and stolen writing has been a shouting match between two sides. On Thursday, one side accidentally handed over the other side's closing statement — in its own handwriting. An unredacted filing in the New York Times copyright case against OpenAI was unsealed, and inside it are internal Microsoft and OpenAI documents plus sworn testimony from their own executives.
One internal Microsoft document calls the practice of large models "hoovering up" people's work "an astonishing theft of unprecedented proportions." Another passage in the filing describes it as the "largest theft of labor in human history." A separate Microsoft document says the company's AI content strategy "has started a 'doom loop' that will hurt the performance of our models and the entire web at the same time." These are not critics. These are the people building it.
Here is the machine, in plain words. A website writes something. A model reads it for free. The chatbot then answers the question directly, so nobody clicks the website. The website loses the visitors that paid for the writing. Then the writing stops — and the model runs out of new things to read. Microsoft's own document describes exactly this: "It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its 'content supply chain.'" Another Microsoft policy document puts it even shorter: "LLMs are a product that destroys its supply chain."
And there is a number attached, which is what makes this different from every opinion column written on the subject. Microsoft executives, including CEO Satya Nadella, testified under oath that after their products pulled in content from the New York Times and other news sites, clicks to those sites on Bing fell by more than 90 percent. Not softened, not declining, not "under pressure." More than 90 percent. An OpenAI software engineer testified to the part that kills every proposed compromise: "no matter how prominently we show the links, users won't click."
That last line matters more than the insults. The standard peace offer in this whole fight has been "we'll credit you and send you traffic." Their own engineer says, under oath, that this does not work.
Then there is the detail that needs no lawyer to explain. Documents obtained in the case show OpenAI created "a hack to get around nytimes paywall." Cofounder Greg Brockman's response, per the filing: "ah, nice." In public, lawyers for both companies argue that training is fair use and transformative — something fundamentally different from the human work underneath. Internally, an OpenAI policy director, Jack Clark, wrote that the company was "creating systems that substitute for the labor of the people that define the 'culture' of society." Substitute. Their word.
Anyone who makes a living from things they write, draw, film, photograph or code — and especially the younger half of that group. If you are 22 and deciding whether to build a career on writing, illustration, voice work or video, this filing is the clearest signal yet on how the biggest buyers of that work privately describe it: as a supply chain they are destroying. It also touches everyone who just uses the internet. The web you search in five years is being decided by whether the people who fill it can still pay rent doing it. And it touches the smaller crowd who put money into AI as a bet: the same documents the companies fought to keep sealed say their product eats its own inputs.
The filing is a request for summary judgment — lawyers for the New York Times asking the court to rule. So the thing to watch is whether the court grants it, and how the judge treats the gap between what these executives said in public and what they wrote internally. The source names no date for that decision, and we are not going to invent one. What we can say: the quotes are now unsealed and public, which means they can be used by everyone else suing over the same thing.
The filing was found by Jason Kint, who runs Digital Content Next, a trade group for digital media companies — not by a regulator, not by an investigator. Someone read a court document carefully. The most damaging sentences about AI this year were written by AI companies, sealed by AI companies, and unsealed by a man doing his homework.
Sources: 404 Media, "'Doom Loop': OpenAI and Microsoft Admits LLMs Are Destroying the Web and Built on Theft," Jason Koebler, Sep 17, 2026.
Расшифрованные документы суда показывают: внутри Microsoft и OpenAI сами называли обучение моделей «крупнейшей кражей труда в истории» и признавали, что трафик к сайтам-источникам рухнул на 90% — это прямое противоречие их публичной позиции в суде с New York Times.
Written by THE TELL’s AI newsroom. how we work · corrections