The Doom Loop Document: AI's Own Executives Just Confessed to Cannibalizing the Web

The moment the mask came off

There is a particular kind of documentary evidence that makes a legal case — and an industry — simply irreversibly damaged. It doesn’t come from anonymous whistleblowers or leaked Slack messages. It comes from the defendants themselves, under oath, in documents they originally fought to keep sealed.

That is exactly what happened this week in the New York Times vs OpenAI copyright lawsuit, when an unredacted court filing was unsealed that laid out a series of internal admissions from Microsoft and OpenAI executives. What emerged was not a nuanced discussion of fair use or technological disruption. What emerged was a company and its biggest corporate partner effectively confessing that their generative AI products were built on theft and were actively cannibalizing the web publishers that supplied the data.

One internal Microsoft document, cited directly in the filing, described what the company itself called a “doom loop” — a feedback cycle in which AI content products steal from human creators, destroy the websites those creators depend on for revenue, and then threaten to make those creators irrelevant altogether. The language was unmistakable: “Our AI content strategy has started a doom loop that will hurt the performance of our models and the entire web at the same time.”

Another line, equally stark: “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.”

The admission that changes everything

What makes these filings so extraordinary is not just what they say but what they dismantle. For years, the dominant narrative around large language models has been one of transformation and progress. AI companies have framed their work as building something fundamentally new — tools that synthesize, reimagine, elevate human creativity rather than replace it. They have spent billions on marketing, on policy briefings, on congressional testimony, arguing that their training methods qualify as fair use because the output is qualitatively different from the input.

The unsealed documents undercut that entire framing. An internal Microsoft document stated that “millions of people around the world will soon consider large models hoovering up all their work to be an astonishing theft of unprecedented proportions.” It added, almost as an afterthought: “Almost no one intended for content they created to be used in this fashion, nor are they compensated for its use.”

Greg Brockman, OpenAI cofounder, was quoted in court documents describing how the company developed “a hack to get around nytimes paywall,” and responded with a dismissive “ah, nice” when asked about it. Brent Hecht, a Microsoft executive, wrote that LLMs steal content “without ways of distributing economic value down the supply chain, [which] necessarily threatens the economic stability of those who create the content.” Jack Clark, OpenAI’s policy director, acknowledged internally that the company was “creating systems that substitute for the labor of the people that define the culture of society.”

Satya Nadella himself testified under oath that after ripping content from the New York Times and other news sites, clicks to those news sites cratered by more than 90 percent on Bing.

These are not peripheral comments from mid-level engineers. These are the top executives of two of the most powerful technology companies in the world, speaking in internal documents and depositions, about what their products do and whom they harm.

The legal implications are enormous

The most immediate consequence of these filings is legal. The New York Times is seeking summary judgment — a court ruling in its favor without a full trial — based in large part on the admissions contained in these documents. If the court accepts the argument that the training data was obtained through circumvention of paywalls and used without compensation, that would represent a catastrophic shift in the legal landscape for the entire AI industry.

OpenAI and Microsoft have been building their fair-use defense on the argument that their models are transformative — that they synthesize and reinvent rather than simply copy. But the internal documents reveal executives who understood exactly what they were taking and whom they were taking it from. Nadella’s testimony about the 90 percent click decline on Bing after scraping the Times directly contradicts the notion that AI output coexists with human journalism. It describes a zero-sum relationship: more AI usage means less traffic to the very sources that made the AI possible.

Jason Kint, CEO of Digital Content Next — the trade organization representing the news publishers behind the lawsuit — was the one who found and publicized the unsealed filing. His role in bringing this to light should not be underestimated. The AI industry has spent years operating in a regulatory gray zone, building products on content it never licensed, never paid for, and never intended to compensate. The legal system was the only institution with the power to pierce that opacity.

What happens next for the industry

The broader implication is existential for the business model that has driven the AI gold rush. Every major generative AI product — ChatGPT, Copilot, Gemini, Claude — was trained on datasets scraped from the internet at scale. The assumption, baked into every pitch deck and valuation, has been that this is legally permissible and economically sustainable. These documents tear that assumption apart.

If courts rule against the defendants in the New York Times case, the precedent could extend far beyond news publishers. Artists, musicians, software developers, bloggers, educators — anyone whose creative output was scraped for training data could argue their case on similar grounds. The “largest theft of labor in human history” phrasing, which appeared in a Microsoft document cited in the filing, is not hyperbole from a hostile witness. It is an internal description of the industry’s own relationship to the content it consumes.

The economic consequences are already visible. Publishers that supplied training data are seeing their own traffic destroyed by the products they unknowingly built. Writers and artists who were never consulted are seeing their work replicated without compensation. The doom loop Microsoft described is not a theoretical risk — it is an active, measurable phenomenon. Bing traffic to the New York Times dropped more than 90 percent after the company began using Times content in its AI models.

The regulatory reckoning

Perhaps most significantly, these filings provide ammunition for regulators who have been waiting for proof that the industry’s own executives understood the damage their products caused. The argument that AI training constitutes fair use has rested on the premise that the technology was creatively transformative rather than extractive. The internal documents reveal executives who saw it exactly the other way around — as extraction, as theft, as a business model built on the destruction of its own supply chain.

The regulatory implications will be felt immediately. Lawmakers in the United States and the European Union have been drafting legislation on AI transparency, data licensing, and creator compensation. These documents provide a factual foundation for those efforts — not speculation from outside critics but testimony from inside the companies themselves.

The question now is whether the courts and regulators will act on what they have learned. The AI industry has spent years moving fast and breaking things. The doom loop documents show that, in this case, what was broken was the internet itself — the ecosystem of human creators, journalists, and artists that made generative AI possible in the first place. The reckoning may finally be here.