The AI ‘Doom Loop’: Internal Warnings That Chatbots Could Gut the Web Economy

Artificial intelligence companies usually talk about the open web as a resource. Behind closed doors, newly unsealed court filings suggest, some of their own executives have described it as something their products might accidentally destroy.

The documents emerged from The New York Times’ copyright lawsuit against Microsoft and OpenAI. Among the internal material quoted in the filings is a Microsoft warning that the company’s AI content strategy had started a ‘doom loop’ that could hurt the performance of its own models and the entire web at the same time. It is an unusually frank admission from inside one of the industry’s biggest AI investors.

How the Doom Loop Is Supposed to Work

The mechanism described in the filings is straightforward. AI models are trained on enormous quantities of published material — journalism, reference works, forums and books. Chatbots then answer questions directly, in their own interface, instead of sending readers onward to the sources that produced the underlying information.

If readers stop visiting, publishers lose advertising and subscription revenue. With less revenue, they produce less original reporting and fewer expert resources. And with less fresh, high-quality material being published, the training data available to future AI models gets thinner and worse. The product, in other words, slowly starves its own supply chain.

The filings cite steep reported falls in click-through rates to publisher sites once AI-generated answers appear above search results, with some figures in the range of 83 to 93 percent, and consultant estimates of search-referral declines of as much as 60 percent. Coverage by 404 Media summarised the filings as an acknowledgement that AI products can cannibalise the businesses they learned from.

Strong Language From Inside the Building

Some of the quoted internal language goes well beyond concern about traffic. One senior Microsoft scientist described large-scale scraping of creative work in terms of historic theft of labour, and questioned whether the industry’s fair-use defence could withstand scrutiny. OpenAI figures, according to the filings, described their products internally as largely substitutive for publisher websites — designed to replace visits, not to route them.

Microsoft’s chief executive, testifying under oath, acknowledged that chatbot conversations had substituted for visits to publisher platforms, because answers now arrive inside the AI interface itself. Both companies have publicly distanced themselves from some of the sharper internal characterisations, describing them as individual views rather than company positions.

Why This Matters Beyond the Media Industry

It would be easy to file this story under ‘newspapers versus tech giants’ and move on. That would miss the wider business stakes. The same referral economics underpin comparison sites, specialist blogs, review publications, documentation portals and much of the small-business web. When answers are absorbed into a chatbot, the commercial model of every site that depends on a visit is weakened — a shift our coverage of the AI spending boom shows is being funded at extraordinary scale.

There are second-order effects too. Sites that survive will optimise harder for the traffic that remains, which can mean more aggressive paywalls, more search-engine gaming and less accessible information. Even user experience work, such as making navigation easier on mobile, matters less if audiences never reach the page at all.

What Could Break the Loop

The filings have strengthened calls for licensing arrangements that pay publishers for training content and for referral traffic, an approach some large publishers have already adopted through commercial deals with AI firms. Supporters argue that a functioning market for content would align incentives: AI companies get reliable, permissioned data, and publishers get a revenue stream that does not depend entirely on clicks.

Critics of that model warn it could entrench the largest media owners, leaving independent creators — whose work also trains models — uncompensated. Regulators in several markets are watching the litigation closely, since its outcome may shape how copyright applies to machine learning far beyond one newspaper’s archive.

The Bottom Line

Whatever the court decides, the unsealed material has already changed the debate. The risk to the web’s economics is no longer just a complaint from publishers; it is documented, in the industry’s own internal words, as a scenario its executives saw coming. For businesses that publish anything online, the practical lesson is to build audiences that do not depend on a single referral channel — because the loop described in these filings is still turning.