Unsealed documents show AI companies privately modeled a ‘doom loop’ for the web

An unredacted summary-judgment filing in The New York Times v OpenAI exposes unusually blunt internal assessments of the economics behind generative AI.

A Microsoft document described a “doom loop”: models depend on publishers and other creators, while AI answers reduce visits and revenue to those same suppliers, eventually damaging both the web and the future content supply on which models rely. Microsoft executives testified that clicks to news sites from Bing could fall by more than 90% when an AI answer appeared. An OpenAI engineer similarly said prominent links would not make most users click.

The filing also cites an internal reference to a “hack” for getting around the Times paywall and statements recognizing that creators did not expect or receive compensation for model training. OpenAI and Microsoft contest the plaintiffs’ legal characterization and continue to argue that training is transformative fair use. The quoted documents establish internal awareness of substitution and economic harm; they do not by themselves decide whether copying was lawful.

The Information’s Martin Peers emphasizes the business contradiction: executives now worry about the survival of publishers while their products capture the publishers’ audience. 404 Media presents the filing more aggressively as an admission of theft. The legally careful reading is narrower but still damaging. The defendants’ own material appears to undermine the claim that AI merely complements original publishing or reliably returns value through links.

The case could turn on doctrine such as fair use, market substitution and evidence about particular copied works. Regardless of the verdict, the documents reveal that the industry understood the supply-chain problem internally: if answer engines absorb demand without financing reporting, model quality can rise in the short run while the information ecosystem that refreshes it deteriorates.