Internal documents reveal OpenAI and Microsoft warned each other about AI's "doom loop" threat to the web

Freshly unsealed documents filed in the New York Times' legal action against OpenAI and Microsoft contain damaging admissions from the companies themselves. Internal records show executives and employees explicitly warned that their AI development strategy risked creating a self-destructive cycle for online publishing, even as they proceeded with large-scale data scraping to train their models.

What the documents reveal#

The court filing contains stark language from company insiders. Microsoft's applied science director called the practice the "largest theft of labor in human history" and said the company's copyright defense made "a complete mockery of fair use." An internal Microsoft strategy document warned of a "doom loop" — a scenario where AI models trained on web content would eventually undermine the economic incentives that generate that content in the first place, damaging both the publishing ecosystem and the models themselves.

Executives including Satya Nadella acknowledged that AI chatbots had essentially replaced traditional search, eliminating the need for users to visit original sources. Yet despite Nadella's later public statements about licensing paywalled content, OpenAI representatives admitted they made no meaningful effort to filter protected articles from their training datasets.

The copyright problem#

Internal discussions at OpenAI showed the company was aware of ChatGPT's tendency to reproduce substantial passages from copyrighted articles verbatim. Employees acknowledged the importance of preventing memorization to limit copyright breaches, but noted that GPT-4 "memorized a ton of data" and would be "insanely good at regurgitation." The court filing cites instances where ChatGPT reproduced long sections directly from publications including the Times, Mercury News, The Denver Post, LifeHacker, and Eurogamer.

How the system breaks itself#

Multiple internal documents described an inherent contradiction: large language models fundamentally depend on training data from human creators, yet their existence as functional products actively discourages the creation of that data. An OpenAI policy director reflected that the company was "creating systems that substitute for the labor of the people that define the 'culture' of society." Microsoft characterized LLMs as "a product that destroys its own supply chain."

OpenAI's own media analysts estimated that AI-powered search summaries from competitors like Google had reduced referral traffic to major publishers by as much as 60 percent — meaning fewer readers reach original sources and fewer incentives exist for quality journalism.