OpenAI and Microsoft Pushed Ahead While Recognizing 'Doom Loop'

Documents disclosed by court order in the lawsuit filed by The New York Times against OpenAI and Microsoft are causing ripples. According to reporting by The Verge’s Terrence O’Brien, internal documents and statements by people connected to both companies were cited across 92 pages. The documents included warnings that their own businesses would harm the web. They also contained harsh criticism of collection for training generative AI.

The disclosure came to light in reporting as of September 18, 2026. Much of the content is unfavorable to both companies. It paints a picture of development moving forward even as risks were shared internally. The confrontation with publishers has entered a new stage.

Contents of the 92-Page Document

Disclosed in Court

The disclosed documents compile internal records supporting the plaintiffs’ claims. Cited are statements by executives, researchers, and developers. Records of statements by management are also included. There are many excerpts from internal strategy documents.

At the center is the legitimacy of collecting and using training data. The shift from search to conversational AI and the blow it deals to publishers is also at issue. The documents show that both companies understood the causal relationship. It was suggested that it may not have been an incidental side effect.

Regarding model evolution, OpenAI Announces GPT-5.6 With Three Models Sol, Terra, and Luna shows the latest milestone. Meanwhile, these documents question the process by which that foundation was formed. The focus is the imbalance between performance gains and rights clearance. The history of the technology is now being examined in court.

largest theft of labor in human history

complete mockery of the idea of fair use

The Structure of the Doom Loop Warned About

Internally

The most attention-grabbing expression was doom loop. The original English term is doom loop. Internal documents stated that their own AI strategy had started this cycle. They warned that it would simultaneously harm model performance and the web as a whole.

The structure can be summarized as follows. Conversational AI shows answers directly. Users no longer visit source sites. Publishers’ revenue base shrinks. Production of new reporting and verified articles declines. High-quality information available for training decreases. Model quality also declines.

This cycle is unusual in that it undermines its own supply side. The internal documents were aware of that abnormality. They noted that the final product threatened the economic foundation of suppliers on which it indispensably depends. The contradiction inherent in the business structure of large language models was exposed.

Collection Called the Largest Theft in

Human History

Statements by Brent Hecht, head of Microsoft’s applied sciences department, were cited. He described collection by ChatGPT and Copilot in harsh terms. The expression largest theft of labor in human history was used. It was recorded as criticism of the reality of the training process.

These remarks are positioned as raising issues from an academic standpoint. Hecht has studied the state of the data ecosystem. He is said to have played a role in presenting an asymmetric and forward-looking perspective internally. The fact that theoretical warnings existed internally carries weight.

The collection allegedly included news articles and explanatory pieces. Fragments of headlines and body text were reportedly used for training. No system for rights holders’ permission or compensation was in place. On data processing transparency, there are later measures such as OpenAI Announces Private Safety Processing With Zero Data Retention. However, the legitimacy of past collection remains as a separate issue.

Internal Criticism Shaking the Principle

of Fair Use

The documents also criticized the treatment of fair use. They included a statement calling it a complete mockery of the idea of fair use. Doubts were raised about simply accepting the U.S. copyright law defense. Voices from within questioned the validity of the legal argument.

Fair use is judged by purpose and character, amount used, and effect on the market. There is a view that it differs in nature from copying for search indexing. Generative AI can produce outputs that substitute for original texts. The judgment of market substitutability is at the heart of the lawsuit.

The plaintiffs claim an existential threat. The cited internal statements reinforce that perception. Statements by OpenAI’s head of ChatGPT were also cited. In reporting, he was presumed to be Nick Turley. There was a shared understanding that it concerned the survival of the publishing business.

Company Stance That Critical Statements Were

Personal Views

Microsoft distanced itself from Hecht’s remarks. Spokesperson Alex Haurek said they were personal views. He explained they were not legal analysis and did not represent the company’s views. This was recorded as a response to inquiries from The Verge.

A similar framing was presented in court filings. Jordan Usdan, an executive in Microsoft AI’s data strategy division, explained. He said Hecht’s role was to present an adversarial perspective. He described him as a person with divergent, academic, and forward-looking views. He argued that theoretical views do not represent the company.

This framing is standard litigation defense. It separates individual statements from organizational intent. It relies on the freedom of research divisions for its positioning. Meanwhile, the fact that warnings existed inside the company does not disappear. The question is how the organization received them.

Regarding corporate posture, attention is also gathering on the response to Microsoft Defender Privilege Escalation Vulnerability ‘RoguePlanet’ Disclosed. Efforts on safety and governance are subject to evaluation. These documents will serve as material to question their effectiveness. The focus is not only on how statements were handled but on operational records.

Executive Statements Acknowledging Search

Replacement

The documents also included statements by Satya Nadella. They showed recognition that conversational AI had largely replaced search. He said the need to go directly to information sources had diminished. It amounts to management acknowledging a change in user behavior.

This change is called Google Zero. It refers to the dwindling of traffic via search. Summary answers by AI take away dwell time. Media outlets’ advertising revenue and subscription referrals are harmed. The premises of the web economy are collapsing.

Statements by Sam Altman and other OpenAI-related figures were also cited. Details of the specific wording will be scrutinized in future proceedings. The relationship between management decisions and awareness is at issue. Whether they proceeded knowingly will affect liability judgments. The 92-page document aims to prove that.

Spreading Damage to Publishers and the Web Economy

Publishers allege their basis for survival is being undermined. Reporting costs labor and time. If AI extracts only the results, compensation does not flow back. Maintaining high-quality journalism becomes difficult. Local and investigative reporting are hit hardest.

There are also concerns about declining quality across the web. Updates of primary information stall. AI-generated secondary information proliferates. There is fear that errors will be learned in a circular manner. The doom loop warning referred to this situation.

The convenience of conversational AI and the maintenance of the information ecosystem must coexist. Permission and compensation, attribution, and referral design are challenges. Talks between tech companies and publishers continue intermittently. The current disclosure will also affect negotiating terms. Court rulings will shape industry practices.

Editorial Opinion

On short-term impact. We see this disclosure hardening negotiations over the next three to six months. We see publishers strengthening moves to seek higher licensing fees. We see AI companies being pressed to explain the scope of collection and present compensation frameworks. We see an increase in contracts, including out-of-court settlements. We see stronger demands for disclosure of contract terms.

On the long-term perspective. We see the procurement structure for training data changing over a one- to three-year horizon. We see a shift from large-scale unlicensed collection to a combination of contracts and synthetic data. We see distribution to providers of primary information becoming institutionalized. We see search-linked revenue models shrinking. We see media outlets leaning toward subscriptions and business information sales.

We present questions from the editorial team. Could the decision to proceed while sharing warnings internally be justified? Who should draw the line on the scope of fair use, and how? Which actors should bear the cost of stopping the doom loop? Can the benefits of technology and the sustainability of information production coexist? Can industry norms be created without waiting for the court’s conclusion?

References

• “OpenAI and Microsoft knew they were starting a ‘doom loop’ for the web”, by Terrence O’Brien — The Verge, 2026-09-18T21:07:24.000Z (ARR)
• Source URL: https://www.theverge.com/ai-artificial-intelligence/997633/openai-microsoft-chatgpt-ai-new-york-times-doom-loop-theft-google-zero