Unsealed Filings Show OpenAI and Microsoft Recognized “Unprecedented Theft”

Unsealed court documents have been made public in the copyright lawsuit filed by The New York Times against OpenAI. They include passages showing that Microsoft and OpenAI executives were critically aware of the realities of large language model training and operation. Reporting by Jason Koebler of 404 Media detailed how unredacted versions of internal documents and deposition transcripts were disclosed on September 17, 2026. The documents accompanied the plaintiffs’ motion for summary judgment, bringing to light portions that had previously been redacted or kept under seal. This adds new material to the dispute over generative AI training data and news organizations’ rights.

At the heart of the documents is a frank reference to the fact that training has depended on human creations. A Microsoft internal document recorded in strong terms the act of large models hoovering up the work of people around the world. According to the plaintiffs’ citation, the view was expressed that the act would be seen as theft on an unprecedented scale. It was also noted that many creators did not anticipate such use and have received no compensation. The content corroborates, through the developers’ own internal records, the structure critics have long alleged.

The relevant passage in the original is quoted as follows.

Millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions.

Furthermore, the same document touched on creators’ intent and the lack of compensation.

almost no one intended for content they created to be used in this fashion, nor are they compensated for its use.

Elsewhere, the phrase the largest theft of labor in human history is said to have appeared. These wordings were organized by the plaintiffs as summaries in the court filings. They are presented not as an isolated metaphor, but as showing multiple internal records and testimonies pointing to the same awareness.

Background and Disclosure of the New York

Times Lawsuit

The lawsuit by The New York Times has continued for a long time over the unauthorized use of its articles. At issue are reproduction and memorization during training, and whether generated outputs substitute for the market. This disclosure consists of the group of documents submitted by the plaintiffs as evidence with their motion for summary judgment. Portions that had been sealed or redacted at the request of Microsoft and OpenAI came to light through court proceedings. As for how they were discovered, Jason Kint, CEO of the digital media trade group Digital Content Next, is said to have obtained the documents and publicized them.

Summary judgment is a procedure asking the court for a ruling on the grounds that there is no dispute over the facts. The plaintiffs built their case by accumulating statements by the defendants’ executives and internal assessments. The sources of training data, the competitiveness of generated outputs, and the impact on the media business were argued together. The reason the defendants wanted to avoid disclosure is seen as lying in the bluntness of the language. Future proceedings will likely see continued exchanges over the context of the quotations and their admissibility as evidence.

As for explanations of safety measures around OpenAI, OpenAI Announces Private Safety Processing, Achieving Zero Data Retention illustrates a known initiative. These documents separately raise the question of how to evaluate the origins of training, apart from such institutional design. The efficacy of the technology and the legitimacy of its sourcing need to be verified on separate dimensions. Factual findings in court could also affect product design philosophy. Record-keeping and accountability in the development system are once again in focus.

Internal Documents Show Awareness of

Unprecedented Theft

The language in the internal documents shows an awareness that training was supported by reproduction without permission. The phrase theft on an unprecedented scale includes an assessment of both volume and method. The scope was described as extending beyond news articles to a wide range of creations, including text and images. Unintended use by creators and a structure without compensation were also clearly noted. This leads to the view that the absence of rights clearance supported the speed and scale of development.

Deposition transcripts reportedly also discussed the potential to displace human writers, artists, and media companies. The shift from training subjects to competitors through outputs was recognized internally. Apart from the future dangers of superintelligence, the pressure current tools place on the labor market was indicated. While narratives of increasing capability ran ahead, substitution on the ground was progressing. This gap adds to the weight of this disclosure.

Regarding Microsoft’s business operations, as illustrated by Microsoft Defender Privilege Escalation Vulnerability “RoguePlanet” Disclosed, situations calling product reliability and management systems into question continue. Management of training data is also emerging as a similar governance challenge. Preserving records, identifying the scope of use, and explaining to rights holders will be essential. Consistency between internal assessments and external explanations will also be subject to scrutiny. Coordination between technology development and legal controls has become a condition for business continuity.

The Doom Loop Eroding the Web Economy

Another key point is the internal assessment that generative AI products harm the structure of the web. The documents referred to a doom loop called “doom loop.” It refers to the flow of ingesting human-made content, then taking views away from the sources and undermining their business foundations. Behind this is a structure in which direct answers via search or summarization reduce traffic to media outlets. Revenue long supported by advertising and subscriptions is being pushed to shrink.

This cycle was described as rebounding on model performance as well. The view is that if the quality and diversity of the open web decline, future training resources will be depleted. Outputs threaten the web, and the decline of the web starves the foundation of outputs. Expanded product use and maintenance of the information ecosystem are in a relationship difficult to reconcile. There is documentary value in the fact that the developers’ internal documents noted this paradox.

Concerning the operating environment for infrastructure, development of testing infrastructure is progressing, such as GNOME OS Test Center, Inspired by Apple TestFlight. Even in evaluating generative AI, reproducible and transparent verification is required. Display of sources, measurement of referrals, and design of revenue sharing are at issue. The convenience of summaries and the sustainability of originals is a relationship to be balanced by design. Whether there are concrete measures to break the cycle will be a future focus.

Threat to the Labor Market and Industry

Group Reaction

The documents do not treat the danger of labor displacement as an exceptional event. Direct pressure on creative labor was noted, including writers, editors, and visual artists. There is a dual effect: absorbing value through training and replacing commissions and views through generation. Media companies have been placed in the position of facing competition while being providers of source material. Uncompensated use combined with market substitution increases the difficulty of sustaining business.

Industry groups’ reactions are also leaning in a severe direction. Digital Content Next, as a group representing digital media companies, has followed the litigation. This disclosure will serve as material supporting calls for a framework of permission and compensation. Apart from individual settlements, a review of contracting practices across the industry may be discussed. Permission for training use, calculation of compensation, and conditions for display and referrals are the main issues.

Technological progress and protection of rights should not be fixed as opposing concepts. Implementation of permission acquisition, opt-outs, revenue sharing, and other measures remains as design challenges. There is a view that returning value to creation is essential for sustaining training resources. Court rulings could provide a reference point for negotiations. Rebuilding the relationship between developers and media will be the next stage’s challenge.

Editorial Opinion

In the next 3-6 months, proceedings over whether summary judgment will be granted will dominate coverage. The language in the disclosed documents could become material for trial or settlement negotiations. AI operators will be pressed to explain disclosure of training data and compensation frameworks. Moves toward contracts, including with news organizations, are also expected to accelerate.

In 1-3 years, the relationship between training data procurement and web publishing will be redesigned. Reliance on permission-based acquisition with rights clearance and on synthetic data will increase. The balance of power between media and foundation developers will likely be reflected in contract terms. Maintaining the information ecosystem will become a condition of development competition.

Remaining questions include how liability will be constructed if awareness of theft existed internally. Who should bear the cost of correcting the doom loop also remains undetermined. How should the danger of superintelligence be weighed against immediate labor displacement? Both proof and institutional design will be tested.

References

• “‘Doom Loop’: OpenAI and Microsoft Admits LLMs Are Destroying the Web and Built on Theft”, by Jason Koebler — 404 Media, 2026-09-17T22:01:43.000Z (ARR)
• Source URL: https://www.404media.co/doom-loop-openai-and-microsoft-admits-llms-are-destroying-the-web-and-built-on-theft/