Newly unsealed documents in the New York Times’ copyright lawsuit against OpenAI and Microsoft show that AI scraping was described internally as “the largest theft of labor in human history.” Filings published September 17, 2026 also reveal that Microsoft’s own Copilot product caused click-through rates to the Times’ website to drop by as much as 93% compared to traditional Bing search.
What did the unsealed filings reveal about AI scraping practices?
The new material comes primarily from The New York Times’ own legal brief, not the underlying exhibits, which remain sealed. Filed in the lawsuit the Times brought three years ago against OpenAI and Microsoft, the unredacted sections include internal communications, deposition testimony, and company documents previously hidden from public view.
Key disclosures, per the filing as reported by TechCrunch on September 17, 2026:
- In a January 2023 internal memo, Brent Hecht, Microsoft’s Director of Applied Science, called the practice “an astonishing theft of unprecedented proportions” and “the largest theft of labor in human history.”
- OpenAI’s mid-training datasets (data used partway through the model-building process) alone contain more than 91,692 copies of works published by the NYT, Daily News, and Center for Investigative Reporting.
- A Common Crawl-derived dataset (Common Crawl is a publicly available repository of web data collected by automated crawlers) included more than 2 million documents from nytimes.com alone.
- A dataset assembled under an initiative called “Project Mango” contained copies of at least 160,903 unique works from the news publishers.
- Researchers allegedly stripped copyright notices from training data before it reached the model, because they “wouldn’t want model outputting” copyright notices to users.
How did they allegedly bypass paywalls?
The filings allege that OpenAI employees devised methods to access paywalled content without detection. Internal communications show that when OpenAI researcher Nick Ryder informed OpenAI President Greg Brockman about a “hack to get around nytimes paywall,” Brockman replied: “ah nice.” Separately, Microsoft provided training data to OpenAI through initiatives called Project Taxi and Project Mango, per the filing.
Microsoft CEO Satya Nadella, testifying in a deposition earlier in 2026, stated that “anything that is paywalled should be licensed by anyone who wants to use it…for grounding (using external data to anchor an AI’s response) or training.” He added that had he known OpenAI scraped paywalled content, he would have “invoked [Microsoft’s right to] require OpenAI to retrain its models.”
What does a 93% click-drop mean for publishers?
An internal Microsoft presentation authored by Hecht in January 2024 showed that Copilot’s answer engine caused click-through rates to the Times’ domain to drop as much as 93% versus traditional Bing search. Hecht’s document described this as a “doom loop”: fewer clicks to publishers means less revenue to fund content creation, which means less quality material available to train future models.
“It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its ‘content supply chain,'” reads the Microsoft document, as quoted in the filing.
OpenAI’s head of ChatGPT, Nick Turley, wrote in internal communication that publishers face an “existential threat” from products like the chatbot, which are “largely substitutive” and “will get more and more substitutive as they get better.” Nadella agreed under oath that conversing with chatbots “has substituted…giving you the information right there on the website on the AI platform versus needing to go to the underlying source.”
Where does the fair use argument stand?
The legal question of whether AI firms can train on copyrighted content without permission remains unresolved. Judges have so far been broadly favorable to AI companies’ “fair use” arguments, and in September 2026 the Trump administration filed a brief supporting OpenAI’s position on unlicensed use of copyrighted material.
Fair use (a U.S. copyright doctrine permitting certain uses of protected work without permission, such as commentary or parody) requires, among other factors, that the use not substitute for or harm the market for the original work. Several of the newly unsealed admissions, which acknowledge substitution and potential market harm in internal documents, could complicate that defense, according to TechCrunch’s reporting.
“The evidence revealed here for the first time shows that OpenAI and Microsoft knew that what they were doing was wrong,” Steven Lieberman, counsel for the New York Daily News, said in a statement. OpenAI and Microsoft did not return requests for comment.
What this means for AI-search visibility
Training an AI on your content and having that AI send traffic back to you are two separate things, and these filings are the most direct evidence yet that they can actively work against each other.
The 93% click-drop figure from Microsoft’s own internal data is the most concrete measurement yet of what GEO (Generative Engine Optimization, the practice of improving a brand’s visibility inside AI-generated answers) practitioners call the substitution problem. An AI that has learned from your content can answer topic-related questions without ever directing anyone to your site. Picture a restaurant competitor who received your full recipe book: their food gets better, and your regulars have less reason to visit.
There is a second, less-discussed implication in the alleged stripping of copyright notices from training data. If attribution signals were removed before content reached the model, many AI systems may have absorbed publisher knowledge without retaining the connection between that knowledge and its original source. The model knows the information; it may not know who produced it. That is a structural problem for any brand trying to build cited authority inside AI answers, because the pathway from “AI learned from you” to “AI credits you” may have been broken at the data-preparation stage.
The “doom loop” framing also extends well beyond news publishing. Any content-heavy vertical, including health, finance, legal, and travel, faces the same dynamic: if AI answer engines reduce click-throughs at scale, the revenue model sustaining content creation weakens, producing less fresh material for future model updates. Hingewise’s assessment: the long-term risk is not just that AI replaces traffic today, but that it degrades the content ecosystem it depends on, eventually reducing the quality of its own answers.
One open question worth tracking: whether successful copyright claims could force retraining of existing models, or establish licensing norms that reshape which content AI systems are permitted to learn from going forward. Either outcome would significantly affect which brands and publishers hold share of model (how prominently a brand appears in AI-generated responses relative to competitors) in the next generation of AI search products.
5 things to check before assuming AI can find and credit your content
- Determine whether key content is behind a paywall or login gate, and check each major AI platform’s stated crawling and licensing policy for gated content.
- Confirm that your pages include clear authorship metadata and structured markup, since the filings suggest some training pipelines removed attribution data.
- Test whether AI answer engines (ChatGPT, Gemini, Perplexity, Google AI Overviews) answer questions in your category by citing your domain or a competitor’s.
- Review referral analytics for changes in AI-adjacent traffic and compare against organic search trends over the same period.
- Note whether AI tools summarize your content in answers without linking back, which is the substitution effect in practice rather than theory.
The case remains active. Further exhibits are sealed and may produce additional disclosures as proceedings continue.
Source: TechCrunch, September 17, 2026
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.