AI copyright filing reveals Microsoft's Project Mango scraper

A less-redacted motion, a sharper picture

The biggest US copyright fight over AI training data got more detailed on October 16. News publishers suing OpenAI and Microsoft filed a less-redacted version of their motion for summary judgment. In it, they argue that the companies' AI products work as substitutes for the journalism used to build them.11 The new disclosures cover more than market harm. They describe how training data was allegedly gathered, how it moved between the two companies, and what each company knew about the consequences.

The most striking new detail is a scraping effort code-named "Project Mango." According to the motion, OpenAI paid Microsoft to build and run a web crawler that copied webpage content, and that material was later used to train models including GPT-4o.1 The filing also says Microsoft told employees to keep the project "secret."1 The publishers say the scraped data contained more than 160,000 of their works. The price OpenAI paid and the total number of pages crawled are still under seal.1

The filing adds to what the parties call "horse trading" deals. Among other things, it says OpenAI handed Microsoft its complete GPT-3 training dataset.1 Microsoft had first pushed to keep this material redacted, but it dropped some of those requests on October 9 after the publishers objected.1

The 'doom loop,' now with more numbers

The market-harm argument is not new. Last month's first unsealing introduced the "doom loop" theme. Coverage at the time quoted a Microsoft document warning that the company's AI content strategy threatened its own "content supply chain."18 It also reported click-through rates from Bing Chat that were 87% to 93% lower for the Times' websites than from traditional Bing search.16

The October version widens that picture. It cites Microsoft data showing click-through drops of 83% to 93% for the domains of The Times and the Daily News plaintiffs, and 51% to 94% for Ziff Davis domains, when Copilot's "answer engine" is compared with classic Bing.11 The publishers also quote Microsoft's lead counsel saying that chatbots are "designed to answer your question" rather than send users to a list of links. They present that as an admission that the products pull readers away from the outlets that wrote the original work.4

The filing also estimates how much demand for news flows through ChatGPT. The publishers take an OpenAI expert's figure that 1.88% of ChatGPT messages in July 2025 sought news and apply it to the chatbot's 2.5 billion daily prompts. That produces their estimate of about 47 million news-seeking prompts a day.1 Earlier unsealed material cited OpenAI's own internal tracking of about 1 million weekly prompts for reliable local news, which the publishers called only part of the picture.18

The figures do not all line up exactly across the coverage. The ranges for the Times shift slightly between the September and October versions, likely because they compare different products (Bing Chat versus Copilot) or different plaintiff groupings.1611 The 47 million figure is the publishers' own extrapolation, not a number OpenAI has accepted.1 Even so, the overall argument holds steady across every account: the defendants' own records show traffic falling.

Why the internal quotes matter legally

Much of the attention has gone to blunt internal statements. Microsoft applied science director Brent Hecht is quoted calling the copying perhaps the "largest theft of labor in human history."14 OpenAI's head of ChatGPT, Nick Turley, reportedly wrote that publishers faced an "existential threat" from products that were "largely substitutive."16 Microsoft CEO Satya Nadella testified in a deposition that paywalled material should be licensed by anyone who wants to use it for grounding or training.6

The quotes matter most under the fourth fair-use factor, which asks whether the use harms the market for the original work. Statements showing that insiders expected chatbots to replace visits to publisher sites help the plaintiffs argue that the harm was foreseeable, not hypothetical.6 Legal opinion on how much this helps is divided. One consultant told Digiday that Turley's remark would be "difficult to walk back." Another lawyer called the documents damaging but not decisive, noting that courts have still treated some AI training as transformative.8

The defendants are pushing back hard. A Microsoft spokesperson said the comments reflect one employee's personal view, not company policy or legal analysis.16 A Microsoft declaration described Hecht as someone paid to argue the contrarian side.18 OpenAI points to two June 2025 rulings, Bartz v. Anthropic and Kadrey v. Meta, which found AI training highly transformative. It also cites its licensing deals with more than 20 publishers.6

My reading: the Project Mango disclosures may end up mattering more than the colorful quotes. If OpenAI paid for content scraped secretly and traded datasets with Microsoft, that supports the publishers' claim that this content has a market price that both companies recognized. One licensing-coalition founder made that point directly.8 Courts can dismiss an employee's ethical worry as opinion. Showing that something was bought and sold is harder to dismiss.

The procedural stakes and the calendar

The case is In re OpenAI, Inc. Copyright Infringement Litigation, a multidistrict proceeding before Judge Sidney Stein in the Southern District of New York.14 Both sides filed cross-motions for summary judgment on September 4.17 The publishers want liability findings at five stages: acquiring articles, training on them, Microsoft's use of Bing-indexed copies to ground answers, reproducing articles in outputs, and the companies supplying copies to each other.2 OpenAI wants a ruling that pretraining is fair use. Microsoft wants judgment on every claim against it.2

The two sides describe the outputs very differently. OpenAI's expert reportedly found 24 cases of verbatim reproduction in a sample of 20 million ChatGPT logs. Microsoft's brief says matching text of 16 or more words appeared in about 0.73% of 8.2 million Copilot conversations.2 The publishers want the court to bar OpenAI from relying on that 20-million-log sample at all. That request is part of a sanctions motion accusing the company of hiding its ability to search training data and output logs.27 The sanctions motion has not been decided.2

Oppositions are due October 23 and replies November 20. No hearing date has been set, and a ruling for the defendants on fair use could end much of the case without a trial.2 Other cases depend on the outcome. A suit by publishers of nearly 400 local and regional newspapers is on hold until the summary judgment motions are decided.15

Washington's thumb on the scale

The dispute now has a political side. The Justice Department filed a statement of interest supporting OpenAI's fair-use defense.13 Coverage quotes it as calling training "exceedingly transformative" and warning against liability that would effectively require licenses for AI training.2 The government based its position partly on "America's global AI dominance" and national security. The publishers asked Judge Stein to give the statement no weight. They argued it ignores the Copyright Office's opposing conclusion13 and was filed before the summary judgment record was unsealed, so it did not consider the evidence.2

That timing argument gets stronger with each unsealing. A policy statement about training in general is a weak answer to specific allegations of secret crawlers and traded datasets. Judges usually decide summary judgment on the record in front of them, not on administration priorities.

The wider litigation map and what legal teams should take from it

This case is one of many. Publishers have also sued Perplexity, Cohere and others, often targeting retrieval-augmented generation, the grounding method at the center of the Bing-index claims here.15 The Supreme Court's March 2026 Cox v. Sony decision has already narrowed this case. The Times and Daily News dropped their material-contribution claims based on user infringement, and the court refused to let them add new theories against Microsoft.2 Contributory liability is now harder to win across AI cases, which makes direct-infringement and fair-use questions about training and grounding more important.

This matters well beyond newsrooms. Lawyers and compliance teams buying generative AI tools are watching a case where the key evidence is about where data came from, not how the model performs. The sanctions fight is also a warning about discovery: companies that say log searches are infeasible may later find their own engineers contradicting them.7 Organizations adopting AI should expect questions about training-data provenance and output logging to become routine in vendor due diligence, whatever Judge Stein decides.

The ruling could still go the defendants' way. Two federal judges have already found training transformative, and the DOJ supports that view.62 Still, the record unsealed over the past month gives the publishers their strongest market-harm evidence yet. Bloomberg Industry Group is seeking to unseal more of it, so further disclosures are likely.2