AI copyright case filing exposes Microsoft Project Mango crawler

A less-redacted brief, and a new code name

The biggest US copyright fight over generative AI has become a steady release of internal documents. On October 16, the news publishers suing Microsoft and OpenAI filed a less-redacted version of their motion for summary judgment. They argue that the companies' AI products act as substitutes for the journalism used to build them.11 The most striking new detail is "Project Mango." According to the publishers' motion, it was a program in which OpenAI paid Microsoft to build and run a web crawler. The content that crawler copied was then used to train models including GPT-4o.13 Microsoft allegedly told employees to keep the project "secret." The publishers say the scraped data included more than 160,000 of their works. How much OpenAI paid and how many pages were collected are still under seal.8

The new filing adds to a record that has been made public in stages. Microsoft first argued that the material should stay redacted. On October 9 it withdrew some of those requests after the publishers objected.8 Bloomberg Industry Group, which owns Bloomberg Law, has also moved to intervene to get more of the record unsealed.17 Judge Sidney H. Stein is weighing cross-motions for summary judgment, and documents keep coming out while he does.1

What the latest unsealing adds

Three new points stand out.

The "horse trading" relationship. The publishers describe the companies' content exchanges using their own label of "horse trading." The latest version says OpenAI gave Microsoft its full GPT-3 training dataset.8 The September version had already claimed that the two companies copied "billions of web pages" for these exchanges instead of licensing the material from news organizations. Project Mango fits that account: Microsoft built and ran the crawler, and OpenAI used the output for training.

How much news people ask chatbots for. The publishers estimate that ChatGPT gets about 47 million news-related prompts a day. They reach that number by taking an OpenAI expert's study, which found that 1.88% of ChatGPT messages in July 2025 sought news, and applying it to 2.5 billion daily prompts.8 This is an extrapolation by the plaintiffs, not a figure OpenAI has confirmed. Still, it is far larger than the internal OpenAI number unsealed in September, which put local-news demand inside ChatGPT at roughly 1 million prompts a week as of February 2026.

Click-through losses across all plaintiffs. The new filing says Microsoft's own data show click-through rates falling 83% to 93% for the Times and the Daily News group, and 51% to 94% for Ziff Davis domains, when Copilot is compared with traditional Bing search.11 The publishers paired those numbers with a quote from Microsoft's lead counsel: chatbots are "designed to answer your question" instead of handing users a list of links. In their reading, that means the chatbots draw readers away from the outlets that produced the copied work.3

Where the coverage agrees, and where the numbers differ

Outlets agree on the basic structure. Cross-motions were filed on September 4. The publishers want liability found at five stages: acquisition, training, Bing-based "grounding" of chatbot answers, Copilot outputs, and the copies the two companies passed to each other.17 The grounding and output claims are aimed only at Microsoft. One footnote ties that to a pending sanctions motion against OpenAI.15 The publishers also want a ruling that statutory damages are counted per article, not as one lump sum.15

The click-through figures are reported differently, though. TheWrap, describing the September version, said the drops were 87% to 93% for the Times and 83% to 91% for the Daily News plaintiffs, and attributed them to Bing Chat.14 The October filing, as quoted, gives 83% to 93% across both groups and attributes the data to Copilot.11 Bing Chat was renamed Copilot, so the gap probably reflects how different exhibits were summarized, not a real contradiction. Readers should still be careful with any single headline number.

Accounts of who the plaintiffs are also vary. The Times reported that eleven other publishers have joined its case.1 Computerworld lists the Daily News, the Center for Investigative Reporting and seven other newspapers.6 Other summaries describe five consolidated cases: the Times, eight Daily News papers, Ziff Davis, the nonprofit behind Mother Jones and Reveal, and The Intercept.15 The differences come down to whether outlets count corporate parents, individual mastheads or separate cases.

The defense, and the government on its side

The publishers' documents are getting attention, but the defendants' position is strong in several ways. OpenAI wants a ruling that pretraining is fair use. Its brief says its expert searched a sample of 20 million ChatGPT logs and found 24 instances of verbatim reproduction.2 Microsoft cites a calculation that matching text of 16 or more words showed up in about 0.73% of 8.2 million Copilot conversation logs.2 OpenAI also points to the June 2025 rulings in the Anthropic and Meta cases, which treated training as transformative. It notes that it has licensing deals with more than 20 publishers.4

The US Justice Department has also stepped in. In a September 1 statement of interest, it argued that training on written works is "exceedingly transformative" and that broad liability requiring licenses would be legally incorrect.7 The statement framed the government's interest in terms of America's global AI dominance and national security. The publishers asked the court to give it "no weight at all." They argue it was filed before the record was unsealed and conflicts with the Copyright Office's May 2025 report.2

Microsoft says the most damaging quotes come from one dissenter. Brent Hecht, its director of applied science, called the copying possibly the "largest theft of labor in human history."13 A Microsoft spokesperson said those remarks reflect one employee's view and are not a legal analysis.14 A declaration from another Microsoft executive described Hecht as someone paid to argue the contrarian side. That explanation fits Hecht less well than it fits other evidence. The publishers also cite OpenAI's head of ChatGPT, Nick Turley, writing that AI products were "largely substitutive,"14 and Satya Nadella testifying that chatbot conversations had replaced visits to publishers' sites.14

Why market substitution is the central issue

The documents matter so much because of the fourth fair-use factor, which asks whether the use harms the market for the original. The publishers' summary judgment strategy depends on showing that ChatGPT and Copilot are direct substitutes that take away traffic and revenue.4 The Anthropic and Meta rulings centered on training. This case puts more weight on output and grounding, the retrieval step in which a chatbot pulls live web content into its answers. That is where Microsoft's click-through data and the "doom loop" memo carry the most force. In that memo, Microsoft warned that its AI content strategy threatened its own "content supply chain."

My view is that the publishers have a stronger case on grounding and outputs than on training alone. The DOJ statement explicitly covers only training.2 The defendants' low verbatim-match rates also speak mostly to regurgitation, not to whether a summarized answer replaces a click. Even if Judge Stein follows the 2025 rulings on pretraining, Microsoft's Copilot claims could still survive. That would be a split result, not a clean win for either side.

The acquisition evidence could matter as well. The publishers say OpenAI trained on the New York Times Annotated Corpus, more than 1.8 million articles, even though its license was limited to non-commercial research.15 They also cite an exchange in which an OpenAI employee told Greg Brockman about "a hack to get around nytimes paywall," and Brockman replied, "ah nice."14 How a work was obtained has become a key question in AI fair-use cases. Evidence like this makes it harder to describe training as a neutral technical step.

What is at stake for companies using AI

The stakes go beyond the newsroom. Statutory damages for willful infringement can reach $150,000 per work.5 By OpenAI's account, the news plaintiffs together assert more than 10.8 million works.2 The Times has also asked for destruction of models and datasets built with its content.6 Analysts at WindowsForum note that even a partial version of that remedy could force model deprecations and API migrations for enterprises building on specific model versions.4 For corporate legal teams and the legal-tech vendors deploying these models, the result in this case will help set how much indemnification and provenance documentation they should require.

The pressure is also coming from more directions. A coalition of nearly 400 local newspapers led by Richner Communications has filed its own suit against the companies.7 The news plaintiffs' sanctions motion is still unresolved. It accuses OpenAI of hiding its ability to search training data and logs and of failing to preserve ChatGPT conversations, which OpenAI denies.17

What comes next

OpenAI and Microsoft must file their oppositions by October 23, and replies are due November 20. The defendants have asked for oral argument, and no trial date has been set.17 None of this is a finding yet. Every quote in the brief was chosen by the plaintiffs, and the judge could deny the motion entirely.15 Still, each round of unsealing has added to the record the publishers are building, and it is clear that some of the companies' own employees anticipated the market harm the publishers now allege.