Microsoft built a secret scraper for OpenAI's AI models1. It copied webpages by the hundred thousand1. Microsoft named the project Project Mango1. The company told its own staff to keep it quiet1. A court filing unsealed Oct. 16 names the project and the order1. Buried in an earlier filing sits a line from Microsoft's own scientist, who called the practice "the largest theft of labor in human history"2. The government disagrees. Its own lawyers call the same copying fair use5.
Project Mango's Secret
Copyright law lets a court excuse copying as fair use when the new work transforms the original into something new. Microsoft built Project Mango to copy webpages at scale1. OpenAI paid for the work, and the two companies fed the harvest into models including GPT-4o1. Microsoft ordered its own staff to keep the project quiet1. The order stayed hidden until a federal judge unsealed it this month1. Secrecy inside one company is routine. An order that survives a federal lawsuit's discovery process becomes evidence in its own right1. It records what the company expected the public's reaction to be1.
Brent Hecht holds the title director of applied science at Microsoft2. He wrote an internal memo now cited in The New York Times's lawsuit2. Training data firms "hoovered up" copyrighted work, Hecht wrote, and millions of people would soon call that "an astonishing theft of unprecedented proportions"3. A court win for Microsoft and OpenAI, he added, would "make a complete mockery of the idea of 'fair use'"3. Hecht wrote those words in 2023, years before the case reached a judge2. He kept his job at Microsoft after writing them, and the company still cites his broader research record on its own site2.
OpenAI already pays for some of the same kind of material it stands accused of taking for free. The company has signed more than 20 publisher licensing deals, covering over 160 outlets8. News Corp holds the largest, a reported $250 million over five years8. Those deals buy OpenAI permission to train on an archive and to show current stories inside ChatGPT, with credit and a link8. The Times and the other plaintiffs skipped that table entirely and went to court instead3.
OpenAI signed 20 publisher deals, and still answers to 5 groups in court
Number of publisher deals OpenAI has signed versus the number of publisher groups suing OpenAI and Microsoft over training, as of Oct. 2026
The first column counts the publisher deals OpenAI has signed; the second counts the plaintiff groups pressing the summary judgment motion.
The 20 figure counts publisher deals tallied by LLM Pulse's running tracker, short of a full survey of every publisher OpenAI has approached.
Sources: MIXED Reality News [3]; LLM Pulse [8]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Value |
|---|---|
| Signed a deal | 20 |
| Suing in court | 5 |
What Discovery Turned Up
The scraped set holds 160,903 works from news publishers, court filings show1. A separate dataset, built from Common Crawl, holds more than 2 million documents from nytimes.com alone3. Publishers estimate ChatGPT fields 47 million prompts a day that ask for news content1.
A separate Microsoft document, cited in the same brief, warned that Copilot had started a "doom loop" for the sites it answers questions from9. The document put the mechanism in six words: "LLMs are a product that destroys its supply chain"9. Microsoft's own data showed why. Click-through rates to the Times fell 87 to 93 percent9. Daily News group sites fell 83 to 91 percent, and Ziff Davis titles fell 51 to 94 percent9. Each figure compares a Copilot answer against an ordinary Bing search result9.
News sites lose up to 94 percent of their clicks when Copilot replaces a Bing search
Click-through rate decline, Copilot answers versus ordinary Bing search results, Microsoft's own data cited in court filings
Each bar is one publisher group's click-through rate decline, plotted at the midpoint of Microsoft's reported range.
Each bar plots the midpoint of the range Microsoft's own document reported; the underlying spread runs wider for Ziff Davis than for the other two groups.
Source: The Next Web [9]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| The Times | 90% | Range: 87 to 93 percent |
| Daily News group | 87% | Range: 83 to 91 percent |
| Ziff Davis | 72.5% | Range: 51 to 94 percent |
Nick Turley runs ChatGPT at OpenAI3. He wrote that publishers face an "existential threat" from the product he runs3. Turley called OpenAI's answers "largely substitutive, period"3. Satya Nadella testified under oath too4. Microsoft's own chief executive told the court that a chatbot conversation stands in for a visit to the page underneath it4. Both men described their own products in the words a copyright plaintiff needs most.
The Publishers' Price Tag
David Buttle founded the publisher coalition Spur4. He reads the scraped number as a receipt that carries real money owed. "The defendants' own conduct establishes that the content has a market price," Buttle said. "The case is about their refusal to pay it to these plaintiffs"4.
Danielle Coffey runs the News/Media Alliance4. She expects Turley's own words to force the issue industry-wide. "Once that's answered, I think it will be an across-the-board acknowledgement from the AI companies that they do need to come to the table," Coffey said4.
Alan Chapell, a privacy consultant who tracks AI licensing fights, reads Turley's sentence as a confession on the record. "The apparent admission by the head of ChatGPT that their products are substitutive will be difficult to walk back, and likely eviscerates the fair use defense," Chapell said4.
Ian Crosby leads the Times's legal team7. He argues Microsoft and OpenAI's own conduct gives away their case. "If OpenAI genuinely believed that copying our clients' journalism was fair and legal, it wouldn't have hid the truth about having done it," Crosby said7.
Washington Takes the Other Side
The U.S. Department of Justice backs the opposite reading5. Its September brief stated: "Consistent with the national interest in promoting innovation and free expression, the training of AI models on copyrighted material, in and of itself, does not violate copyright laws"5. The department called the training "exceedingly transformative," because a model processes an article for patterns, a different use than a reader seeking the story6. Fair use law's fourth factor asks whether a copy substitutes for the market of the original. That single question now carries the whole case6.
Summary judgment lets a judge decide a case on the papers alone, skipping a trial, when both sides already agree on the facts. Opposition briefs from OpenAI and Microsoft land Oct. 231.
The department's brief reaches past this one case too. It warned that a ruling against AI training could threaten national security, slowing American labs while rivals abroad train under looser rules6. That argument treats national speed as the stake, a bigger number than any single publisher's lost subscription revenue6. The brief stays advisory, outside the judge's binding law. A federal agency's view still shapes how a court reads a hard legal question56.
What Each Side Needs to Be True
The two sides need opposite readings of the same memos25. Publishers need the court to treat Hecht's and Turley's own words as admissions. Those admissions would carry real legal weight. If the court agrees, the fourth fair-use factor tips toward the plaintiffs. AI companies that trained on news then face a bill set by the market's own price. That price runs far above a licensing deal struck after the fact.
OpenAI and Microsoft need the court to treat transformative use as the only test that counts. The memos, in that reading, show two employees thinking out loud. If the court agrees, the scale of the copying stops mattering. The reach of a model that answers any question outweighs the lost clicks underneath it.
By the numbers
- 160,903 works: copied from news publishers for Project Mango's training set, court filings show1.
- 2,000,000 documents: pulled from nytimes.com alone in a separate Common Crawl-derived set3.
- 47 million prompts: the daily ChatGPT queries publishers estimate ask for news content1.
- Five publisher groups joined the suit, from The New York Times to the nonprofit newsroom behind Mother Jones and Reveal3.
- Sept. 17, 2026: publishers filed their 92-page summary judgment brief that date3.
- Oct. 16 to Oct. 23, 2026: the window from the order's unsealing to the opposition deadline1.
What to watch
Watch the Southern District of New York for a ruling on summary judgment1. Opposition briefs land Oct. 23, and the court's reading of the fourth fair-use factor will show which side's evidence carried the day16. Watch whether other publishers press Buttle's math toward a licensing price, the path that skips a verdict entirely4. A ruling either way will set the price AI companies pay for the news industry's work, in this case or the next one15.
Sources
- Annelise Levy, "Microsoft Told Staff to Keep OpenAI Scraping Project 'Secret'," Bloomberg Law, Oct. 16, 2026, https://news.bloomberglaw.com/ip-law/microsoft-told-staff-to-keep-openai-scraping-project-secret
- Rebecca Bellan, "Microsoft exec called AI scraping 'the largest theft of labor in human history,' new unredacted filings reveal," TechCrunch, Sept. 17, 2026, https://techcrunch.com/2026/09/17/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history-new-unredacted-filings-reveal/
- Michael Zelich, "News publishers ask a judge to reject OpenAI's fair use defence, quoting Microsoft's staff," MIXED Reality News, Sept. 21, 2026, https://mixed-news.com/en/news-publishers-summary-judgment-openai-microsoft-fair-use/
- Jessica Davies and Sara Guaglione, "Paywall violation 'eviscerates' fair use defense: What OpenAI-NYT twist means for publishers," Digiday, Sept. 21, 2026, https://digiday.com/media/paywall-violation-eviscerates-fair-use-defense-what-openai-nyt-twist-means-for-publishers/
- Wendy Davis, "Trump Admin To Court: AI Training Is 'Fair Use'," MediaPost, Sept. 2, 2026, https://www.mediapost.com/publications/article/417630/trump-admin-to-judge-ai-training-is-fair-use.html
- Eileen McDermott, "DOJ Sides with OpenAI, Warns Obstacles to AI Development Threaten National Security," IPWatchdog, Sept. 3, 2026, https://ipwatchdog.com/2026/09/03/doj-sides-with-openai-warns-obstacles-to-ai-development-threaten-national-security/
- Rebecca Bellan, "New York Times says OpenAI hid evidence in ChatGPT copyright trial," TechCrunch, July 9, 2026, https://techcrunch.com/2026/07/09/new-york-times-says-openai-hid-evidence-in-chatgpt-copyright-trial/
- LLM Pulse, "OpenAI Publisher Deals: The Complete Map of Every Media Partnership (2023-2026)," LLM Pulse, Oct. 12, 2026, https://llmpulse.ai/blog/openai-publisher-deals/
- Ana Maria Constantin, "A Microsoft scientist called AI training the largest theft of labor," The Next Web, Sept. 21, 2026, https://thenextweb.com/news/microsoft-theft-of-labor-doom-loop-unsealed-brief-openai