News publishers’ brief: Project Mango, ‘largely substitutive’ ChatGPT
Partly confirmed
The takeaway
On Sept 17, 2026 the New York Times, Daily News titles, Ziff Davis, the Center for Investigative Reporting and The Intercept filed a less-redacted version of their summary-judgment brief in the OpenAI/Microsoft copyright MDL (S.D.N.Y., 25-md-3143).
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 17 October 2026
Your AI and this story
- GPT-6 Astra140 days after its cutoff
- Claude Opus 5.579 days after its cutoff
- Gemini 4 Argon170 days after its cutoff
- Grok 4.7109 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 79 days before it.
Key facts
- Filing: public version of the news plaintiffs’ combined summary-judgment brief, Doc. 1977-1, MDL 25-md-3143 (Judge Sidney H. Stein), filed Sept 17, 2026; 92 pages; fewer redactions than the Sept 4 versions (PPC Land)
- Brent Hecht (Microsoft Director of Applied Science), January 2023 memo: people would see models ‘hoovering up’ their work as ‘an astonishing theft of unprecedented proportions’; January 2024: a self-inflicted ‘doom loop’, ‘highly unusual that an end-product threatens the economic foundations of its essential suppliers’ (Techstrong.ai)
- Nick Turley (OpenAI head of ChatGPT): the products ‘are largely substitutive, period’; ‘no good reason to click’ once browsing gave an answer. An unnamed OpenAI engineer: ‘no matter how prominently we show the links, users won’t click’
- Satya Nadella (deposition): agreed chatbots substitute for ‘giving you the information right there on the website’; ‘anything that is paywalled should be licensed by anyone who wants to use it’; said he would have required retraining had he known paywalled material was used without permission (Techstrong.ai)
- Microsoft data cited: Copilot produced 87–93% fewer referral clicks than Bing for Times sites, 83–91% for Daily News sites, 51–94% for Ziff Davis sites
- Project Taxi (2019–2022): Microsoft gave OpenAI a copy of the Bing index (price redacted); Bob McGrew (Oct 2021): ‘we’re giving data to them and they are giving data to us’
- Project Mango: Microsoft ran the ‘Mango crawler’ for OpenAI, which paid a redacted sum; the dataset holds at least 160,903 unique plaintiff works and allegedly trained models including GPT-4o; the brief says copyright notices were removed before training
- Paywall: OpenAI researcher Nick Ryder allegedly told Greg Brockman of a ‘hack to get around [the] nytimes paywall’; Brockman replied ‘ah nice’ (Techstrong.ai)
Show 3 more
- Counts claimed: Times 3,924,653 training copies (944,655 unique works), Daily News 7,387,394 (1,388,384 unique); nytimes.com 2,064,805 articles in a filtered Common Crawl top-100 set
- Oct 16, 2026 (Bloomberg Law): part of the publishers’ motion was unsealed further: Microsoft ordered employees to keep the scraping project ‘secret’; OpenAI gave Microsoft its full GPT-3 training set; the filing calls the companies’ deals ‘horse trading’; the publishers estimate ChatGPT gets 47M news-seeking prompts a day (1.88% of July 2025 messages × 2.5B daily prompts); Hecht said copying millions of news articles could be the ‘largest theft of labor in human history’. Payment amounts, page counts, licensing-deal figures and some Slack messages remain sealed; Microsoft withdrew some redaction requests on Oct 9 and Bloomberg Industry Group moved to intervene against sealing
- Defence: OpenAI and Microsoft argue training on public web content is transformative fair use (cross-motions filed Sept 2026); OpenAI says its models are ‘trained on publicly available data’
What happened
The New York Times’ 2023 suit is part of the multidistrict litigation that also holds the authors’ cases (see the LibGen entry). In September 2026 both sides moved for summary judgment on whether training was fair use. The publishers’ brief, made public in a less redacted form on Sept 17, leans on internal documents to argue that ChatGPT and Copilot substitute for news sites (the fourth fair-use factor, market harm). On Oct 16 more of the motion was unsealed after publishers and Bloomberg Industry Group objected to redactions.
Gap: the dataset missed the Sept 17 filing. It was found on Oct 17 through Bloomberg Law’s report on the second unsealing (sweep, Google News). This entry is written from PPC Land’s, Techstrong.ai’s and Bloomberg Law’s accounts; the brief itself was not read.
Why it matters
These are the most detailed public claims about how OpenAI got news data: a Microsoft-run crawler, a Bing-index exchange and Common Crawl. The quotes from Microsoft’s own scientist and from ChatGPT’s head on substitution go to the market-harm factor that decided earlier AI fair-use rulings. The court has not ruled on any of it.
Sources
3 sources from 3 sites. Numbers match the chips in the text.
3 sources: 3 press
Press
- PPC Land: Copilot cut publisher click-through rates up to 94%, internal Bing data shows (Sept 20; brief details)ppc.land, press
- Techstrong.ai: Unredacted court filings reveal tech executives privately termed AI data scraping ‘theft’ (Sept 18)techstrong.ai, press
- Bloomberg Law: Microsoft told staff to keep OpenAI scraping project ‘secret’ (Oct 16)news.bloomberglaw.com, press
Changes
- Filed as a review, found through Bloomberg Law’s Oct 16 report