The story. A tranche of internal emails and analysis unsealed September 16 in The New York Times’ lawsuit against OpenAI and Microsoft lays out, in the companies’ own words, that they treated news journalism as training and retrieval fuel for products built to replace the sites that published it. (Digiday, “Paywall violation ‘eviscerates’ fair use defense: What OpenAI-NYT twist means for publishers,” Jessica Davies and Sara Guaglione, 21 September 2026)
01What happened
- OpenAI’s own head of ChatGPT called the substitution risk existential. Nick Turley wrote that publishers face an “existential threat” from OpenAI’s products, which are “largely substitutive, period” and “will get more and more substitutive as they get better.”
- OpenAI’s president said the models were built to be good at exactly this. Greg Brockman wrote that OpenAI’s models are “particularly good at predicting text of news articles,” “excellent at news,” and “very good at any news task.”
- A Microsoft executive’s own words describe the harm as intentional and mutual. The unsealed material quotes an executive saying the companies “started a ‘doom loop’ that will hurt the performance of our models and the entire web at the same time.”
- Microsoft has the click-through data, and it’s stark. Comparing its Bing Search engine to its Copilot chatbot, Microsoft recorded 83-93% drops in click-through rates on The New York Times’ and Daily News’ domains.
- Satya Nadella testified to the substitution effect directly. Microsoft’s CEO said 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.”
- A privacy attorney says the paywall detail changes the legal calculus, not just the optics. Alan Chapell: paywall violations “almost always” defeat a fair-use defense — the evidence isn’t just embarrassing, it’s a specific legal liability distinct from the general scraping argument.
02What it means inside a GAM network
Every publisher weighing whether to license, block, or litigate against an AI platform has been operating on inference — modeled substitution risk, estimated referral-traffic decline, projected AI-answer cannibalization. This unsealed tranche replaces inference with the platforms’ own internal admissions and Microsoft’s own measured numbers. The 83-93% click-through gap is the figure worth internalizing: it’s not a publisher’s own attribution model, it’s Microsoft’s data on its own product, produced in litigation. Chapell’s point about paywall violations “almost always” defeating fair use also narrows the legal question in a way that matters for licensing leverage specifically. A general “was this fair use” argument is a multi-factor balancing test AI platforms have won pieces of elsewhere. Evidence of a mechanism that circumvented an access control — a paywall — to reach the content in the first place is a narrower, more fact-specific claim, and harder for a platform to argue its way around. That distinction is what shifts a publisher’s negotiating position from “we think you’re substituting us” to “you documented that you were.”
Publishers have argued for years that AI platforms were built on their journalism and designed to replace the destinations that produced it.
03What publishers should do about it
04The bottom line
Publishers have argued for years that AI platforms were built on their journalism and designed to replace the destinations that produced it. What changed on September 16 isn’t the argument — it’s that OpenAI’s and Microsoft’s own executives said versions of it in writing, and Microsoft’s own data now puts a number on the substitution effect. Combined with a paywall-circumvention fact pattern a privacy attorney says nearly defeats fair use on its own, this tranche gives publishers something inference never could: the platforms’ own record.