Microsoft measured a 93 percent click-through drop to nytimes.com from its own answer engine

A reader asks Copilot a question, reads the answer, and never opens the tab. Microsoft's own instrumentation counted the result: for every 100 clicks the traditional Bing results page sent to the Times' domain, the answer engine sent as few as 7 ^[15].

Brent Hecht, Microsoft's Director of Applied Science, wrote in a January 2024 internal presentation that the decline was a "doom loop" that would "hurt the performance of our models and the entire web at the same time" ^[5]. The same document, as quoted in the Times' brief, said: "It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its 'content supply chain'" ^[6].

Every team shipping a summary or answer surface over someone else's content tells itself the same thing: the answer filters the user, so the clicks that survive are worth more. Microsoft counted the clicks. Nick Turley, OpenAI's Head of ChatGPT, wrote in internal communication that products like the chatbot are "largely substitutive" and "will get more and more substitutive as they get better" ^[3].

The volume behind that is now on the docket. One Common Crawl-derived dataset held more than 2 million documents from nytimes.com alone, roughly 22 times the number of copies logged across the three publishers in OpenAI's mid-training data ^[14][16]. The Times also alleges the companies bypassed paywalls undetected and deliberately stripped copyright notices out of training data ^[7].

TechCrunch reported that much of the new material comes from the Times' own brief rather than the exhibits, which remain sealed, and that the quotes appear without their original context ^[8]. Judges have so far been receptive to the argument that training is fair use, and earlier this month the Trump administration filed a brief defending OpenAI's unlicensed use of copyrighted work for training ^[9][10]. Near-term exposure for anyone building on these models is licensing cost and settlement leverage.

Does your surface reduce outbound clicks to the source? Measure that in clicks you can see in your own logs. Then ask whether your model vendor will state in writing that paywalled material in the training and grounding path was licensed. Substitutive plus unwarranted is the pairing that ends up in someone's discovery request. Most retrieval-augmented features shipped this year would fall into it.

Satya Nadella testified in a deposition earlier this year that "anything that is paywalled should be licensed by anyone who wants to use it...for grounding or training" ^[11], and said that had he been made aware OpenAI trained on paywalled information, he would have invoked Microsoft's right to require OpenAI to retrain its models ^[12].