They drew the loop

Eighty-three to ninety-three percent. That is how far click-through rates fell for some of the news sites suing OpenAI and Microsoft, and the number did not come from the plaintiffs. It came from Microsoft. On Thursday the New York Times and its co-plaintiffs unsealed their motion for summary judgment, and Ars Technica's Ashley Belanger pulled the internal documents out of it.

The line that will travel is Brent Hecht's. A Microsoft director of applied science wrote that scraping news for training was "an astonishing theft of unprecedented proportions," possibly "the largest theft of labor in human history," and that the plan made "a complete mockery of the idea of 'fair use.'" Microsoft had a disclaimer out within hours: one employee's perspective, not a legal analysis, not the company's view. Fair enough. It is also the wrong document to read.

The one that matters has no author to disclaim. It is a Microsoft memo about what it calls the "content supply chain," and it reads like a risk register, not a conscience.

The same memo names a "doom loop" that "will hurt the performance of our models and the entire web at the same time." Microsoft drew it as a cartoon. So the companies did not miss the harm. They modeled it, as a threat to their own inputs, and they modeled why they would keep going anyway. The plaintiffs' motion states the logic plainly: the industry as a whole would benefit if every company paid to sustain the works its technology depends on, but each individual company is better off taking content for free while the others pay. That is a prisoner's dilemma, described from inside the prison.

Read that way, the last two years of licensing deals look different. OpenAI signed News Corp, Axel Springer, and the Associated Press. Microsoft signed its own. Those were cooperate moves in a game the memo says no single player can afford to play alone, which is why they changed the traffic numbers not at all. Nick Turley, who runs ChatGPT, wrote that chatbots are "largely substitutive, period" and "will get more and more substitutive as they get better." An OpenAI engineer put it shorter: "no matter how prominently we show the links, users won't click." Substitution is not a defect they were working to patch. It is the product curve.

That is a prisoner's dilemma, described from inside the prison.

Microsoft's real defense is better than its disclaimer, and it deserves the honest version. Fair use is decided on four statutory factors, not on what a director typed into a memo. Intent is not one of the factors. A court could find every quote in this motion sincere and still rule that training is transformative. All true. But the fourth factor is the effect on the market for the original work, and the evidence on that factor here is not anybody's opinion. It is Microsoft's own click-through measurements, 51 to 94 percent declines across the plaintiff set, and OpenAI's own forecast that the decline steepens as the model improves. The memos' job in this motion is not to prove bad faith. It is to show that the defendants measured the substitution themselves, before the plaintiffs did. When the defendant's dashboard and the plaintiff's complaint report the same number, that factor stops being contested.

Greg Brockman, told that a staffer had found "a hack" to get OpenAI's crawlers past the Times paywall, replied "Ah, nice." That is the quote for the jury. The supply-chain memo is the quote for the judge, because it concedes the thing the whole case is about: nobody in the industry can stop free-riding unless everyone has to. The plaintiffs are asking for exactly that, a rule that binds every lab at once, and they are quoting Microsoft's own analysis to explain why nothing weaker will work.

Microsoft drew the loop as a cartoon. The motion asks a judge to draw it as law.