Tech Companies Knew Their AI Tools Posed ‘Existential Threat' to Publishers

The documents sat in a folder. Redacted. Waiting.

Then a judge said, fine, open them. And what fell out was not a surprise to anyone who had been watching the numbers. It was a confirmation. A quiet one. The kind that arrives in an email written by a director of applied science at Microsoft in January 2024, addressed to colleagues, describing what the company was doing to the web it depended on.

Brent Hecht called it a “doom loop.” He wrote that the AI content strategy would “hurt the performance of our models and the entire web at the same time.” He noted, almost as an aside, that it is “highly unusual that an end-product threatens the economic foundations of its essential suppliers.”

Highly unusual. That is one way to put it.

Nick Turley, the head of ChatGPT at OpenAI, used different words. He wrote that publishers face an “existential threat” from AI products. That the products are “largely substitutive.” That they “will get more and more substitutive as they get better.”

Substitutive. That is the word that matters. Not competitive. Not adjacent. Substitutive. A reader no longer needs the original thing because the machine has absorbed it and can hand back a version of it on demand.

The court filing came from the New York Times' copyright lawsuit against OpenAI and Microsoft. Filed in December 2023. Consolidated with suits from other publishers. The unsealed portions landed on a Thursday. By Friday, the quotes were everywhere.

The “Modern Newsstand”

OpenAI's internal documents referred to ChatGPT as a “modern newsstand,” according to the filing. A newsstand does not write the news. It displays it. People walk past. They pick up a paper. They pay. The analogy holds up to a point. Beyond that point, it collapses. A newsstand does not read the paper for you and then summarize it back in three paragraphs with no attribution.

The filing also detailed internal discussions about circumventing paywalls. In one exchange, OpenAI President Greg Brockman responded “ah nice” after a researcher described finding a hack around the New York Times' paywall. This is not the language of a company agonizing over fair use. It is the language of a company that has already decided the rules do not apply and is now comparing notes on technique.

Microsoft CEO Satya Nadella testified that “anything that is paywalled should be licensed by anyone who wants to use it” for AI training. He also acknowledged under oath that conversing with chatbots “has substituted” the act of going to the underlying source for information.

So the record now contains this: the people building the tools said the tools would replace the thing they were training on. They knew. They wrote it down. They sent emails. Then they kept going.

The Numbers That Confirm the Emails

You do not need a court filing to see what happened. You need a dashboard.

Microsoft's own data showed that its Copilot “answer engine” caused click-through rates for the New York Times' domain to drop as much as 93% compared to traditional Bing search. Let that sit for a second. Ninety-three percent. Not a dip. Not a seasonal adjustment. A near-total elimination of the pathway between a user and the thing they were searching for.

Globally, Google search traffic to publishers fell 33% in the year to November 2025. Chartbeat data reported page views from Google Search fell 34% across its publisher network between December 2024 and December 2025. Small publishers lost roughly 60% of their search referral traffic over two years. Mid-sized outlets lost about 47%. Large sites lost 22%.

The Reuters Institute surveyed 280 media leaders in 51 countries. They expect search engine referrals to fall another 43% over the coming years.

These are not projections from worried publishers. These are the numbers the platforms' own analytics show. The doom loop Hecht described is not theoretical. It is operating.

The Fair-Use Collision

OpenAI and Microsoft have argued in court that training AI models on copyrighted news content constitutes fair use. The argument rests on several pillars. One of them is that the use is “transformative” — the model does not reproduce the original work but creates something new from it. Another is that the use does not substitute for the original or harm its market.

The internal emails cut directly against that second pillar.

If your own head of product writes that the product is “largely substitutive,” it becomes difficult to argue in front of a judge that the product does not substitute. If your own director of applied science describes a “doom loop” in which the end-product threatens its suppliers' economic foundations, the market-harm question answers itself.

The publishers' filing argues that these quotes “eviscerate” the fair-use defense. That is lawyers being lawyers. But the underlying point is not legalistic. It is plain. The people who built the tools understood what the tools would do to the people who made the raw material. They did not stop. They wrote memos.

A Microsoft spokesperson responded that Hecht's comments “reflect one employee's individual perspective” and do not represent the company's views. Maybe. But the filing also quotes Nadella under oath. And Brockman's “ah nice” was not an individual perspective. It was the president of the company reacting to a paywall bypass.

The Department of Justice filed a statement of interest in the case supporting the AI companies, arguing that preventing training on copyrighted material would “thwart creative and scientific progress.” The Trump administration has taken the same position.

So the legal battle is not really about whether the internal emails exist. It is about whether what they describe is illegal. The publishers say the emails prove substitution and market harm. The AI companies say fair use covers transformative training regardless of internal speculation about downstream effects.

Judge Sidney Stein will decide.

What Publishers Are Doing

Some publishers have signed licensing deals with OpenAI. Others have not. The ones who have signed get paid. The ones who have not get scraped.

The deeper problem is structural. Even if every publisher signed a licensing deal tomorrow, the traffic economics would not revert. A user who asks ChatGPT for a summary of a news article is not going to click through to the article afterward. The summary is the destination. The article was the source material. The payment, if it exists, covers the training. It does not cover the lost subscription. It does not cover the lost ad impression. It does not cover the reader relationship that no longer forms.

Ben Thompson at Stratechery put it bluntly: publishers have been doomed for a long time, and AI just makes the long-term outlook for advertising-based publishers worse. That is not a hopeful sentence. But it is an accurate one.

Some publishers are pivoting to video. Some are building direct relationships through newsletters and podcasts. Some are experimenting with AI tools of their own — using the same technology to produce content faster, cheaper, in more formats. The Reuters Institute found that publishers are ditching SEO as a primary strategy and pivoting toward YouTube and other platforms where the relationship with the audience is more direct.

This may work for some. It will not work for all. The local newspaper that covered city council meetings for forty years does not have a YouTube strategy. It had a search traffic strategy. That strategy is collapsing.

The Thing Nobody Says Out Loud

The internal emails are embarrassing because they are honest. They show that the people closest to the technology understood the consequences before the consequences arrived. They understood that an answer engine is not a search engine. A search engine sends you somewhere. An answer engine sends you nowhere. It answers. You leave. The publisher gets nothing.

Hecht called it “the largest theft of labor in human history.” That is a heavy phrase. But consider what actually happened. Millions of articles. Written by reporters. Edited by editors. Fact-checked. Published. Crawled. Absorbed. And now available as a synthesized answer that does not pay the reporter, does not pay the editor, does not pay the publication.

The AI companies say this is fair use. The publishers say it is theft. The emails show that at least some people inside the AI companies agreed with the publishers — privately, in writing, before the lawsuits.

The court will decide the legal question. The market has already decided the practical one.

Publishers are losing traffic. They are losing revenue. They are losing the reader relationship that made the whole system function. The internal documents confirm that the companies building the tools knew this would happen. They built them anyway.

The emails are now exhibits. The exhibits are now public. And the public can read, in the plain language of a Microsoft director writing to colleagues in January 2024, exactly what the industry was doing to itself and to the web it depended on.

He called it a doom loop.

He was right.