AI answers leave sites without referrals: Copilot cut NYT clicks by 93%

He said this not in an interview but at an internal presentation in January 2024. The slides emerged on 17.09.2026, when the court unsealed briefs in the consolidated case brought by publishers and writers against OpenAI and Microsoft.

The language came not from AI critics but from inside the companies themselves. An internal Microsoft document from 2023 includes this line: millions of people will soon see models sucking up all their work as an astonishing theft on an unprecedented scale.

Why they called it a loop

Doom loop. This is what Hecht called the loop Microsoft’s AI content strategy had set in motion: it hits both the quality of its own models and the entire web. In the same presentation, he called large models a product that destroys its supply chain.

Minus 93%. A standard Bing result sent a click to the site that wrote the text. Copilot answers cut visits to nytimes.com by 93%, and Microsoft itself made that measurement.

How much other people’s text they found in the data

In OpenAI training data, they found more than 91 thousand copies of material from The New York Times, Daily News and the Center for Investigative Reporting. A separate Common Crawl-based set contained more than 2 mln documents from nytimes.com alone.

Microsoft supplied the data to OpenAI through the Taxi and Mango projects. The Mango dataset contained more than 160 thousand individual works by news publishers.

Nick Turley, head of ChatGPT at OpenAI, in internal correspondence described chatbots as largely a replacement for publishers and wrote that publishers faced an existential threat.

Publicly, Microsoft distanced itself from its employees’ remarks. The company maintains that copyright permits model training as transformative use.

The briefs have been submitted at the stage where the court can decide the case without a jury.