Internal OpenAI and Microsoft documents admitted what critics have been saying out loud for three years: scraping at scale is destroying the web the models depend on.
404 Media published on September 17 the report that names the phenomenon: doom loop. The internal documentation, tied to the New York Times litigation and the Google Zero context, records executives recognizing that millions of people will soon consider large models "an astonishing theft" of other people’s work. The Verge and Futurism documented the same admission the next day, with a Microsoft director acknowledging the plunder with nothing given back.
The loop, drawn
The mechanics are simple and closed: models trained on web content generate more content, which returns to the web, which trains the next generation. Each turn, the original human fraction shrinks and the synthesis-of-synthesis fraction grows.
The outcome predicted in their own documentation: the web empties of human work exactly for the models that need it to avoid degenerating. It is model collapse at internet scale: the input running out in the middle of the factory that consumed it.
The admission changes the public game
Accused companies spent years answering with the same line: we train on public data, it is fair use, the courts will decide. What 404 Media revealed is that internally the diagnosis was different: executives knew the public would read "hoovering up" as theft, and that the loop would damage the ecosystem.
That changes the game three ways: the "public data is free" narrative lost its owner; the litigation (NYT at the front) gained the argument that not even the developers believed the generous reading; and the content market gained the value ruler it was missing.
For publishers: the original is the scarce good
The practical consequence for companies is the value inversion: in an increasingly synthetic web, original human content with authorship, sources, and method becomes a scarce asset.
Four moves the doom loop rewards:
- Original at your address: your own analysis lives on the company’s site; social media is distribution, not residence.
- Marked authorship: signature, dates, sources, and visible method, what separates human work from model output.
- Declared rights: robots.txt, terms, and licenses stating what third-party AI may do, the legal base the ongoing litigation is pricing.
- AI as assistant, never as source: use the model to accelerate human work, never to replace the input that gives the piece its value.
Conclusion
The doom loop is the first AI crisis the AI admitted in internal documents: the input is running out, and the input is us.
For publishers, the read is not panic, it is pricing: original is worth more today than when the web was people writing for people without machine competition. October’s question is how many original pieces your company published this year, and at whose address they live. If the answer is "on the others’ platforms", the loop is already eating yours.