In three years, AI has created a third of the new websites

We have been promised for months a "dead internet," overrun by robots talking to other robots. Researchers decided to check. A team from Stanford, Imperial College London, and the Internet Archive has just published a study titled The Impact of AI-Generated Text on the Internet. It is the largest ever conducted on this topic. His main figure: 35 % of websites created since the end of 2022 would be generated or assisted by AI, according to 404 Media who revealed the results. From zero to a third in three years. The "theory of the dead internet" has just received its first empirical validation.

Also read: Internet Archive and its Wayback Machine are in danger due to AI

How researchers measured the contamination

The team exploited the archives of the Wayback Machine to extract samples of websites published between August 2022 and May 2025. That is, 33 months of web production scrutinized. The raw HTML of each page was downloaded, cleaned, and then analyzed by Pangram v3, an AI text detector judged the most reliable after comparison with three competitors. Once the websites were classified as "AI" or "human," the researchers tested six common hypotheses about the effects of synthetic content.

Of the six, only two were confirmed. AI makes the web less semantically diverse: same phrases, same structures, same smoothed vocabulary from one site to another. And more positive in tone. The AI-generated web smiles permanently, as it were. On the other hand, no measurable increase in factually false statements. No more verifiable lies, no fewer links to sources. The AI web is not more deceptive than the human web (which may say as much about us as about the machines).

35% or 74%: why the numbers differ

The 35% figure is not the only one circulating. Another study, conducted on nearly a million web pages in 2025, estimated that 74.2% of new pages contained AI-generated content. The discrepancy is explained by the scope. Stanford measures entire sites classified as mostly AI. Ahrefs counts individual pages containing as little as a trace of assisted text. It's the difference between "this house was built by a robot" and "a robot laid at least one brick." Both numbers are true, but they tell different stories.

For researchers, the most concerning aspect is not the quantity. It's the stylistic convergence. A third of the new web is written by the same language models, trained on the same data. Models calibrated to be "perfectly compliant and pleasant," according to the words of one of the study's co-authors. The risk is not misinformation. It's the monoculture. A web where all voices sound alike and where rough edges disappear in favor of mass-produced consensus.

👉🏻 Follow tech news in real-time: add 01net to your Google sources, and subscribe to our WhatsApp channel.

Source : 404 Media