Dead Internet? A Third of New Websites Are AI-Generated, Says Stanford

A caller survey has a fig for however overmuch of the net is present AI-generated: 35%. That's the stock of recently published websites classified arsenic AI-generated oregon AI-assisted by mid-2025, according to probe from Stanford University, Imperial College London, and the Internet Archive. The fig was fundamentally zero earlier ChatGPT launched successful November 2022.

"I find the sheer velocity of the AI takeover of the web rather staggering," Jonáš Doležal, researcher astatine Imperial College London and co-author of the paper, told 404 Media. "After decades of humans shaping it, a important information of the net has go defined by AI successful conscionable 3 years."

The study, titled “The Impact of AI-Generated Text connected the Internet,” drew connected 33 months of website snapshots from the Internet Archive's Wayback Machine and utilized an AI substance detector called Pangram v3 to classify each page.

The confirmed harms: vibes, not facts

Researchers tested six hypotheses astir what AI contented does to the web. Only 2 held up nether information scrutiny.

The first: We’re turning into a horde of dumb NPCs acting successful the aforesaid way… Or much scientifically put, the web is becoming little semantically diverse.

AI-generated sites showed pairwise semantic similarity scores 33% higher than human-written ones. The aforesaid ideas support getting expressed successful astir the aforesaid ways.

The insubstantial suggests the online Overton model whitethorn beryllium narrowing, not done censorship oregon coordinated campaigns, but due to the fact that connection models optimize for outputs adjacent to their grooming distribution.

The second: The web is getting aggressively cheerful.

AI contented showed affirmative sentiment scores much than 107% higher than quality content. Researchers necktie this to the well-documented sycophantic tendencies of LLMs—trained connected quality support signals, they nutrient substance that feels sanitized, friction-free, and relentlessly upbeat.

An net flooded with cheerful, homogenized contented whitethorn marginalize quality dissent astatine standard without anyone pulling a lever.

Despite wide nationalist belief, the survey recovered nary statistically important grounds that AI contented is making the net little factually accurate. Researchers recovered nary meaningful correlation betwixt AI prevalence and factual mistake rate.

The stylistic monoculture hypothesis—AI flattening idiosyncratic voices into a generic azygous register—was the content respondents held astir powerfully (83% agreed). The information didn't corroborate it. Character-level investigation recovered nary statistically important summation successful stylistic homogeneity tied to AI prevalence.

The exemplary illness occupation conscionable got real

The broader stakes spell beyond sermon quality. At 35% AI prevalence, the theoretical hazard of model collapse—where aboriginal models degrade aft grooming connected AI-generated data—shifts from world interest to empirical reality. Future instauration models trained connected modern web crawls volition inevitably ingest information that is substantially AI-generated and measurably little semantically diverse.

The squad is present moving with the Internet Archive to crook the survey into a continuous, unrecorded monitoring tool, tracking AI's stock of the web successful existent clip alternatively than arsenic a one-off snapshot.

A U.S. survey conducted alongside the survey recovered astir Americans already judge each six antagonistic hypotheses, including the ones the information doesn't support. People who usage AI infrequently were 12% much apt to judge successful the harms than predominant users. Dead Internet Theory believers, conscionable the data: The net isn't dead, but 35% of what's caller is astir apt zombie contented successful immoderate way.