Via niu.edu
Researchers from Imperial College London, Stanford, and the Internet Archive found 35% of new websites are now AI-generated or AI-assisted, up from zero before ChatGPT launched.
The web is being rewritten, quite literally, by machines. A peer-reviewed study from researchers at Imperial College London, Stanford University, and the Internet Archive found that roughly 35% of newly published websites by mid-2025 can be classified as AI-generated or AI-assisted. More than 20% are fully AI-generated, with no meaningful human editing detected.
That figure was effectively zero before ChatGPT’s public launch in late November 2022. In about two and a half years, AI went from writing none of the web to writing more than a third of its new pages.
What the data actually shows #
The research team, led by Jonas Dolezal, Sawood Alam, Mark Graham, and Maty Bohacek, analyzed a stratified sample of websites archived through the Wayback Machine from mid-2022 to mid-2025. They used the Pangram v3 detector to classify content, scanning for linguistic patterns characteristic of large language model output.
Two findings stood out. AI-generated sites showed a 33% increase in semantic similarity compared to human-written pages. The second notable result: positive sentiment scores on AI-generated pages were 107% higher than on human-written ones.
But the study also pushed back against several popular fears. Four commonly cited concerns about AI content, including increased factual inaccuracies, fewer outbound links, less stylistic variety, and lower semantic density in longer texts, showed no statistically significant correlation with AI prevalence.
The perception gap #
The researchers surveyed 853 US adults about their beliefs regarding AI-generated content online. A majority believed in all six negative impacts they were asked about, including the four that the quantitative analysis couldn’t substantiate.
Why verification infrastructure matters #
The researchers described this transformation as one of the most significant shifts in the history of the web. Their recommendations pointed toward two specific interventions: cryptographic verification of human authorship and algorithmic changes that prioritize diverse, human-generated content.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our