AI text is not the same on all web pages, a study of 5 million finds

Researchers made a study of a sample of more than 5 million web pages that a tool can read. The share of pages with AI text increased quickly in less than 5 years, but it is not the same in all areas. English pages have more AI text than pages in other languages, and .gov and .edu domains have less than .com domains.

How it works

The authors use some AI detection tools. A Bayesian hierarchical model (a math method) corrects for the errors of the tools. It also lets the skill of a tool be different for each group.

The findings

• The results stay the same after the authors adjust for the type of pages.
• AI text is most frequent on the pages that a small number of pages connect to, and it increases most quickly there.
• In the tiers of pages that most other pages connect to, AI text is less frequent than in all other tiers.

What is not known

The text gives no numbers for the shares. The detection tools are not always correct, and the results come from 1 sample.

Insight: AI text grows at the edge of the web

The study finds the most AI text on pages that a small number of pages connect to. Other researchers found that 9 frontier models write text that is not easy to read. The AI text reads as volume at the edge of the web, and thus one average share does not show where it is. Measure the AI share for each source in your data.