More than one-third of English language webpages published since ChatGPT’s launch show signs of being written or substantially edited using AI, according to a new study.

The Pew Research Center analysed about 490,000 webpages collected through the Common Crawl web archive between January 2021 and July 2026.

Researchers used the Open Pangram detection model to identify linguistic patterns associated with AI-generated text.

Across a random sample of 10,000 webpages collected in July, 10% displayed significant signs of AI authorship. However, that figure climbed to 35% when researchers considered only pages published after ChatGPT was publicly launched in November 2022.

The findings indicate that tools including ChatGPT, Google Gemini and Anthropic’s Claude are having an increasingly visible impact on the content available online.

AI authorship was most prevalent across commercial websites. Around one in 10 pages using a .com domain displayed AI signals, compared with 4.6% of .org pages and about 1% of .edu and .gov pages.

Before ChatGPT’s launch, detection rates across all four domain types were 1% or lower.

Pew also found that writing traits commonly associated with generative AI have become more frequent. The use of em dashes has roughly doubled since early 2023, while Oxford commas increased 63%.

Words frequently favoured by AI models, including “delve”, “interplay” and “testament”, more than doubled in usage.

However, Pew said these stylistic features do not prove an individual article was produced by AI. Humans also use the same language, while detection systems can misclassify both human and machine-generated material.

The study’s threshold also captures pages substantially edited with AI, meaning it does not distinguish between entirely machine-generated content and human writing polished using tools now built into platforms such as Microsoft Word and Google Docs.