Pew Finds AI Authorship Signs in 10% of Webpages

Summary

Pew Research Center analyzed nearly 490,000 English-language webpages collected from Common Crawl snapshots between January 2021 and July 2026. Using the open-weight Open Pangram detection model, researchers looked for statistical patterns associated with AI authorship or substantial AI editing. In the July 2026 snapshot, 10% of all sampled pages showed significant signs of AI authorship, while the share exceeded one-third among pages published after ChatGPT’s release. The pattern varied sharply by domain: about 10% of .com pages showed such signs, compared with 4.6% of .org pages and roughly 1% of both .edu and .gov pages. The study also found that AI-associated language features became more common on newer pages: em dashes appeared about twice as often as in a 2023 comparison, Oxford commas increased 63%, selected AI-associated vocabulary more than doubled, and negative-parallelism constructions nearly tripled. Pew cautions that individual signals do not prove AI authorship and that detection models can misclassify human and AI-written text. The findings describe statistical signs across large collections, rather than a definitive authorship judgment for every webpage.