AI Signs Appear on 35% of Post-ChatGPT Webpages, Pew Finds
Pew found signs of AI writing or heavy editing on 35% of English webpages published after ChatGPT launched, with .com sites far ahead of .edu and .gov.
Summary
On August 20, 2026, Pew Research reported that 35% of English-language webpages published after ChatGPT’s November 2022 launch showed significant signs of AI authorship or substantial editing. Pew collected nearly 500,000 pages spanning roughly five years from the Common Crawl archive, including about two pre-ChatGPT years, then analyzed them using Open Pangram’s detection technology. In a random 10,000-page sample collected in July 2026, 10% showed significant signs, but that snapshot included older pages created before AI writing tools existed.
After excluding prelaunch pages, the rate reached 35%. .com URLs showed signs at roughly 10 times the rate for .edu and .gov, each around 1%, while .org reached 4.6%. Pew also tracked increased use of purported markers including em dashes, Oxford commas and “it’s not X, it’s Y” phrasing. It cautioned that Pangram can falsely classify human-written pages, although the large-scale pattern is likely directionally accurate. The findings support other studies and follow Cloudflare’s report that bot traffic overtook human traffic earlier than expected, leaving bots increasingly likely to browse content created by bots.
Positives
- Nearly 500,000 English-language pages gave Pew a dataset extending about two years before ChatGPT’s November 2022 launch.
- Filtering out pre-ChatGPT pages prevented older material from diluting the estimate for content published after generative AI became widely available.
- .edu and .gov URLs each showed only around 1% AI authorship or substantial editing.
- .org URLs recorded a 4.6% rate, remaining below the prevalence Pew observed for .com domains.
Risks & concerns
- Thirty-five percent of webpages published after ChatGPT’s launch showed significant signs of AI authorship or substantial editing.
- .com URLs displayed AI authorship signs at roughly 10 times the rate recorded for .edu and .gov domains.
- Cloudflare reported that bot traffic overtook human web traffic sooner than the infrastructure provider had estimated.
- Open Pangram and other AI detectors can falsely classify human-written pages, limiting certainty about individual results.
- Em dashes, Oxford commas and “it’s not X, it’s Y” constructions increased despite being imperfect indicators of AI authorship.