More Than a Third of Web Pages Published Since ChatGPT Show Signs of AI Writing

More than a third of English-language web pages published since ChatGPT's release show signs of AI authorship, according to an analysis the Pew Research Center published Aug. 20 that ran a detection model over nearly 490,000 pages taken from the Common Crawl web archive.

Across the whole July 2026 sample, including pages too old to have been written by a chatbot, Pew found that 10% showed significant signs of AI authorship. The pages were drawn from Common Crawl snapshots collected between January 2021 and July 2026 and classified with Open Pangram, an open-weight detection model built by a company called Pangram, which looks for word and sentence patterns that appear more often in machine-generated text.

Pew is explicit that the figure is a detector's estimate. "AI detection models aren't perfect," Pew writes, and "sometimes misclassify individual documents that were written by humans as including signs of AI authorship, and vice versa"; the report argues the patterns hold across very large collections of text rather than for any single page. Pew's measure also counts pages likely written or substantially edited by AI, so a human draft heavily revised with a chatbot falls inside the number.

The signals are not spread evenly. In the 2026 samples, about one in 10 .com pages showed signs of AI authorship, against 4.6% of .org pages and about 1% of .edu and .gov pages. In the January 2021 samples, all four categories sat at or below about 1%.

Comparing today's web with a 2023 snapshot, em dashes were about twice as frequent, Oxford commas increased by 63%, and words grouped as AI-typical (such as "delve," "interplay" and "testament") more than doubled. Constructions of the form "it's not just X, it's Y" nearly tripled, though they remain rare overall.

The analysis was written by Samuel Bestvater, a senior data scientist at Pew, with four other listed authors.