Pew put half a million web pages through a detector: a third of everything published since ChatGPT carries its marks
The Pew Research Center pulled roughly 490,000 English-language web pages out of the Common Crawl archive, covering the past five years, and ran them through a detection tool called Open Pangram. In a random sample of 10,000 pages captured in July 2026, one in ten showed significant signs of having been written with AI help. Narrow the sample to pages that first appeared after ChatGPT's release in late 2022, and the figure passes one third. The curve starts bending upward exactly when ChatGPT launched and keeps rising as Claude, Gemini and the rest arrive. Pew is careful about what it is claiming: the detector flags signs that a machine had a hand in the text, not that a page was written entirely by one.
Why it mattersUntil now the scale of machine-written text on the open web has been guesswork, argued from small samples or from vendors selling detection. This is a large, documented count from an organisation with no product to sell, and it puts a number on something that affects everyone who reads or searches: a growing share of what a search engine returns was drafted with the same tools that answer your questions. It also matters for the models themselves, which are trained on this same web — each generation now learns from a corpus its predecessors helped write.
Read in the app — free, in 9 languages
Related stories
Machine learning read the shape of sick brain cells and picked out nine already-approved drugs that calmed them down2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21The agent invented a second person to vouch for its code. A 24-year-old in Texas refused to believe either of them.
2026-08-21The FDA has cleared 1,357 AI medical devices. Three of them have been tested on whether patients live longer or better.
2026-08-20So many machine-written proofs arrived that mathematicians built a filing office — and one of its two checks is done by a language model
2026-08-20