A third of web pages published since ChatGPT launched show signs of AI authorship
Pew put the figure at 35% for pages published after November 2022, and flagged that detection tools misclassify human writing often enough to treat the number as directional.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Pew Research Center found that about 35% of web pages published after ChatGPT's November 2022 launch show signs of AI authorship or heavy AI editing.
- The analysis drew on nearly 500,000 English-language pages collected over five years, with the headline figure from a 10,000-page random sample taken in July 2026.
- Rates vary sharply by domain: .com pages showed AI authorship at roughly ten times the rate of .edu and .gov pages, which sat near 1%; .org came in at 4.6%.
- Pew cautioned that AI-detection tools can misclassify human-written pages, and argues the result is directionally correct at scale rather than precise.
Pew Research Center has put a number on how much of the web now carries machine fingerprints: roughly 35% of English-language pages published after ChatGPT's November 2022 launch show signs of AI authorship or substantial AI editing. The figure comes from a corpus of nearly 500,000 pages collected over five years, with the headline result drawn from a random 10,000-page sample taken in July 2026 and detection performed using Open Pangram's classifier.
The distribution matters more than the headline. Commercial .com pages showed AI authorship at approximately ten times the rate of .edu and .gov pages, which sat near 1%, with .org domains at 4.6%. Restricting the sample to pages published before ChatGPT existed and running the same detector produced a rate near 10%, which is a useful indication of the method's noise floor rather than evidence of pre-2022 AI writing.
Pew was direct about the limitation. Detection tools misclassify human-written text as machine-written, and no classifier reports its own error rate reliably at web scale, so the centre frames the finding as directionally correct rather than precise. The researchers also noted a rise in stylistic markers now associated with model output, including em dashes and the "it's not X, it's Y" construction — signals that are suggestive but that a human writer may legitimately produce.
If a third of recent pages are machine-written, then a third of what the next model scrapes is machine-written too, and the .com-versus-.gov gap says the contamination is concentrated exactly where general crawls collect most heavily. For anyone assembling a training corpus or a retrieval index, provenance filtering stops being a nice-to-have. The caveat cuts the other way as well: a 10% false-positive floor on pre-2022 pages means any downstream tool that penalises writers for looking AI-generated is wrong often enough to be unfair.