AI Slop Found in Over a Third of New Web Pages
A new study by the Pew Research Center estimates that over a third of sampled English webpages published after ChatGPT’s November 2022 launch contain signs of AI-generated material or substantial AI editing.
Researchers took 490,000 English web pages from snapshots taken between January 2021 and July 2026. In July 2026, roughly one in every ten pages of any age showed signs of AI authorship or editing. For pages published after ChatGPT launched in November 2022, that number increased to 35%. Interestingly, about 10% of .com pages showed evidence of AI authorship, 4.6% of .org pages did, and only about 1% of .edu and .gov pages did.
Pew used Pangram’s open-weight AI detection model, editlens_Llama-3.2-3B, to score the pages. This “AI test” estimates whether AI created, rewrote, translated, or edited, but can’t clearly tell which of these actions were taken on a specific page. The researchers considered a page to have AI authorship if it received a score of 0.2 or more. (Pangram’s model places fully human-written text at 0 and fully AI-generated text at 1.)
A study like this has a couple of limitations. The 35% estimate applies only to pages with publication dates that researchers could identify in their web code, and those dated pages only made up about 10% to 15% of the full sample. AI detection models like Pangram’s can also make mistakes; after all, humans were the first to use commonly cited AI “tells” like em-dashes, Oxford commas, and the “rule of threes.”
Nevertheless, Pew’s deep dive undeniably demonstrates that automation is on its way to dominating the web.
