Pew Research Center ran nearly half a million webpages through an AI detector and found signs of AI authorship or editing on about 10% of them. Among pages published after ChatGPT's launch, that share jumps to 35%. The analysis, published August 20 by Pew's Data Labs team, lands less than a year after two other major estimates of AI-generated text on the web.
For writers, the takeaway is straightforward: AI-written content is now a measurable part of the web's commercial core, and the tools you use to edit are part of the reason detectors can't tell full automation from light cleanup.
What the data shows
Pages on .com domains show signs of AI authorship at about 10 times the rate of pages on .edu and .gov domains. In a six-month average, the detection rate was 9.35% on .com, against 4.59% on .org, 1.03% on .edu, and 0.76% on .gov. All four domain types sat at or below 1% in the sample collected before ChatGPT launched. They have separated since, with .com climbing fastest while .edu and .gov stayed close to the 1% mark.
Pew examined specific markers in pages published after ChatGPT's release. Em dashes increased from 5.79 uses per 10,000 words in early 2023 to 11.19 in early 2026. Oxford commas saw a 63% rise. The use of words common to AI writing, such as "delve," "interplay," and "testament," has more than doubled. The negative parallelism style, also known as the "it's not just X, it's Y" structure, increased from 0.87 to 2.36 uses per 10,000 pages, though it still remains rare.
Pew says none of these traits can pin down a single document, since human writers use all of them. The claim is about rates across large sets of text.
Other estimates disagree
Graphite, an SEO firm, estimated that the share of newly published English-language articles primarily AI-generated was 49.9% in the first quarter of 2026. A preprint from Imperial College London, Internet Archive, and Stanford found that by mid-2025, 35% of newly published websites are detected as AI-generated or AI-assisted (not peer-reviewed). All estimates lean on Pangram in some form, and Graphite also used Copyleaks and GPTZero.
The differences matter because each study uses different detection tools and thresholds. Pew's threshold catches AI editing too, so a page a person wrote and then cleaned up with a tool lands in the same category as one AI produced start to finish.
Why this matters for writers
Signs of AI text are most prevalent on the commercial web. Pew's .com detection percentage has climbed in every reading since ChatGPT launched, while .edu and .gov have stayed under 2%. Commercial pages are where the concentration sits, and that is the part of the web most SEO work touches.
AI editing is now a native feature in Google Docs and Microsoft Word. Pew's threshold already considers this type of writing, but none of the studies distinguish it from text entirely generated by AI. That means a writer who drafts an article and runs it through an AI cleanup tool is statistically indistinguishable from a writer who pastes a prompt and publishes the output.
Whether a page is accurate, useful, and worth publishing is still decided by reading it. Detector scores can't tell you if a piece is good - they can only tell you how it was probably made. For working writers, the practical question isn't whether your work trips a detector. It's whether you can defend the editorial choices you made, regardless of which tools helped you make them.
Your membership also unlocks:
