AI authorship markers are rising fast across the web, Pew finds

Pew Research ran nearly half a million webpages through an AI detector and found signs of AI authorship or editing on about 10% of the sample. Among pages carrying a post-ChatGPT publication date, the share reaches 35%. Pew's Data Labs team published the analysis on August 20.

Where it concentrates

The split by domain type, a six-month average:

• .com: 9.35%

• .org: 4.59%

• .edu: 1.03%

• .gov: 0.76%

Before ChatGPT launched, all four domain types sat at or below 1%. Only .com and .org have climbed since; .edu and .gov have stayed close to that mark.

The writing tells are getting more common

Pew tracked specific markers on post-ChatGPT pages. Em dashes rose from 5.79 uses per 10,000 words in early 2023 to 11.19 in early 2026. Oxford commas are up 63%. Words like "delve," "interplay," and "testament" have more than doubled in frequency. The "it's not just X, it's Y" construction rose from 0.87 to 2.36 uses per 10,000 pages, though Pew says it remains rare.

Other groups measuring the same thing land on different numbers. Graphite put the share of newly published English-language articles that are primarily AI-generated at 49.9% for the first quarter of 2026. A preprint from Imperial College London, Internet Archive, and Stanford, not yet peer reviewed, found 35% of newly published websites detected as AI-generated or AI-assisted by mid-2025. All three estimates lean on the Pangram detector in some form.

What the score does not tell you

Pew is explicit that none of these markers identify a single document. The claim only holds across large sets of text, and the same threshold catches a page a person wrote and then edited with a tool as it catches one a model produced from nothing.

That is the gap worth naming. A detector score describes how a page was assembled. It says nothing about whether the page carries a fact, a data point, or a piece of first-hand experience a model did not already have. A page can trip every marker Pew tracked and still be the only source with the number a reader, or an AI answer engine, actually needed. Pew's own conclusion lands there too: whether a page is accurate, useful, and worth publishing is still decided by reading it.