Pew Research Center's data-labs team has released a web-sampling report that pulled 10,000 English-language pages from each of 49 Common Crawl snapshots spanning January 2021 to July 2026 — 490,000 pages in total — and ran AI-writing detection on every one. Across the full sample, 10% of pages were flagged as showing clear signs of AI involvement. Strip out the pages that predate ChatGPT's launch, and the share jumps to 35%.
10% and 35% Are Two Different Numbers
These two figures get cited interchangeably, but they answer very different questions. The 10% figure covers the entire sample pool, which is loaded with leftover pages from 2021 and 2022, before any model could write at scale — that dilutes the denominator. The 35% figure looks only at pages that first appeared after November 2022, which is the number that actually answers “how much of new content carries an AI hand.” Pew itself cautions that its detection model runs a higher false-positive rate for years before AI was widely used, so the 2021-2022 portion of the curve shouldn't be taken at face value.
The Cutoff Is Set at 0.2
The detection tool is Pangram's open-source editlens_Llama-3.2-3B, which outputs a score from 0 to 1 — 0 for purely human-written, 1 for purely machine-generated. Pew drew the line at 0.2: anything above that counts as “meaningful AI involvement.” That's a loose threshold — a human-written draft that's simply had its grammar cleaned up by a model would likely cross it too. So what the report is really measuring isn't so much “how much AI wrote” as “how much AI touched.”
To check whether the open-source model holds up, Pew ran 62,370 pages through Pangram's commercial version 3.3 as well; the two tools agreed 96% of the time. But Cohen's kappa on the test sample was only 0.61 — moderate agreement. A single page's score shouldn't be treated as proof of anything; what's reliable is the aggregate trend.
Em Dashes, Oxford Commas, and “Delve”
The report also tracked several stylistic markers. Between January 2023 and January 2026, em dashes per 10,000 words rose from 5.79 to 11.19 — nearly doubling. Oxford commas climbed from 34.04 to 55.51, up 63%. Words like “delve,” “interplay,” and “testament” more than doubled in frequency. And “negative parallelism” — the construction that first rules out one framing before offering another — rose from 0.87 to 2.36, nearly tripling.
None of that is good news for human writers. Em dashes and Oxford commas are both perfectly legitimate habits of English punctuation; it's only because models lean on them so heavily that they've become circumstantial evidence. The more detection tools spread, the more human writers have to prove their own habits are innocent.
.com vs. .gov: a Tenfold Gap
Broken down by domain, roughly one in ten .com pages in the 2026 sample carried AI traces, versus 4.6% for .org and about 1% each for .edu and .gov. The gap roughly tracks each domain's content-production pressure: commercial sites chase traffic through publishing frequency, while academic and government sites face no such incentive.
The sample covers only English-language pages, and anything behind a login wall or paywall is excluded — so this is best read as a snapshot of the public English-language internet. No one has run a comparable study for the Chinese-language web yet, but a rough estimate suggests the share would be higher still, given how industrialized content farms already are. There's a thornier issue underneath all this: these very pages are becoming training data for the next generation of models, and the 0.2 detection threshold will likely need to keep moving upward.
Sources: Pew Research Center Data Labs report and methodology page, CocoLoop, The Independent, TechCrunch; the 490,000-page sample size, 0.2 detection threshold, domain breakdown, and per-10,000-word punctuation frequencies all follow Pew's methodology page.