Pew Research Finds Signs of AI Authorship in a Third of Web Pages Published Since ChatGPT

Pew Research Center says just over a third of English-language web pages published since ChatGPT’s November 2022 launch show measurable signs of AI authorship, a finding the nonpartisan research organization published Thursday in a new data essay built on nearly half a million pages pulled from the Common Crawl web archive.

The report, “How Much of the Internet Is Written With AI?”, is credited to Pew’s Samuel Bestvater, Aaron Smith, Carson TerBush, Chris Baronavski and Janakee Chavda. It sets out to answer a narrower question than the one usually asked about AI and the web: not how much of everything ever posted online carries an AI’s fingerprints, but how much of what’s been published since generative writing tools became widely available does.

A Random Sample Undersold the Trend

Pew’s researchers pulled nearly 500,000 English-language pages from Common Crawl, an open, continuously updated archive of the web, spanning publication dates from January 2021, roughly two years before ChatGPT’s debut, through July 2026. To flag likely AI involvement, they ran the pages through an open-weight detection model called Open Pangram, built by Pangram to spot statistical differences in how human authors and AI models tend to use language. The same detection technology now powers Substack’s reader-facing AI scanner, which lets readers check individual posts and replies for AI-generated text.

A simple random sample of 10,000 pages collected in July 2026 put the AI-authorship rate at a modest 10 percent. Pew flagged that number as misleading on its own: a random sample pulled from a five-year archive will inevitably include pages written years before any AI writing tool existed, dragging the average down. Once the researchers isolated only pages published after ChatGPT’s release and excluded everything older, the rate jumped to just over a third, 35 percent, of pages showing significant signs of AI authorship or heavy AI editing.

Commercial Sites Are Driving the Shift

The rate varies sharply by where a page lives online. Pages on .com domains showed AI-authorship signs at roughly 10 times the rate of pages on .edu or .gov domains, both of which sat near 1 percent in Pew’s July 2026 snapshot. Nonprofit .org domains landed in between, at 4.6 percent. The pattern lines up with the incentive structure around these tools: commercial publishers chasing search traffic and content volume have the clearest reason to lean on AI writing, while universities and government agencies mostly don’t.

Pew is upfront that Open Pangram, like any AI-detection tool, will misclassify some individual pages in both directions. The researchers argue that at the scale of hundreds of thousands of documents, the pattern holds even if any single flagged page might be a false positive. They also tracked a handful of stylistic tells that grew more common across the same period, including heavier use of the em dash character, more frequent Oxford commas, and the construction “it’s not X, it’s Y,” patterns AI writing models favor more than typical human prose.

A Web Increasingly Written and Read by Machines

TechCrunch, which covered the study Thursday, tied the finding to a separate milestone Cloudflare disclosed earlier this year: automated traffic, meaning AI agents, scrapers, and other bots, has overtaken human visitors on the open web. Cloudflare’s own Radar dashboard put the split at 57.5 percent bot traffic to 42.5 percent human by early June 2026, according to Forbes, a threshold CEO Matthew Prince had not expected to cross until the end of 2027.

Put together, the two data points describe a web increasingly written by one kind of automated system and read by another. Pew’s report stops short of drawing conclusions about what that means for search rankings, content moderation, or professional writing online, but the trajectory in its own numbers is stark: the AI-authorship rate more than triples, from 10 percent across the full five-year archive to 35 percent in the slice published after ChatGPT arrived.