The dash is one of several AI tells behind the study's estimate of how much of the English-language web is now machine-written.
Among the English-language web pages that appeared after ChatGPT's release, over a third carry the markers of AI authorship. Pew Research Center, a nonpartisan research centre in Washington best known for its opinion polling, found them on samples taken in July 2026, in an analysis published last week. Across the full sample, which reaches back to pages far older than any chatbot, the share is about one in ten.
Individual tells have grown much more common since 2023. Em dashes turn up about twice as often, the Oxford comma is up 63%, and vocabulary the models favour — delve, interplay, testament — has more than doubled in frequency. Phrasings built on negative parallelism, where one characterisation is brushed aside so a second can be put in its place, now appear at close to triple their 2023 rate, though they remain relatively rare overall.
Roughly one page in ten under a .com address shows the same markers. That share stood at about 1% as of January 2021 and 9.35% as of January 2026 — a climb that began in late 2022, when ChatGPT was first opened to the public, and continued as further chatbots, Claude and Gemini among them, followed.
The rise has not been even, and the unevenness is recent. Around the time ChatGPT appeared, the word patterns that hint at machine authorship showed up at practically indistinguishable rates on .com, .org, .edu and .gov alike. In the 2026 samples, .org stands at about 4.6%, while .edu and .gov each stay near 1% — roughly a tenth of the commercial rate.
The gap tracks who is publishing, and why. University and government pages move through editors, need institutional sign-off and answer to slower schedules. The .com category takes in everything from news organisations to affiliate-marketing content mills that push out pages faster than any human editor could keep up with.
Researchers at Pew Research Center's Data Labs assembled about 490,000 English-language webpages from Common Crawl's archive covering January 2021 to July 2026 and ran the text through a detector called Open Pangram. It treats no single word or turn of phrase as evidence on its own, weighing statistical regularities across large bodies of writing instead.
The analysis states its own limits: detectors of this kind can get an individual page wrong in either direction, marking a person's writing as machine-made and machine writing as human. A page flagged for significant signs of AI authorship was also not necessarily produced start to finish by a machine — a large share of the flagged text was probably written with AI assistance rather than generated wholly by it.
Spotting AI writing may not stay a matter of reading style. AI firms including Anthropic are already developing fingerprinting built into the models themselves, an approach that would make machine-written text simpler to identify while holding detection mistakes to a minimum.