A new study from Pew Research Center finds that roughly a third of English-language web pages published since ChatGPT’s late-2022 debut carry clear statistical signs of AI authorship, in one of the largest attempts yet to measure how much of the open web is now machine-written.
What the study found
Pew’s Data Labs team analyzed close to 500,000 English-language pages pulled from the Common Crawl web archive, covering snapshots from January 2021 through July 2026. Researchers ran the pages through Pangram’s open-weight AI detection model to flag writing patterns associated with large language models.
Across the full July 2026 sample, 10% of pages showed significant signs of AI authorship. Narrowing the sample to pages published after ChatGPT launched in November 2022, that share jumped to more than a third.
AI-flagged writing was not evenly spread across the web. Roughly one in ten .com pages showed AI signals, compared with 4.6% of .org pages and about 1% each for .edu and .gov domains, according to the report.
Writing style has shifted, too
The researchers also tracked how web writing style changed between 2023 and 2026, a period that roughly tracks generative AI’s mainstream adoption. Em dash usage roughly doubled, Oxford comma usage rose 63%, and certain words associated with AI-generated prose — such as “delve,” “interplay,” and “testament” — more than doubled in frequency. Constructions like “it’s not just X, it’s Y” nearly tripled.
Pew’s researchers caution against reading too much into any single marker. An em dash or an Oxford comma on its own says nothing about who wrote a page; the detection model instead relies on statistical patterns across large collections of text, and even those patterns can misfire on individual pages.
Why it matters
The findings add texture to an ongoing debate about how AI detection tools work and how much to trust them — a question we’ve covered in more detail in our explainer on how AI text detectors work. They also bear on search engines, publishers, and platforms trying to distinguish human from machine writing at scale, and add to a growing body of AI research documenting how quickly generative tools — built on the GPT model family and its rivals — have reshaped everyday content on the internet.