The Pew Research Center ran close to 500,000 English-language pages through the Pangram detector and found strong signs of machine drafting on more than a third of everything posted since ChatGPT shipped. Measured against the whole archived web, AI writing drops to 10 percent of pages. The distance between those two numbers is the story.
Key Takeaways
- More than a third of pages published after November 2022 carry markers of generated text, against 10 percent across the full sample.
- The .com space runs roughly ten times the rate of .edu and .gov, both of which sit near 1 percent.
- Since 2023, em dash frequency has doubled in published text and negative parallelism has nearly tripled.
Have an AI Sum Up This Article
ChatGPTHalf a million pages run through the Pangram detector
The sample comes from Common Crawl, the archive that takes periodic snapshots of the open web. The Pew Research Center team pulled close to 500,000 English-language pages spread from January 2021 to July 2026, which gives the study a baseline that predates mainstream access to language models.
Detection runs on Pangram, a machine learning model trained to catch the regularities in vocabulary, phrasing and structure that turn up more often in machine-drafted text than in human copy. The same tool has already been used to measure machine drafting inside a single social platform, which makes the two readings directly comparable.
The headline finding is two numbers that answer different questions. On the July 2026 snapshot, 10 percent of pages across all dates show strong signs of machine drafting. Filter down to pages posted after November 2022 and the share passes one third, and the Pew Research Center walks through that gap in its count of how much of the internet is now written with AI.
The composition of the archive explains the spread. A large slice of the sampled pages predates generative models and could not have been written by one. AI writing is therefore packed into a short window, which mechanically raises its density inside recent publishing.
The team states its limits up front rather than burying them. The detector misfires on individual documents and only becomes dependable across large volumes of text. The authors, Samuel Bestvater, Aaron Smith, Carson TerBush, Chris Baronavski and Janakee Chavda, present the figures as a mass trend, not a page-by-page verdict.
The .com side flipped while .gov stayed put
The breakdown by domain extension is the most useful part of the count. .com sits near 10 percent on the July 2026 snapshot, .org at 4.6 percent, while .edu and .gov both hover around 1 percent.
The order of magnitude matters more than the decimal. A commercial domain carries roughly ten times the traces of machine drafting that a university or government domain does. The shift tracks the money rather than the tooling, since all three categories have had access to the same models since the same date. The same platform-level split showed up when LinkedIn hosted 41 percent of the long posts written by AI.
The second measurement looks at writing habits themselves. Since 2023, em dash frequency has doubled in published text, the Oxford comma is up 63 percent, vocabulary associated with language models has more than doubled and negative parallelism constructions have nearly tripled.
None of those markers proves anything on its own, and that is exactly what makes them useful statistically. An Oxford comma convicts nobody. Four separate markers bending at the same time, measured across hundreds of thousands of pages, describes a shift in the average style of the English-language web.
Usage fills in the rest of the picture. Around half of American adults say they use a chatbot and 24 percent use one every day. Assisted writing has stopped being a professional specialism and turned into an ordinary habit.
That spread explains why AI writing no longer sits inside one category of site. Commercial pages pull the total up, but no measured extension held its 2021 level. The move runs across the whole corpus, at very different intensities.
More articles on Horizon
- OpenAI Watches for Abuse Without Keeping Your Data
- Claude Designed Proteins That Worked in the Lab
- Cursor Origin Takes On GitHub During a Six-Hour Outage
What the count changes for publishers and search engines
For a publisher, the number moves the question. Shipping assisted copy is no longer a differentiator once it is the statistical norm on .com. What sets a page apart now sits beyond the prose itself, in proprietary data, original reporting, or a named source the reader can go and check.
Disclosure is the neighbouring problem. It came up in very concrete terms when eight Pulitzer 2026 winners disclosed AI use in their work, with wildly different levels of involvement from one entry to the next.
On the search side, the count lands as reader patience thins out. The rise of DuckDuckGo, whose 30 percent jump tracked a rejection of forced AI answers, feeds on the same fatigue. An index filling up with generated text makes sorting more expensive as the volume grows.
Detector vendors inherit their own version of the squeeze. Pangram earns its keep by measuring a mass rather than convicting a document, and that distinction will be hard to hold in front of buyers who want a binary ruling on a student essay or a job application.
Regulators now have an order of magnitude where they previously had estimates. A disclosure rule no longer targets a fringe practice but a majority share of what ships on commercial domains, which changes the compliance load for the platforms in scope.
For editorial teams, the practical effect is a shift in what quality control means. Checking whether a piece was assisted matters less than checking what it claims, since AI writing covers both competent synthesis and pages with nothing in them.
Then there is the training corpus. Tomorrow’s models will learn from a web where a third of recent pages carries the fingerprint of yesterday’s models, and none of the labs measured here has published what that loop does across several generations.
Follow the story on Horizon.
