Pew Put a Number on AI-Written Web Pages, and Published the Tells

Ten percent of the web pages Pew Research Center sampled show significant signs of AI authorship as of July 2026. Narrow the sample to pages published after ChatGPT’s release on 30 November 2022 and the share climbs past one-third. The corpus was almost 500,000 English-language pages pulled from the Common Crawl archive, spanning January 2021 to July 2026.

The headline number will get quoted. The more useful part of the study is underneath it. Pew published the specific linguistic markers its classifier leans on, and how far each one has moved since a 2023 baseline. That converts “sound less like a machine,” which is a taste judgment two reviewers can argue about forever, into something a pipeline can count.

What the classifier measures, and what it cannot do

The pages were scored by Open Pangram, a machine learning model that, in Pew’s description, “looks at patterns in language, identifying words, phrases and linguistic quirks that are more commonly used by AI than by human authors.” Nobody read the pages. There were no survey respondents, no interviews, no room. The unit of analysis is an archived document and a model’s score on it.

Pew states the limitation plainly: “AI detection models aren’t perfect - they sometimes misclassify individual documents that were written by humans as including signs of AI authorship, and vice versa.” So the 10% and the one-third are detector-derived estimates of a corpus, with error running in both directions.

That constraint decides what you are allowed to do with the finding. Measuring the composition of a large body of text is supported by this method. Rendering a verdict on the single document in front of you is not, and Pew’s own caveat is the reason. If a vendor tells you their detector can settle authorship of one page, they are claiming more than this research supports.

Four markers, and how fast they moved

Pew reports the change in each marker against a 2023 baseline, measured per 10,000 words. Em dashes rose roughly 193%, the sharpest move in the set. Oxford commas rose 63%. AI-typical vocabulary more than doubled. Negative parallelism nearly tripled.

That last one deserves a definition, because it is a structure rather than a word. Negative parallelism sets up an option, rejects it, and hands you the preferred one in the same breath. A specimen: “It’s not about speed, it’s about trust.” The shape flatters the writer with a feeling of insight and gives the reader an argument that was never made.

Pew is careful about a point that anyone building checks on top of this needs to internalize. Human authors use every one of these features. Writers have always used the em dash, the Oxford comma is a standing house preference in English style guides, and the negative-parallelism construction long predates the transformer. Detection works on aggregate statistical patterns across many documents, so no individual signal identifies anything. A page dense with em dashes is a page dense with em dashes.

The domain split is the more useful half of the study

Broken out by top-level domain, the estimated share of pages with signs of AI authorship diverges sharply. Commercial pages sit at approximately 10% for .com. Nonprofits come in at 4.6% for .org. The two domain classes that carry institutional credibility sit far below both: approximately 1% for .edu and approximately 1% for .gov.

Pew does not say why the domains diverge, and the study cannot answer it: it measures pages, not the processes behind them. The most economical explanation available to a content operation is review load, and it is a hypothesis rather than a finding. Where more hands touch a page before it publishes, fewer pages carry the markers.

Which means the domain gradient is a rough map of where editorial process still binds. If your content is meant to carry the weight of a .edu page while shipping at the speed of a .com page, this is the number to put in front of the team.

A style rule can double as an auditable control

Our house style has banned the em dash since before anyone could measure it, on the argument that it reads as machine rhythm. Pew now shows the em dash as the fastest-moving marker in the set, up roughly 193%. The rule did not get smarter. It got a measurement behind it, which is the difference between a preference a reviewer defends in a comment thread and a check that runs before the comment thread exists.

Counting a marker and running a detector are two different exercises, and the distinction survives the whole argument. A detector returns a probability that a document was machine-written, which Pew’s caveat tells you to distrust one document at a time. A marker count returns a fact about the text: how many em dashes it carries per 10,000 words. The fact is verifiable, reproducible by anyone with the same file, and free of any claim about who typed it. A review process can act on the second kind of number honestly. It cannot act on the first.

That is the transferable move here. Three of the four markers are countable with a regular expression and a word count. Once a preference is expressed as a count per 10,000 words, it stops consuming review attention, the same way a lint rule stopped anyone from arguing about indentation.

Do this now

Take your last twenty published pieces and count four things in each: em dashes, Oxford commas, your own list of AI-typical vocabulary, and negative-parallelism constructions. Normalize every count per 10,000 words so a long piece and a short one are comparable. You now have a baseline for your own corpus, which is the only baseline that matters, because Pew’s numbers describe the open web and not your team.

Then wire the counts into review as a flag with a threshold you set yourself, and keep them out of the merge decision. The count tells a reviewer where to look. It never tells anyone who wrote the page, and a pipeline that treats a threshold as a verdict will reject good human writing, which is the failure Pew warned about.

For a content operation, the instrument is what to take from this study. Pew made authorship style a measurable property of a document this year, with published markers and published error behavior. Governance that runs on taste can now run on counts, and the part that still requires judgment is smaller and better defined than it was before this study.

---

This analysis draws on How much of the internet is written with AI? (Pew Research Center, August 2026).

Victorino Group helps content and engineering teams turn editorial standards into automated checks that run before review. Let’s talk.