Superpower Daily
AWS’s RAG Cost Cut Comes With a 19% Latency BillMeta Is a Major Microsoft AI Customer—and a Potential Foundry R...Google Lets Publishers Add a Preferred Sources Button to Their...GOP Senate Arm Tells AI Companies to Fix Data Centers’ Ohio Pro...Broadcom Gives Its vDefend AI Assistant Power to Push Security...Nvidia Maps AI Memory Between Models Instead of Making Them Sta...AWS’s RAG Cost Cut Comes With a 19% Latency BillMeta Is a Major Microsoft AI Customer—and a Potential Foundry R...Google Lets Publishers Add a Preferred Sources Button to Their...GOP Senate Arm Tells AI Companies to Fix Data Centers’ Ohio Pro...Broadcom Gives Its vDefend AI Assistant Power to Push Security...Nvidia Maps AI Memory Between Models Instead of Making Them Sta...
Culturepublished

Pew’s 35% Web Figure Counts AI Editing, Not Just AI Writing

The detector-based estimate suggests AI has become a substantial part of newer online publishing, but it cannot determine the origin of an individual page.

By 3 min read
Pew’s 35% Web Figure Counts AI Editing, Not Just AI Writing
Sign in to save and react

Listen to this story

The audio brief

About 1:23
0:001:23
Read transcript
Thirty-five percent of newer English-language web pages showed signs of AI authorship or substantial AI editing, according to Pew Research. But that number is narrower—and more uncertain—than it first sounds. Pew removed pages published before ChatGPT launched in November 2022, then analyzed nearly half a million pages from the Common Crawl archive. Among that newer subset, 35 percent triggered Open Pangram’s detector. A separate random sample of 10,000 pages, collected across the web’s full age range, came in at about 10 percent. So this is not a claim that 35 percent of the entire web was written by machines. It is a population-level signal about newer publishing, and the detector can incorrectly flag human writing. It also cannot tell whether a page was fully generated or started as a human draft and then substantially revised with AI. The pattern varied by domain. Dot-com pages showed signals at roughly ten times the rate of dot-edu and dot-gov pages, each near one percent, while dot-org pages were at 4.6 percent. Pew also found more em dashes, Oxford commas, and “it’s not X, it’s Y” phrasing over time—possible tells, not proof. The key constraint is that readers, publishers, and search systems still lack a reliable page-level way to separate machine generation from heavy human-led editing.

Story brief

3 key points

Pew’s estimate depends heavily on what gets counted: 35% of pages published after ChatGPT’s November 2022 launch showed detector signals of AI authorship or substantial editing, versus about 10% across a mixed-age 10,000-page sample. The analysis covered nearly 500,000 English-language pages and used Open Pangram, which can misclassify human writing. The result is best read as a population-level publishing...

  1. 01

    The 35% figure applies only after pre-November 2022 pages were removed; it is not a measure of the entire web.

  2. 02

    .com pages showed signals at roughly 10 times the rate of .edu and .gov pages, each near 1%; .org pages were 4.6%.

  3. 03

    Pew identified rising em dashes, Oxford commas, and “it’s not X, it’s Y” phrasing as additional purported AI tells.

Pew Research found significant signs of AI authorship or heavy AI editing on 35% of the English-language web pages in its post-ChatGPT sample. That is not a count of pages written entirely by machines. It is a detector-based estimate of AI’s combined role in drafting and substantially revising newer web content, and Pew says individual pages can be misclassified.

The sample behind the number

The study starts with ChatGPT’s November 2022 release as its dividing line. Pew analyzed nearly half a million English-language pages from roughly the prior five years, drawn from the Common Crawl web archive. It then used Open Pangram’s technology to identify pages likely written by AI or heavily edited with it.

Pew’s two published percentages answer different questions. A random collection of 10,000 pages gathered in July 2026 included material from across the web’s age range, and about 10% showed significant AI-authorship signs. Once Pew excluded pages published before ChatGPT’s release, the share was 35%. The higher figure therefore describes the analyzed newer-web subset, not the web archive as a whole.

Two cuts of Pew’s July collection
About 10%Random sample, all page ages

About 10% of the random sample of 10,000 pages collected in July 2026 showed significant signs of AI authorship.

35%Pages published after ChatGPT’s release

After the pre-ChatGPT pages were excluded, 35% of analyzed pages showed significant signs of AI authorship or editing.

The signals are not evenly distributed

The detected pattern also varied sharply by domain suffix. Pages on .com domains showed AI-authorship signs at roughly 10 times the rate of .edu and .gov pages, which were each around 1%. .org pages registered a 4.6% rate. Those figures describe detected signals by domain type; they do not establish why the rates differ.

Other writing patterns Pew said increased over time

  • Use of em dashes.
  • Use of Oxford commas.
  • Phrasing built around “it’s not X, it’s Y.”

Pew identified those as purported tells of AI authorship, alongside the detector results. They add a time-based observation to the study, but they are not a substitute for a direct determination of who wrote any particular page or how much editing an AI system performed.

A population-level signal, not a page-level verdict

Pew cautioned that AI detectors such as Pangram can classify human-written pages as AI-written. It nevertheless characterized the large-scale result as directionally reliable. That makes the 35% figure useful as an estimate of a publishing pattern in the sample, while leaving the status of any single flagged page uncertain.

The next question is less whether newer pages bear AI-associated signals than how readers, publishers and search systems will distinguish full generation from substantial revision. Pew’s result makes clear that treating web publishing as a simple human-versus-machine split misses the blended activity its measure is designed to capture.

Sources

  1. techcrunch.comA third of web pages published since ChatGPT's launch show signs of AI authorship, study finds | TechCrunch

Related coverage

Continue reading