Anyone reading long posts on LinkedIn today runs into texts that look like model output. The detection firm Pangram in July labelled more than 40% of long posts on the network (over 250 words) as fully AI-generated. The same detector has now been used for a much broader measurement. According to a Pew Research Center study published on Thursday, more than a third of pages created after ChatGPT launched carry linguistic traces of AI authorship.
What Pew measured
Pew took almost half a million English-language webpages from the Common Crawl archive, 10,000 randomly selected from each of 49 crawls between January 2021 and July 2026. It then ran the body text of each page through the editlens_Llama-3.2-3B model from Pangram, which, according to the methodology description, returns a score from zero (fully human text) to one (fully AI).
Across the whole sample, 10% of pages showed marked signs of AI authorship. For pages published after November 2022, when OpenAI launched ChatGPT, the share in the July 2026 snapshot is substantially higher, at 35%. Pew does not speak of proof, however, but of a linguistic fingerprint. “The model looks at patterns in language, identifying words, phrases and linguistic quirks that are more commonly used by AI than by human authors,” the study says of what the detector looks for.
Where AI text sits
The gap between domains is large. On commercial .com addresses, AI authorship appeared on roughly 10% of pages, and on .org it was 4.6%. Educational .edu and government .gov domains came in around one percent. Most machine-written text therefore sits on the commercial web, while university and government pages stay low.
Pew itself notes where the limits of this measurement lie. “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,” the study says. The figure is a snapshot of one moment and an estimate from a detector, not a tallied state of the web. It should be taken as an indication of scale, with the same caution Pangram shows when it speaks only of “traces”.
Human signature and European regulation
For a European brand that builds its communication on content, the value of a verified human voice is rising. When machine-written text becomes a standard layer of the commercial web and professional networks, the author who is clearly a specific person with real experience stands out. The same concentration that Pangram's data showed on LinkedIn appears in Pew's data for the commercial web.
European regulation is heading into the same space. Article 50 of the AI Act applies from 2 August 2026 and requires providers of AI systems to mark synthetic text, images and audio in a machine-readable way and as artificially generated. The rule targets exactly what Pew and Pangram measure from the outside, namely distinguishing human from machine authorship. How precisely the obligation will be enforced will be settled in practice.
Less than three years after ChatGPT, machine-written text is a standard layer of the commercial web, and a human signature is becoming a recognisable exception.
Comments