Pew Research Finds a Third of Web Pages Written Since ChatGPT Show AI Signs

Pew Research Center reported on August 20 that 35% of English-language web pages published since ChatGPT's November 2022 launch show significant signs of AI authorship, based on an analysis of nearly 490,000 pages from the Common Crawl archive. The study, which used the AI detection model Open Pangram, also found that 10% of a July 2026 snapshot of all web pages shows such signs, with commercial .com sites at roughly 10% versus 4.6% for .org domains and near 1% for .edu and .gov sites.

Pew Research just put a hard number on the "dead internet" meme: more than a third of web pages published since ChatGPT launched carry clear signs of AI authorship. You've probably felt it before you ever saw a statistic for it: the sameness of search results, the blog posts that all reach for "delve," the reviews and how-tos that read like they came from nowhere in particular. Now there's a number. Pew Research Center reported on August 20 that 35% of English-language web pages published since ChatGPT's November 2022 launch show significant signs of AI authorship. The finding comes from an analysis of nearly 490,000 pages pulled from the Common Crawl archive, spanning 49 separate crawls between January 2021 and July 2026. Zoom out to the whole web, old pages and new, and the number drops but doesn't disappear. Pew found that 10% of a July 2026 snapshot of pages shows significant signs of AI writing. That's the baseline the "dead internet theory" crowd has been arguing about for years, mostly without data. Now they have some. Pew ran its sample through Open Pangram, an AI detection model, rather than relying on guesswork. That matters, because detection tools can misfire on any single page in either direction. Pew itself is careful to note that "significant signs of AI authorship" doesn't mean a page was written entirely by a machine. Plenty of what got flagged was probably AI-assisted rather than AI-generated outright: a person drafting with a chatbot open in another tab, not a bot working unsupervised. The domain breakdown is the most useful part of the study, and it's not the one you'd guess. Commercial .com sites show AI-authorship signs at roughly 10%, about double the 4.6% rate on .org domains, according to Pew's data as reported by TechCrunch. Government and education sites are barely touched. Both .edu and .gov domains sit near 1%, about a tenth of the .com rate. Read plainly, that means the parts of the web most exposed to advertising and SEO pressure are the parts leaning hardest on AI to fill pages. Nonprofits and institutions mostly still write their own copy. ChatGPT Can Now Read and Send Your iMessages on a Mac https://startupfortune.com/chatgpt-can-now-read-and-send-your-imessages-on-a-mac/ OpenAI's new ChatGPT plugin for Apple Messages can read, draft, sort and send iMessages, SMS and RCS texts on a Mac after users grant Full Disk Access and contacts permissions. A default approval step checks each outgoing message, but users can switch it off, a setting privacy researchers are already flagging as risky. - chatgpt read and send imessages on mac https://startupfortune.com/chatgpt-can-now-read-and-send-your-imessages-on-a-mac/ - how to use chatgpt with apple messages app https://startupfortune.com/chatgpt-can-now-read-and-send-your-imessages-on-a-mac/ Pew's researchers also found the tics of AI writing showing up in aggregate, across the web as a whole and not just on flagged pages. Em dashes have roughly doubled in frequency. Oxford commas are up 63%. Words chatbots favor, like "delve" and "interplay," now appear more than twice as often as they did before ChatGPT arrived. That's not a handful of outlier pages. It's a shift in how the entire web sounds. The story has legs. TechCrunch, Slashdot and Yahoo Tech have all picked it up in the days since Pew's release, and it's been reframed everywhere from "the dead internet theory may be coming true" to blunter takes calling a third of the web AI slop. Why founders building on web data should pay attention For anyone training models or building products on scraped web text, this isn't a curiosity. It's a data-quality problem. Search engines and large language models increasingly train on, and rank, text that a language model itself produced. Researchers have warned for years about model collapse, the risk that AI systems trained repeatedly on AI-generated output degrade in quality and diversity over successive generations. Pew's numbers are the first large, systematic evidence of how fast that raw material is piling up in the wild, not in a lab simulation. There's a practical angle too. If a third of new pages already carry AI's fingerprints, telling a real product review, a founder's actual blog post, or a genuine customer testimonial apart from synthetic filler is about to get harder, for readers and for the crawlers that rank them. That's a real opening for anyone willing to publish first-hand reporting, real data and named sources instead of the confident, fact-free prose a chatbot produces by default. The scarce resource on the internet isn't fluent writing anymore. It's proof someone actually did the work. Also read: TCS Will Buy Porsche's MHP Consulting Unit for $373 Million https://startupfortune.com/tcs-will-buy-porsches-mhp-consulting-unit-for-373-million/ • Visa and Mastercard Join 26 Firms to Set Rules for AI Agent Payments https://startupfortune.com/visa-and-mastercard-join-26-firms-to-set-rules-for-ai-agent-payments/ • General Intuition Hits A $6 Billion Valuation Betting Video Games Teach Robots https://startupfortune.com/general-intuition-hits-a-6-billion-valuation-betting-video-games-teach-robots/