The Web Is Eating Itself: AI Now Writes a Third of New Pages

Over a third of the web pages published since ChatGPT launched in November 2022 now show signs of being written or substantially edited by AI. That is the headline finding of a new study from Pew Research Center, which sampled nearly half a million English-language web pages from the Common Crawl archive and ran their text through an AI-detection tool called Open Pangram.

What happened

Pew's researchers drew a random sample of 10,000 web pages collected in July 2026 and found that roughly 10% showed "significant signs of AI authorship." That figure alone understates the trend, because a random sample of the whole web inevitably includes millions of pages that predate AI writing tools entirely. When Pew filtered out pages published before ChatGPT's release, the share jumped to 35%.

The distribution is not even. Commercial .com domains showed AI authorship at nearly ten times the rate of .edu and .gov domains, around 9.4% of .com pages versus roughly 1% each for .edu and .gov. Nonprofit .org domains sat in the middle at 4.6%. In other words, AI-written text is concentrated exactly where the economic incentive to publish is highest: marketing pages, SEO content farms, product listings, and affiliate blogs, not universities or government sites.

The study is not an outlier. It joins a growing body of independent research, aggregated under the "AI on the Internet" project, that has tracked the same pattern: the newest layers of the public web are disproportionately machine-generated.

Why it matters

The finding lands the same month Cloudflare reported that bot web traffic has overtaken human traffic for the first time. Put the two together and the picture is stark: the modern web is increasingly a machine-to-machine medium, where bots read pages written by other bots.

That shift has consequences that ripple far beyond content quality. The web is the largest training corpus in existence, and every frontier model scrapes it. If a growing share of new content is itself AI-generated, then future models are increasingly trained on their own output, a feedback loop researchers have been warning about for years under labels like "model collapse" or the "Habsburg AI" problem.

The stakes are higher for smaller players than for frontier labs. Hyperscalers can afford human-curated and licensed corpora. Independent builders and fine-tuning teams, who lean hardest on open web data, are the ones most exposed to a corpus that is quietly filling with synthetic text.

Pew also documented the linguistic fingerprints of AI writing and found them spreading across the web. Em dashes now appear roughly twice as often as they did in 2023. Oxford commas are up 63%. Words that AI models overuse, "delve," "interplay," "testament," have more than doubled in frequency. The "it's not X, it's Y" rhetorical structure has nearly tripled.

The detection arms race

None of this detection is foolproof. Pew itself notes that AI-detection tools like Open Pangram can misclassify individual documents. But at the scale of half a million pages, the signal is directionally reliable.

The deeper problem is that detection is an arms race. The linguistic tells Pew measured, em dashes, Oxford commas, a specific overused vocabulary, are a snapshot of how current models write. As models improve, and as writers learn to strip those tells, detection becomes harder. Anthropic is moving in the opposite direction, toward baked-in provenance: the company recently detailed how its Claude watermarks will work, signaling that the industry's long-term answer may be authentication rather than detection.

The two approaches are complementary, and the gap between them matters. Detection is reactive and probabilistic. Provenance is proactive and cryptographically strong. For a web that is increasingly written by machines, the only durable answer is to know where content came from, not to guess whether a model made it.

Builder impact

For AI builders and automation engineers, this is not an academic concern. It changes the substrate on which agents, retrieval systems, and models all depend.

First, retrieval-augmented generation. Every RAG pipeline and web-searching agent now pulls from an index where a growing fraction of "knowledge" is synthetic. Without provenance filtering, an agent answering a question is as likely to ground itself in AI slop as in a primary source. Prefer .edu and .gov sources where authority matters, and treat .com content as a signal to verify, not a source of truth.

Second, training data. Teams fine-tuning or training models on scraped web data need provenance signals, domain, authorship likelihood, recency, to avoid poisoning their datasets with the same synthetic patterns they are trying to escape. A simple domain-tier filter is the cheapest first step; AI-likelihood scoring is the more aggressive one.

Third, content pipelines. The same detection signals Pew used can be repurposed defensively. Builders shipping content-generation tools can add provenance and disclosure by default, rather than waiting for detection to catch up. The teams that build disclosure into their output now will be ahead of the coming regulation.

What to watch next

The trajectory points toward provenance infrastructure. Expect more watermarking and signed-content standards, pressure on platforms to disclose AI-generated material, and a growing market for detection and filtering tooling. Anthropic's watermarking work is an early signal of where the frontier labs are headed.

The key question for builders is whether the web's training-data layer, the raw material for every model that comes next, gets provenance built in before the synthetic share compounds further. The study gives a number to a problem everyone suspected. The next chapter is about what the industry does with it.