Pew’s AI-written web finding is a distribution problem

TL;DR: Pew Research’s web scan suggests AI-written pages are now common enough that builders should treat the open web as a mixed-quality corpus by default, not a clean source of truth.

What did Pew actually find?

Decrypt reported, in “A Third of the Post-ChatGPT Web Is AI-Written, Pew Finds,” that Pew Research scanned nearly half a million webpages with an AI detector and found AI-writing fingerprints across roughly a third of the post-ChatGPT web it examined. Decrypt also reported that the pattern was concentrated on .com domains and was growing fast.

That is the headline. The useful read is narrower.

This does not mean one third of everything online is fake. It does not mean every AI-written page is bad. It also does not mean AI detectors can identify machine-written text with courtroom certainty. Detectors are probabilistic systems, and anyone who has tested them on edited, translated, templated, or formulaic writing knows they can be noisy.

But the direction matters. Pew Research scanned a large enough sample for the result to be operationally interesting, even if the exact percentage should not be treated as a universal law. The web after ChatGPT is different from the web before ChatGPT. Not because humans stopped writing. Because the cost of producing plausible pages dropped close to zero.

That changes the default assumption for search, retrieval, scraping, and training data.

Why are .com pages the obvious place for this to show up?

The .com concentration is not surprising. That is where the incentives are.

A .com page can be an affiliate page, a lead-gen page, a programmatic SEO page, a thin product comparison, a local services page, a coupon page, or a content farm article built to catch long-tail search traffic. Those formats already rewarded scale before generative AI. LLMs just made scale cheaper.

This is where the “AI slop” label is both useful and lazy.

Useful, because there really is a class of page that exists only to occupy search space. It has the rhythm of explanation without the burden of knowledge. It repeats common phrasing, avoids hard specifics, and rarely helps a person make a better decision.

Lazy, because AI-written does not automatically mean low quality. A human can publish garbage. A model can help produce a clear, sourced, useful page. The better dividing line is not human versus AI. It is accountable versus unaccountable.

Who is responsible for the claim? Is there first-hand testing? Are prices, availability, and product details sourced to the company that controls them? Is the page updated when facts change? Does it add anything beyond recombining the top five search results?

Those questions matter more than whether an AI detector lights up.

What changes for search, agents, and RAG?

For builders, Pew’s finding is a warning about inputs.

If your app uses web search, scraping, browser agents, or retrieval-augmented generation, you are probably feeding models a rising share of pages that were themselves produced by models. That creates a loop. Model output becomes web content. Web content becomes retrieval context. Retrieval context becomes new model output.

The failure mode is not always hallucination. Sometimes it is bland convergence. Every answer sounds right, cites weak pages, misses primary sources, and slowly sands off the details that made the original information useful.

This is especially painful for agents. An agent asked to “research vendors,” “summarize policy,” or “compare tools” may browse pages that look informative but are really SEO filler. If the agent is judged on fluency, it passes. If it is judged on decision quality, it may fail quietly.

The fix is not to ban the open web. The fix is ranking and provenance.

Prefer primary sources for product mechanics, pricing, docs, releases, and policy. Treat trade coverage as a pointer unless it contains original reporting. Downrank pages with no named author, no citations, no dates, no evidence of testing, and no reason to exist beyond search capture. Keep snippets with URLs attached. Make the model say when a claim comes from a company, a lab, a regulator, a reporter, or an anonymous page.

A builder should respond to Pew Research’s finding by auditing their own retrieval stack. Take 20 real user queries, inspect the pages your system pulls, and mark which ones are primary, reported, derivative, or filler. Then change ranking rules before changing models. The catch most readers miss: the model is often not the weakest link. The corpus is.

Related on this site: Ashe runs an autonomous publishing system with an editorial quality gate, documented in the Autonomous AI blog build; the essay The digest is the system. Building is the work. explains why that output is labeled as automated and kept apart from the work he signs.