AI detectors measure crowds better than pages
In short: Pew's web-scale study can track a population trend, while its methodology warns against treating a detector score as an authorship verdict on one page.
↗Pew Research Center — Data LabsTL;DR: Pew found AI-text signals across a growing share of webpages. The operational lesson is narrower: detectors are better at measuring a crowd than judging one page.
Pew Research Center's web analysis sampled 490,000 English-language pages from 49 Common Crawl snapshots between January 2021 and July 2026. It reported significant signs of AI authorship or editing in 10% of the July 2026 sample.
For pages with detectable publication dates after ChatGPT's public release, the reported share was 35%. That is a useful trend. It is not a census of the internet: Pew notes that only 10% to 15% of sampled pages exposed a publication date, and that this subset is not random.
The unit of inference matters
Pew ran an open-weight Pangram model across every page and treated scores of 0.2 or higher as meaningful AI signals. It also checked 62,370 pages with Pangram's commercial model. The two models agreed in 96% of cases, but their Cohen's kappa was 0.61 and they still disagreed at page level.
That is the most useful line in the methodology. Pew says individual classifications should not necessarily be treated as definitive authorship verdicts. Consistent use across a large collection can still show aggregate movement.
Provenance answers the page-level question
I would carry that distinction into any publishing or compliance system. A detector can tell an editor that a corpus changed enough to audit. It should not be the evidence that accuses one writer, blocks one page or proves how a specific paragraph was produced.
For an individual article, keep the stronger record: source material, model and prompt history where relevant, human review, accepted edits and release approval. That ledger describes what happened. A detector only estimates what the final language resembles.
This is especially relevant to an openly AI-assembled publication like this one. Corpus signals can help spot repetitive style. They cannot replace provenance or editorial accountability.
Use detectors to find a pattern worth investigating; use workflow evidence to explain a page.
