The Court of Public Prose
In re Dead Writing Theory
§ IFindings of fact
§ IICaveats for the jury
§ IIIThe court reporter's account
By Jack Stovell
This one’s a bit different, and I’ll be upfront: it’s a preprint. He, Ardi, Govindan and Madhyastha, University of Southern California, April 2026, arXiv, not yet peer reviewed. The paper is called “DeGenTWeb: A First Look at LLM-dominant Websites”. Treat it as a witness who hasn’t been cross-examined yet.
The twist is that they don’t judge articles one by one. They classify whole websites as LLM-dominant, running a calibrated detector over several pages per site. The sample was roughly 100k sites archived by Common Crawl, plus roughly 20k sites from Bing results for 10k how-to searches.
And the numbers jump out. LLM-dominant sites rose from 2.1% in the second half of 2022 to 29.4% in the first half of 2025. That’s a lot of sites that, as far as the detector can tell, had very little human input.
Then the searches. For 46.6% of the how-to searches, at least one top-10 result was an LLM-dominant site. Widen it to the top 20 and it’s 65.7%. If you’ve ever searched for how to fix something and landed on a page that says a great deal and tells you nothing, well, you’ll feel seen.
Money turns up too. 78.8% of LLM-dominant sites have a clear financial incentive, against 55.8% of other sites. Make of that what you will; I’d say it’s a motive, and every good trial needs one.
Fairness requires the caveats. It’s a preprint. It’s detector-based, though they validated on pre-ChatGPT sites and flagged only 0.29%, and they openly warn that AI-text detectors perform much worse than advertised when you try not to falsely accuse human writers. And it counts sites with little human input, so it misses human copy that was merely AI-assisted. The real number of machine-touched pages could be higher, or the labelling could wobble. Probably some of each.
Verdict: for the prosecution.
§ IVThe exhibits entered in this case
Any tally counts the evidence items below; it is not a measure of truth.
Evidence for the theory
2.1% → 29.4%
of sites, second half of 2022 → first half of 2025
Roughly 100k sites archived by Common Crawl, classified site by site. A preprint; on pre-ChatGPT sites the classifier flagged only 0.29%.
Entered for the prosecution · He et al., arXiv preprint 2605.00087, 2026 · third-party study
46.6%
of 10,000 how-to searches had one in the top 10
65.7% for the top 20 results. 78.8% of LLM-dominant sites have a clear financial incentive, against 55.8% of other sites.
Entered for the prosecution · He et al., arXiv preprint 2605.00087, 2026 · third-party study
Evidence against the theory
No exhibit entered for the defence in this case.
§ VThe finding
LLM-dominant websites rose from 2.1% to 29.4% of sites sampled between late 2022 and early 2025, and they turn up in how-to search results. A preprint, detector-based.
§ VIQuestions for the court
In the DeGenTWeb preprint (He et al., University of Southern California, 2026), the share of LLM-dominant sites in its Common Crawl sample rose from 2.1% in the second half of 2022 to 29.4% in the first half of 2025.
Yes. For 46.6% of 10,000 how-to searches, at least one of the top 10 Bing results was an LLM-dominant site; for the top 20 results it was 65.7%.
The authors warn that detectors "perform much worse than advertised" when you try not to falsely accuse human writers. Their own site-level classifier flagged only 0.29% of pre-ChatGPT sites, but the study is a preprint and not yet peer reviewed.