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
The headline is a good one, I’ll give it that: “74% of New Webpages Include AI Content (Study of 900k Pages)”. It’s by Ryan Law, with Xibeijia Guan and Tim Soulo, of Ahrefs, an SEO software company, published on the Ahrefs blog on 19 May 2025. It’s a vendor study, not peer reviewed, and it was published as Ahrefs was about to launch its own AI detector. The court notes the timing and says nothing further.
The method: Ahrefs’ own detector, codenamed bot_or_not, run on 900,000 English-language pages newly found by its crawler in April 2025, one page per domain. No error rate is published. The authors say it “will never be 100% accurate”, which is honest, I guess, if not exactly reassuring.
For the prosecution: 74.2% of the pages contained AI-generated content. That’s the number everyone repeats.
For the defence: only 2.5% were “pure AI”. And 25.8% were “pure human”, while 71.7% were a mix. Ahrefs’ own chart splits the mix, as shares of all pages: 9.87% minimal AI use (1 to 10% of the page), 25.86% moderate (11 to 40%), 20.50% substantial (41 to 70%) and 15.51% dominant (71 to 99%). So the headline counts any AI at all, even a light touch.
The point for the court: compare case DWT-006, where Graphite found about half of new articles “primarily” AI. Two vendors, two definitions of an AI page. One counts a trace; the other counts a majority. Put side by side, they measure different things, and nobody should treat them as one finding.
The caveats are heavy. It’s a vendor study with a product to sell. It’s detector-based, and partial detection is where detectors are weakest, which is awkward, because partial is where most of these pages landed. One page per domain tilts the sample towards small sites. And this is prevalence, not sameness: it says how many pages involved AI, not whether those pages read alike.
So what does the court have? A large sample and a real number, plus a measuring stick nobody outside the company has checked. Three quarters touched is a striking figure. Three quarters converged is a different claim, and nothing here supports it.
The prosecution can take the prevalence. The defence keeps the doubt about the instrument. Neither gets the whole web.
Verdict: split.
§ IVThe exhibits entered in this case
Any tally counts the evidence items below; it is not a measure of truth.
Evidence for the theory
74.2%
of 900,000 new English pages, April 2025
Ahrefs' own detector, one page per domain. Vendor study; no error rate published.
Entered for the prosecution · Law et al., Ahrefs blog, 2025 · third-party study
Evidence against the theory
2.5%
of new pages "pure AI", April 2025
25.8% "pure human". Another 9.87% showed minimal AI use (1 to 10% of the page) and 25.86% moderate use (11 to 40%).
Entered for the defence · Law et al., Ahrefs blog, 2025 · third-party study
§ VIQuestions for the court
It depends on the definition. Ahrefs (May 2025) found 74.2% of 900,000 new pages contained some AI-generated content, but only 2.5% were pure AI. Graphite (case DWT-006) counts articles that are primarily AI and finds about half.
They measure different things. Ahrefs counts a page with any detected AI; Graphite counts a page that is mostly AI. Both are vendor studies built on detectors.
Ahrefs publishes no error rate and says its detector will never be 100% accurate. Detectors are weakest at partial AI, which is where most of these pages fell.
The prosecution gets the prevalence: most new pages carry some AI text. The defence gets the rest: 2.5% were pure AI, and nothing in the study measures whether the pages became alike.