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
Case DWT-006 found that about half of new articles are primarily AI-generated. That came from a Graphite study. This is the other Graphite study, and it asks the obvious follow-up: fine, but does anyone see that stuff?
A flag before I go on. The authors are Paredes, Druck and Smith, of Graphite, an SEO and growth agency. The piece is “How Does AI-Generated Content Perform in Search and Answer Engines?”, on the Graphite research page, 14 October 2025. It’s a vendor study, not peer reviewed. The court should weigh it accordingly.
The method: 31,493 keywords across 10 categories. For each keyword, the first two Google results pages in June 2025. Citations from ChatGPT and Perplexity for 100 keywords per category, turned into questions. English articles and listicles were classified in 500-word chunks with Surfer’s AI detector, whose false-positive rate Graphite measured at 4.2%.
The findings run the other way from what the doom-mongers expect. 86% of articles ranking in Google Search were written by humans, and 14% were generated with AI. For answer engines it was the same split twice: 82% of articles cited by ChatGPT and 82% cited by Perplexity were human-written, 18% AI. Only 7% of number-one Google results were AI-generated, and human pages ranked higher than AI pages for the same keyword.
Topic matters, though. AI share varies: Food sat at 4.5% and News at 4.7% “high confidence AI”, while Productivity reached 14.1% and Crypto 13.3%. So the machine’s footprint is uneven.
The point for the court: the machine may write half of what is published, but what readers are actually shown is mostly human. Publication and visibility are different things. This study is about the second one, and it measures prevalence in what gets surfaced, not whether the writing converged.
Caveats, naturally. It’s a vendor study, and detector-based, so every number inherits the detector’s flaws (see Case DWT-009, which I’d reread). It doesn’t measure AI drafts heavily edited by humans, which the authors think may perform well. And ranking depends on many hidden variables, so no one should read causes into a snapshot.
Even so, the defence gets something real here. If the flood is mostly unread, the flood matters less than the headline suggests.
Verdict: for the defence.
§ IVThe exhibits entered in this case
Any tally counts the evidence items below; it is not a measure of truth.
Evidence for the theory
No exhibit entered for the prosecution in this case.
Evidence against the theory
86%
of articles ranking in Google Search written by humans, June 2025
31,493 keywords across 10 categories. Only 7% of number-one results were AI-generated. Vendor study, detector-based.
Entered for the defence · Paredes et al., Graphite research (Five Percent), 2025 · third-party study
82%
of articles cited by ChatGPT, and by Perplexity, written by humans
18% AI-generated. Surfer's detector, with a false-positive rate Graphite measured at 4.2%.
Entered for the defence · Paredes et al., Graphite research (Five Percent), 2025 · third-party study
§ VThe finding
A vendor study: 86% of articles ranking in Google and 82% cited by ChatGPT and Perplexity were human-written. Published is not the same as read. Detector-based.
§ VIQuestions for the court
About 14% of articles ranking in Google Search in June 2025, according to Graphite (Paredes, Druck and Smith, October 2025), a vendor study based on an AI detector. 86% were written by humans.
Less often than human ones: 82% of the articles each cited were human-written and 18% AI-generated, in the same study.
That figure (case DWT-006) is about what gets published. This one is about what gets shown to readers. Both come from Graphite and both rely on detectors.
Because the theory is about the writing people meet on the web, and on this evidence most of what search and answer engines surface is still human-written. It is a vendor study, so a point, not a knockout.