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
Right, first witness. A team at Stanford (Liang, Zhang, Codreanu, Wang, Cao and Zou) published a paper in Patterns in 2025, peer reviewed, called “The widespread adoption of large language model-assisted writing across society”. Mouthful. The finding is simpler than the title.
Here’s the method, and it matters. They didn’t run an AI detector over each document and count the guilty ones. They built a population-level statistical model of word distributions, then asked how much of the whole pile looks like it involved an LLM. The pile was big: 687,241 consumer complaints, 537,413 corporate press releases, 304.3 million job postings and 15,919 UN press releases, running from January 2022 to September 2024.
So what did they find? By late 2024, roughly 18% of financial consumer complaints involved LLM-assisted writing. For corporate press releases it was 24%. Nearly 10% of job postings in small firms, and 14% of UN press releases. Newswire press releases peaked at 24.3% in December 2023 and stabilised at 23.8% through September 2024.
If you write press releases for a living, that’s roughly a quarter of the copy landing on the wire next to yours.
Growth appears to have levelled off by 2024. The authors say that might be saturation, or it might be that more advanced models are simply harder to spot. They don’t claim to know which, and I like them for it.
Now the caveat, and it’s a big one. This study measures adoption. It does not measure whether the writing became more alike. Lots of people using a tool is one thing; lots of text sounding identical is another, and the data stops in September 2024 anyway. So the court can’t convict on sameness here.
But as evidence that the machine is in the room, and in a lot of rooms, it’s hard to argue with.
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
24%
of corporate press releases, late 2024
537,413 corporate press releases, January 2022 to September 2024. A population-level estimate from word distributions, not a per-document detector.
Entered for the prosecution · Liang et al., Patterns 6(12), 2025 · third-party study
18%
of financial consumer complaints, late 2024
Roughly 18% of 687,241 consumer complaints. Also nearly 10% of job postings in small firms and 14% of UN press releases.
Entered for the prosecution · Liang et al., Patterns 6(12), 2025 · third-party study
Evidence against the theory
No exhibit entered for the defence in this case.
§ VThe finding
By late 2024, roughly 24% of corporate press releases involved LLM-assisted writing. The study measures adoption, not sameness, but the machine is in a lot of rooms.
§ VIQuestions for the court
By late 2024, roughly 24% of corporate press releases involved LLM-assisted writing, according to Liang et al. (Stanford, Patterns 2025). The figure was roughly 18% for financial consumer complaints, nearly 10% for job postings in small firms and 14% for UN press releases.
No. Newswire press releases peaked at 24.3% in December 2023 and stabilised at 23.8% through September 2024. The authors say growth appears to have stabilised by 2024, through saturation or because more advanced models are harder to detect.
No. The study measures how much text is LLM-assisted, not whether it became more alike. It is evidence that the machine is widely used, which is why it is entered for the prosecution, but it cannot convict on sameness.
With a population-level statistical model of word distributions, not a per-document AI detector, over 687,241 consumer complaints, 537,413 corporate press releases, 304.3 million job postings and 15,919 UN press releases from January 2022 to September 2024. It is a third-party study, not ours.