How fast did LLM writing adoption actually spread?

The authors measure LLM-assisted writing in four domains from January 2022 to September 2024: 687,241 consumer complaints, 537,413 corporate press releases, 304.3 million job postings and 15,919 United Nations press releases. Their central finding is that "LLM usage surged following the release of ChatGPT in November 2022," and that by 2024 "growth appears to have stabilized." The shares are estimates of text that "appears to be LLM-assisted," not verdicts on individual documents: roughly 18% of financial consumer complaint text, up to 24% of corporate press release text, just below 10% of job posting text in small firms, and nearly 14% of UN press release content.

The estimates come from what the excerpt calls a "robust population-level statistical framework" that the authors say they validated in earlier work. The excerpt credits it with "superior robustness, transparency (and lower cost)" than commercial AI content detectors, but it does not describe how the framework works, so the method has to be taken on the authors' validation. The discussion gives the shape of the curve: "after an initial lag of 3–4 months following the ChatGPT launch, there was a sharp surge," followed by stabilization. Adoption also varied by organization. The authors name organizational age and size as "the most important predictor of differential adoption," with smaller and younger firms using LLMs more.

Set against the neighboring notes, this excerpt is a population measure of output, and it sits apart from process-level work. Can process data distinguish AI delegation from ordinary collaboration? looks at when contributions arrive within an individual's work. That is a distinction this excerpt does not attempt: it counts how much text is LLM-modified, not whether a writer delegated it or collaborated. The two methods answer different questions and would be worth comparing on the same domain. Where have workers actually delegated tasks to AI? also locates AI use by domain, but from committed agent skills on GitHub rather than from text that reaches readers. The plateau raises its own open question, carried in Is the 2024 LLM writing plateau real saturation or measurement artifact?.

The excerpt does not establish that the shares are accurate in absolute terms. It points to the framework's validation without giving error rates, and its own figures disagree in places. The abstract gives "just below 10%" for small-firm job postings, while the introduction gives "up to 15% for young and small companies job postings." The abstract dates stabilization to 2024, while the discussion says "late 2023." The authors say they "did not directly measure homogenization," and the excerpt says "further study is needed" on whether higher adoption yields better consumer outcomes. The implication is narrower than the headline: adoption was broad and fast across these domains, and the size of each share is a framework-dependent estimate rather than a fact about each text.