Empirical evidence of Large Language Models' influence on human spoken communication
Yakura, Lopez-Lopez, Brinkmann, de la Serna, Kirfel, Gupta, Soraperra, Eisenmann, Wulff & Rahwan. arXiv:2409.01754 (cs.CY). First posted September 2024; a v4 revision followed in July 2026. Cited by the-joy-of-writing for the claim that people pick up AI-favored vocabulary in their own speech — the piece's strongest citation, and undersold by it.
What it argues
Chatbots are a new communication medium — the authors' comparison is to the printing press and social media — that absorb patterns from the human text they were trained on and feed them back to hundreds of millions of users. Word choice in spoken, unscripted language is used as an early, measurable signal of that feedback loop, on the logic that a chatbot's cultural influence is otherwise too diffuse to see directly.
Method — two studies, not one
- Observational. A synthetic-control analysis over 737,083 hours of conversation across 824,634 podcast episodes, screened to unscripted speech, tracking the frequency of words ChatGPT favors (delve, showcase, boast, intricacies, meticulous) before and after its release. The rise in these words is tied causally to the release date by the synthetic-control design, not just correlated with it.
- Experimental. A preregistered study, N = 496, in which participants had a short conversation with a chatbot, completed a distractor task, then took a forced-choice lexical test. Participants who had talked to the chatbot picked its words afterward — and the effect survived the distractor task and showed up in a forced choice, which the authors read as the words entering active vocabulary rather than staying as passing mimicry.
Why this is the load-bearing citation in the-joy-of-writing
Danckert's third argument — that reliance on AI stunts a writer's own voice — leans on this paper for one line about "buzzwords." What the paper actually shows is stronger and more specific: exposure to a chatbot changes what words a person reaches for afterward, unprompted, in an unrelated task — a tested claim about linguistic influence, not an inference from stylistic resemblance. It is also the one citation in the source article this ingest could read and check directly, rather than take from an abstract or a record ID.
What it doesn't establish
It measures word choice, not the fluency-vs-effort tradeoff Danckert's broader argument rests on, and it says nothing about whether the borrowed vocabulary makes writing worse — only that it is borrowed. The authors' own concern is macro-level: a handful of AI providers exerting outsized, largely invisible influence over how people talk, which is a claim about linguistic homogenization at the population level, the same word Bieleke & Wolff use for writing in the article this page supports.
Related
the-joy-of-writing · cognitive-miser · james-danckert · sapir-whorf-hypothesis