People started saying "delve". Then they stopped.

This one’s odd, because it’s about speech, not writing, and that matters. You cannot copy and paste into a spontaneous conversation. If the machine’s words turn up there, somebody absorbed them.

The paper is “Empirical evidence of Large Language Model’s influence on human spoken communication”, an arXiv preprint, revised July 2026, not peer reviewed. The authors are Yakura, Lopez-Lopez, Brinkmann, de la Serna, Kirfel, Gupta, Soraperra, Eisenmann, Wulff and Rahwan, of the Max Planck Institute for Human Development.

The method has three parts. First, 737,083 hours of conversation from 824,634 podcast episodes, screened for unscripted speech. Second, a synthetic-control comparison against words ChatGPT does not favour. Third, a preregistered experiment with 496 people.

For the prosecution: words ChatGPT prefers (delve, showcase, boast, intricacies, meticulous) increased abruptly in spontaneous speech. In Science and Technology podcasts, “delve” rose to about 44% above its expected level 13 to 18 months after ChatGPT’s release. And in the experiment, a short chat with a bot that used particular words led people to use those words themselves after a distraction task. So the bot’s vocabulary got into people. That’s the claim, and the evidence is decent.

For the defence, and I’ll quote the authors here: “the elevation did not persist”. Use of “delve” peaked around mid-2024, then fell below its expected level. It sat 15% below in Science and Technology, 30% below in Business, and 35% below across all categories. The turn coincides with the free release of GPT-4o and growing public awareness of “delve” as a sign of AI language. The authors also note “delve” may become stigmatised. Which is a strange thing to say about a word, but there it is.

The point for the court is simple. The machine’s words do get into people, and once people notice them, they back away. Both halves are in the record.

Now the caveats, because I’m bound to read them out. It’s a preprint. It covers English podcasts only, a self-selected public-facing group, so don’t stretch it to everyone’s kitchen table. And lexical shifts are not a measure of sameness: a handful of favoured words rising and falling tells you about those words, not about whether speech as a whole became more alike. This is neither a clean prevalence study nor a clean convergence study. It’s about contagion, and then about retreat.

The prosecution gets its contagion. The defence gets its retreat.

Verdict: split.