How AI Affects How We All Talk

In July 2026, Hiromu Yakura and colleagues updated a study of spoken English that had first appeared in 2024, now drawing on 737,083 hours of conversation from 824,634 podcast episodes. After chatbots came into wide use at the end of 2022, certain words rose in spontaneous human speech: delve, showcase, boast, intricacies, meticulous. These are words the researchers had identified as typical of chatbot writing. In a preregistered experiment with 496 people, a brief conversation with a chatbot was enough for participants to start using its words themselves. Writing shows the same drift, and it showed it first. Dmitry Kobak and colleagues, writing in Science Advances in July 2025, tracked vocabulary across more than 15 million biomedical abstracts published between 2010 and 2024. They estimated that at least 13.5 per cent of 2024 abstracts had been processed with a language model, rising to 40 per cent in some groups of papers, and that the shift in vocabulary was larger than the one caused by the Covid pandemic.

For two decades the machines that mattered most to public conversation were recommenders. They chose what you saw: which post, which video, which headline rose to the top. The worry they raised was the filter bubble, a narrowed slice of the world. Generative tools sit somewhere new. They do not only choose what reaches you; they suggest what you say, in the draft, the reply and the rewritten email. That makes them a different kind of gatekeeper, standing between a thought and its wording. Our previous issue looked at the opposite pressure, people changing their words to get past the machines that judge them; this one looks at people taking on the machines' words. The effect goes beyond vocabulary. Maurice Jakesch and colleagues (CHI 2023) asked 1,506 people to write about whether social media is good for society, some of them helped by a writing assistant set to argue one side. The assistant's view turned up in what they wrote and in their own answers to a survey afterwards: writing with it shifted what they said they believed. Jess Hohenstein and colleagues (Scientific Reports, 2023) found that suggested replies made conversations faster and warmer, and partners rated each other as closer, unless one suspected the other was using them, in which case they rated that person as less cooperative. For a measurement lab this matters in a specific way. Any test that compares a machine with how people write or talk assumes the human side stays put. If people drift toward the machines, the baseline moves, and the gap between the two can narrow without the machine changing at all.

Three things. First, whether the drift lasts. Words that spread through novelty can fade, and these studies measure the first few years; the next few will show whether chatbot vocabulary settles into everyday English or passes like other fashions. Second, how much of daily communication runs through a suggestion. Reply suggestions, rewrite buttons and summary tools are being built into email, messaging and office software, and each one adds a moment where a machine offers the words before a person does. Third, disclosure. The suggested-reply result points to a tension: AI help can make a conversation warmer, but people trust each other less when they think it is being used. Whether products tell the other person that a message was machine-assisted may decide which of those effects wins. Our part is the machine side of the conversation: measuring, under fixed conditions, what models do, so that a change can be traced to the model rather than to the people judging it.