Here is an interesting study by Yakura et al on how AI is influencing our linguistic output. The authors present what they describe as the first large-scale causal evidence that an AI model’s language patterns are surfacing in spontaneous human speech. People are saying these words out loud, in talks and in conversations, more often than they did before ChatGPT existed.
Yakura et al. analyzed 740,249 hours of human discourse, drawn from 360,445 academic talks on YouTube and 771,591 podcast episodes, all transcribed in-house into over 7.35 billion words. They didn’t trust the platform captions. The team ran its own transcription with WhisperX, partly because YouTube had swapped its transcription models around 2020 and started spelling out filler words like “um,” which would have muddied the signal they were chasing.
Yakura et al. borrowed the synthetic control technique, the same approach economists used to estimate the effect of California’s tobacco control program. The logic is to build a counterfactual, a model of what word usage would look like in a world where ChatGPT never launched, then measure the gap between that imagined world and the real one.
To define an “AI word,” the team had GPT models edit thousands of pre-ChatGPT human texts and measured which words the model reached for more often than the human original. The overused ones became the suspects: “delve,” “underscore,” “comprehend,” “bolster,” “boast,” “swift,” “meticulous,” “pinpoint.”
The headline result lands on “delve.” Its use in academic talks climbed significantly after launch, and the effect was strongest exactly when the team set the treatment date to November 30, 2022, the real release date. That date-matching is the hidden engine of the whole argument. A correlation wouldn’t care which day you picked. A causal effect should peak at the true one, and it did. Yakura et al. report that many of the top GPT words show an annual growth in usage of roughly 25 to 50 percent.
The Podcast Test That Makes the Causal Story Hold
I came into this paper ready with the obvious objection. People read scripts. If a YouTube lecturer drafts a talk in ChatGPT and reads it aloud, of course the AI’s vocabulary shows up. That’s contamination, not influence.
Yakura et al. saw it coming. They filtered the podcast set down to genuine back-and-forth conversation, episodes with two or more speakers and at least four exchanges, the kind of talk you can’t pre-write. “Delve” still rose significantly in Science and Technology, in Business, and in Education.
It didn’t move in Sports or in Religion and Spirituality. That split is the part I keep thinking about. The word climbed fastest in the fields where people spend their days near ChatGPT, then started filtering outward. Yakura et al. read this as a two-phase diffusion, where AI-preferred words land first in heavy-exposure domains and seep into casual talk afterward.
The Feedback Loop and the Homogenization Risk
The idea that gives the paper its weight is the feedback loop. Machines were trained on human culture. They developed their own quirks along the way. And now humans are picking those quirks back up. Yakura et al. frame this as a bidirectional human-machine cultural loop, building on Brinkmann and colleagues’ work on machine culture.
They put it plainly, writing that “machines trained on human culture are now generating cultural traits that humans adopt.” We taught the machine to talk like us, and it’s started teaching us to talk like it.
The risk they name is homogenization. If AI keeps favoring a narrow band of words and phrasings, and humans keep absorbing them, linguistic and cultural variety could thin out. It compounds, because tomorrow’s models will train on data already shaped by today’s AI and amplified by human adoption. Yakura et al. raise the specter of model collapse arriving through a new door. The threat isn’t corrupted data. It’s a culture that has converged on the machine’s preferences.
This connects directly to what Sourati and colleagues (2026) found about the homogenizing effect of LLMs on human thought, and to Abdulhai et al. (2026) on how these models distort written language.
Yakura et al. extend that line into spoken communication, and that’s the part that unsettles me. Writing is deliberate. You can edit a tell out of an essay before anyone sees it. Speech is closer to the bone. If the vocabulary is shifting there, the influence has reached a layer we don’t usually supervise.
I’ll give the authors credit for being open about what they can’t yet explain. Yakura et al. are clear that the mechanism is unresolved. Adoption could be plain imitation. It could be cognitive ease, the human pull toward whatever feels fluent. Or it could be something deeper.
They float a genuinely uncomfortable possibility, that if AI-preferred words carry AI-preferred reasoning patterns, then borrowing the vocabulary might mean borrowing the cognition too. They call it speculative, and it is. But it rhymes with what Shaw and Nave (2026) describe as cognitive surrender, the slow handing-over of our thinking to the machine. Vocabulary might be the first thing to go, and the easiest to miss.
How AI Is Changing Language, and What AI Literacy Should Do About It
So what should teachers do with this? Panic isn’t the answer. The point isn’t that “delve” is dangerous. It’s that AI shapes us in ways that don’t announce themselves, and noticing that pull is a literacy skill in its own right.
I’d put it to students directly. Track the words. Show them the list Yakura et al. built and ask them to listen for those words in their own speech, in their feeds, in the videos they watch. Ask them where the words came from and whether they chose them on purpose. That small act of attention is the muscle AI literacy is supposed to build, the habit of seeing the tool’s hand in your own output.
There’s a final thread here. Yakura et al. note that ChatGPT’s pull toward “delve” actually weakened across model generations, from an odds ratio past 300 to 1 in the early versions down to roughly 40 to 1 in GPT-4o, as OpenAI sanded the artifact down. By the time they smoothed it out, the word was already loose in human speech. The company can patch the model. It can’t recall the word.
Language has always told other people who we are. Yakura et al. speculate that an AI-flavored word could someday read as a marker of borrowed authority, a sign that someone outsourced their thinking. That’s plausible.
What I’m more certain of is simpler. The machine is no longer just answering us. It’s starting to sound like us, and we’re starting to sound like it. The question Yakura et al. leave us with is the right one, that “our next question is no longer whether machines influence us, but how profoundly and through which channels.”
References
- Abdulhai, M., White, I., Wan, Y., Qureshi, I., Leibo, J., Kleiman-Weiner, M., & Jaques, N. (2026). How LLMs distort our written language. arXiv preprint arXiv:2603.18161. https://arxiv.org/abs/2603.18161
- Sourati, Z., Ziabari, A. S., & Dehghani, M. (2026). The homogenizing effect of large language models on human expression and thought. Trends in Cognitive Sciences. Advance online publication. https://doi.org/10.1016/j.tics.2026.01.003
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646
- Yakura, H., Lopez-Lopez, E., Brinkmann, L., Serna, I., Gupta, P., Soraperra, I., & Rahwan, I. (2025). Empirical evidence of Large Language Model’s influence on human spoken communication (Version 3) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2409.01754