People are using the word “delve” far more than they did a few years ago, and the timing points straight at ChatGPT. Nobody taught them the word in school. They picked it up from the chatbot without noticing, and the pattern is strong enough to measure. Researchers tracked the words ChatGPT overuses, then counted those same words across hundreds of thousands of podcasts and recorded talks. The jump came right after the tool went live.

The etymologist Adam Aleksic sees this one word as a warning sign for something much bigger. His argument runs in two steps. First, the tools you type into and scroll through are steadily changing the way you talk. Second, the way you talk feeds back into the way you think. In his TED talk, he puts it plainly: these are not neutral tools, and you should treat them that way.

Start with why a chatbot fixates on a single word. ChatGPT uses “delve” far more often than most English speakers do, and Aleksic offers a likely reason. Part of its training was handled by workers in Nigeria, where the word is common in daily speech. Their habit slipped into the system, and the model then leaned on the word even harder than those workers ever would. That happens because a model copies whatever it sees most. A small quirk, repeated across enough text, becomes the way the machine talks by default.

From there, the loop starts to feed itself. You read the chatbot’s wording, you start using a little of it yourself, and you type it into your own emails and posts. That writing then becomes training material for the next version of the model, which serves the same wording back to the next round of users. The machine learns from us, we copy the machine, and the machine learns from us again. With every loop the effect grows, and ordinary speech drifts closer to the way the software writes.

The words are only the visible part. The same force that shapes your vocabulary also decides what fills your feed, and that is the bigger problem. Posts that make people angry travel further than calm ones, because outrage earns clicks. So the version of the world on your screen ends up louder and more extreme than the world outside. You only ever see the posts that did well, never the ones that flopped, which makes the extreme version look like the whole story. Spend enough time there and it starts to feel normal.

None of this reaches you untouched. Every post and every chatbot answer is filtered first, sorted by what holds your attention, what makes the company money, and its rough guess about who you are. That guess is crude. The platform builds a thin sketch of you from your clicks, then keeps feeding you whatever fits that sketch, and slowly you shape yourself to match it. No feed captures the full range of a real person, yet the picture it serves still passes for reality.

The deepest effect lands on how you think, and it travels through the words you use. Borrow a machine’s words and you slowly borrow its sense of what is worth saying. The same goes for letting TikTok decide what you watch or letting Spotify decide what you hear. Use their words, follow their suggestions, and their version of the world quietly becomes yours. As more people pick up the same phrasing, the small quirks that make voices distinct flatten out, and the shift is slow enough to feel like your own taste rather than something done to you.

Aleksic’s defense is a single question, asked often. Why am I seeing this? Why am I saying this? Why am I thinking this? And why does the platform want me to? He pairs the question with simple habits: read beyond your usual sources, talk to people who do not sound like your feed, and notice the words you reach for before they turn automatic. Skip that, and the platform ends up writing your version of reality for you.

The research is still fresh. These are among the first studies to track spoken language this way, and the explanation for where the word started remains a best guess. The direction, though, is hard to dispute. The way people talk already leans toward the machine, and the more we assume these tools are neutral, the less we notice the pull.

Sources

TED, Why are people starting to sound like ChatGPT? – ted.com
arXiv, Empirical evidence of Large Language Model’s influence on human spoken communication – arxiv.org