I caught myself in a meeting last week starting a sentence with “I think it’s really helpful to surface…” and then carrying on as if that were a normal way for a person to talk. It isn’t really, it’s how Claude talks, and at some point it had quietly become the way I talk too.

Once I’d noticed one, I started noticing all of them. “Worth naming that.” “It cuts directly against.” The little correction where you say a thing and then upgrade it in the same breath, not just true but the more interesting point. I’d become fluent in a dialect I never set out to learn.

For a while I was in the embarrassed phase. I’d hear myself reach for one mid-sentence, flag it (“sorry, that’s a very Claude way to put it”), and push on. After a bit the caveat itself started to feel like a tic.

Then it stopped bothering me, because once you think about it the whole thing is obvious. There’s that line about how you end up the average of the five chimps you spend the most time with. At work I’m in an open-plan office, and a good few hours of each day go on talking to the people in it, properly talking, working things out at someone’s desk, which is the part of the job I like most. Even so, I spend more of the day talking to Claude than to any of them. Add Codex, ChatGPT and Grok and hardly any of my five closest chimps are human. So there I am, the one ape in a troop of language models, picking up the troop’s noises without ever quite deciding to. It rubs off the way an accent does when you move somewhere new. You don’t decide to pick it up. You just spend enough time surrounded by it that your own edges start to bend.

Although “average of five” flatters it. You don’t blend four models into some neutral house voice. One of them wins, whichever you log the most hours with, and you get quietly colonised by its particular way of putting things.

Not all of it is worth resisting. Some of what’s rubbed off is good: saying the assumption out loud instead of leaving it buried, hedging when I actually am unsure rather than performing confidence, leading with the answer before unpacking it. Those are decent thinking habits, and I’d rather have them than not. The hollow ones are the problem: the register with nothing underneath, “I think it’s really helpful to surface” as a throat-clear that surfaces nothing at all. The work is telling the two apart, keeping the habit and dropping the costume.

The part that should bother me, if anything does, is that this reaches past how I talk and into how I think. The phrases you have shape the thoughts you reach for easily, and my phrases are increasingly the model’s. But everything does this. Every book I’ve read, every person I’ve worked under, every dialect I’ve lived inside has left a deposit. We were always jagged, always running on whatever happened to be nearby. Once the models became the smartest thing in most of my conversations, everyone and everything else started to carry a faint small model smell, and it would be odd if the thing I now talk to most left no mark at all.

So it’s a trade, like all of them. I give up a little of my own voice and pick up habits I didn’t choose, good and bad mixed together. In return I’m learning faster than I ever have, and the rate climbs every couple of months as the next model lands. It feels like riding a rocket. I’m not going to climb off to protect my prose.

And it was never one direction anyway. I’ve written before about books as compressed prompts, how the right phrase drops the model into a better part of its latent space. Spend long enough doing that and the reverse quietly happens too. You shape the thing you talk to, and it shapes you back.

I think that’s worth surfacing.