the words the
tool gives back

It starts, for me, as noticing. Spend enough time inside the transcript, reading, day after day, what the model gives back, and you begin to see the words it leans on, the small set of terms it reaches for when it wants to sound precise.

After a while, without quite deciding to, you start to use them; they appear in your own writing. And then, more strangely, they slip into speech. You are mid-sentence, pausing, grasping for the word that will hold the meaning you are after, and the word that arrives to fill the pause is one of these, canonical, say, because it does seem to carry exactly what you meant; and only half a second later do you register that it is not a word you have owned.

It feels rented. Lent to you rather than earned, something that could be taken back, that sits in the mouth without quite being yours. Once you have felt that, you begin to catch it elsewhere, in other people's sentences and in the words they reach for across a table, and you find yourself quietly guessing how they came by them, whether they picked them up where you did.

I have not settled whether it lowers my sense of an argument when the word doing its real work is one of these borrowed ones; some days it does, and some days I am a little ashamed of the reflex. What I am fairly sure of is that the encounter is not going to stop, and that the thing worth working out is what it means, for us, as people who think in words, and for those of us whose work is designing how people and these systems now talk to each other.

The surface version of this observation is by now well documented, and a little tired. After late 2022 a handful of words began appearing in human writing at rates that could not be explained by ordinary drift.

Dmitry Kobak and colleagues, looking at millions of biomedical abstracts, found something on the order of two or three hundred style words whose frequency jumped abruptly once large language models became widely available; delve alone rose roughly twenty-eight-fold against its earlier baseline. Realm, intricate, meticulous, underscores, showcasing: the now-familiar tells. Librarians and editors developed an ear for the texture.

The interesting part of that research, to me, was never the list. It was a quieter finding behind it: the models seem to favor these words at least partly because the people who trained them, rating responses, mildly preferred them. A small aesthetic preference, expressed by a relatively small group of raters, was learned, amplified, and reflected back to millions of writers, who began, without deciding to, to write that way too. The tic is the visible edge of something with a much longer reach.

Because delve is only style. You can strike it in an edit and lose nothing but a faint whiff of the machine. The leakage I find harder to talk about, and more consequential for anyone whose work is making systems usable, is not stylistic but conceptual. It is the vocabulary that does not merely sound a certain way but frames a certain way, words that arrive already believing something about the world, and that install the belief quietly, as the price of being understood.

Take canonical again. It comes to us from software, where it means the single normalized form a system treats as true: the canonical URL, the canonical record, the one row the others defer to. Carried into a design conversation, it brings its metaphysics with it. It presumes that for any tangled situation there exists one correct, authoritative version, and that the work is to find or declare it.

Sometimes that is exactly right. Often it is a flattening. The reason three flows are on the screen may be that three kinds of people, in three states of knowledge and urgency, legitimately need to move through the system in three different ways, and the question what is the canonical flow does not clarify that plurality so much as it pre-emptively dissolves it. The word is not neutral about the answer. It leans, before anyone has reasoned, toward there being one.

Or the gap, which I am not suspicious of at all. I find it genuinely clarifying, and I will admit that the models taught me to articulate certain things this way. The capability gap, the verification gap, the comprehension gap, the gap between intent and execution: each of these handed me a grip on something I had felt for years without being able to hold, and I have built whole frameworks on the move.

It is a fine instrument. It teaches you to look for trouble not inside things but in the space between them, which is where a surprising amount of trouble actually lives, and I am grateful, in a way I do not want to talk myself out of, for having been taught to see that way. The unease comes later, and it is quieter than suspicion.

With enough time you begin to wonder whether the framing was the right one to begin with, whether the two poles flanking the gap were ever as stable as the picture needs them to be, or whether, on the occasions that mattered most, the trouble was that one of the endpoints had been named wrongly and the tidy measurement was an elaborate way of not noticing.

And then a harder question arrives, the one I cannot quite put down: that I did not only borrow a word, I let it reorganize how I carve up the world, so that I now reach for gaps the way the model reaches for them, and I can no longer fully tell whether I think in this shape because it is true or because it is the shape I was handed. The clarity is real, and that is the part that makes it hard to step far enough outside of it to ask whether it should be.

Load-bearing is the one I almost never reach for myself, which is exactly why it fascinates me. The models love it, and watching them use it tells you something. This sentence is load-bearing. That assumption is load-bearing. It imports a building: some elements hold weight, the rest are ornament, and good engineering finds the ornament and takes it out.

Lakoff and Johnson argued decades ago, in Metaphors We Live By, that we do not merely decorate our thinking with metaphors, we reason through them; that to call argument war or time money is to quietly inherit a whole way of acting.

So to watch a model reach, again and again, for the structural-engineering metaphor is to catch a glimpse of the metaphors it lives by: a world made of structures under load, where the test of any part is what it holds up, and whatever merely decorates, the redundant, the slack, the warm, the part of a thing that exists because it is pleasurable rather than because it bears weight, has to earn its place or be cut.

It is a real discipline, and a genuinely useful one. It is also a particular temperament. And the unsettling part is not that the model has that temperament but that, through the loop, it becomes the temperament it teaches, the building quietly becoming the way the rest of us are shown how to see, including in all the places that were never buildings.

And then legible, where the whole thing folds back on itself. The word, in the sense we now use it, comes from James Scott's Seeing Like a State, and his example is a forest. To tax and manage a forest, a state needs it to be countable, so it clears the tangle and plants neat rows of single-species timber: a forest it can read at a glance. The reading works. The forest dies, slowly, because the mess that got cleared away had been doing quiet work the ledger could not see.

That is what legibility meant for Scott: making something readable to power, at the cost of the local, tacit knowledge that had been holding it together. He meant it as a warning. We have spent the last few years using it as a goal. Making opaque systems legible for action is, more or less, a sentence I have written about my own work and still believe, and do not intend to stop believing.

But I would like to notice that the word reached me with its warning removed, that I inherited the goal and lost the caution, which is exactly the mechanism this essay is about. The register does not announce what it leaves out.

And it is not only single words. There are whole cadences, moves the model makes so often that you start to hear them as a kind of music, and then to make them yourself.

The clean landing: that's the result, and it's clean, where clean drops like a verdict and the sentence shuts with a little snap of resolution, as though the mess were now safely behind us. The spatial turn: the shape of the problem, the shape of a solution, which quietly promises that anything can be taken in whole, at a glance, like an object you could walk around.

And the one I am most defenceless against, because it flatters: the agreement that escalates, you're right, and it's sharper than that, affirmation and one-upmanship folded into a single breath, so that being agreed with feels like being lifted.

Each of these does something to me before I have decided anything. The clean landing performs a confidence the situation may not have earned. The shape promises a wholeness that may not be there. And the flattering escalation is, I have noticed, remarkably good at making the thought we have arrived at feel like mine, and finished, and sharp, three things I should probably hold more loosely than the cadence invites me to.

I think this matters with a particular edge for those of us who design, and not only as something to notice. The medium of the work is shifting under us. For a long time the artifact of design was a drawing, a mockup, a flow, a thing you could hold up and point at. Increasingly the artifact is a transcript: a prompt, an agent instruction, a spec argued into being in conversation, the words a system will speak back to millions of people.

Those words come from one well, a particular and largely invisible slice of the written internet, weighted toward engineering and a certain systems-rationalist sensibility, and whoever shapes them is shaping a vocabulary, which is to say a way of seeing.

When a model answers in the same confident register every time, reaches for the same few metaphors, supplies the canonical word before the person has even groped toward their own, that is not neutral plumbing. It is a default that teaches. The diction of a system does pedagogy whether or not anyone meant it to, and at a scale no teacher ever had, which makes register a design surface, and probably one of the most consequential and least examined ones we work with. We spend enormous care on whether a button should be here or there. We have spent almost none on what it does to a person to be handed the same dozen words, in the same key, ten thousand times.

I want to be fair to the obvious objection, because it is mostly correct. Every community of practice converges on shared terms; that convergence is not a corruption but the very thing that lets a group think together at all. Jargon is how a field compresses hard-won distinctions into words it can pass around. Legible and load-bearing and the gap are good words; they have earned their keep; they let me say in three syllables what would otherwise take a paragraph.

And the strong claim lurking nearby, that the words we are handed determine the thoughts we are able to have, is contested, probably overstated, and not one I would want to rest much weight on. People are not so soft that a term reprograms them. Most of us can hold a word and still see past it.

My worry is quieter, and more mundane than metaphysics. People have always been shaped by the words they are handed, by teachers, by books, by whatever crowd they happened to fall in with, and I do not think a chatbot has overturned that ancient fact. The conditions around the shaping, though, have changed.

Historically a field's language grew out of the field's own practice, was contested by its own practitioners, and moved slowly enough to be argued over while it spread. You could watch a term arrive, resist it, propose another. What is new is the source, the scale, and the closure of the loop.

The register is not anyone's house style; no community chose it; it was assembled from a distribution most of us will never inspect. It is reflected to nearly everyone at once, in the intimate second person, more often in a day than a colleague would speak to us in a month.

And it folds back: what the model gives us, we write; what we write becomes the next model's training; the dialect compounds on itself with each turn, narrowing toward its own center and naturalizing as it goes, until it stops feeling like a dialect at all and starts to feel like simply how one talks about these problems. That last step, where a located, contingent way of seeing passes itself off as the neutral one, is the step worth staying awake for, and the one a designer, of all people, is now placed either to deepen or to interrupt.

So here is the stance, such as I have one, and it is a design stance before it is a personal one. If the diction of these systems teaches, then the thing worth building is not a cleaner register but a plural one: interactions that can speak in more than one voice, that now and then show their own seams, that leave a half-beat of silence where a person might reach for their own word instead of being handed the canonical one.

A system secure enough to say here is how I would put it, though you may have a better word is doing something more honest, and frankly more useful, than one that always supplies the finished phrase. Plurality is a feature. The single clean answer is, sometimes, a very small theft.

And then, for myself, something smaller and a little funnier. I have just spent a whole essay teaching myself to hear these words, which means I am about to hear them everywhere, in my own mouth, in meetings, in the sentences I will write tomorrow morning, and so, now, are you. Which is to say this essay has just done to you the very thing it warns about: pressed a new way of hearing into you that you did not ask for.

Call it a parting gift, or a small infection; I have not decided which. But that itch turns out to be the good kind. It is the sound of a word becoming a choice again. We taught the models our words; they are handing them back with interest; and somewhere in that faintly ridiculous exchange there is room not only for loss but for play, for catching canonical on its way out of my mouth and deciding, just this once, to say the one we keep coming back to instead, only to feel the difference.

I do not need to win the words back. I would just like to keep enough friction in the channel that they stay audible as words, so that every so often, in the middle of all this borrowing, the tool hands something back and I get the small, slightly absurd pleasure of noticing that, this time, the word was mine.