A coding recently told us it had avoided a “grep-and-hope” fix. We'd never heard the phrase, and didn't need to have. It means what it says: search for the pattern, change what matches, hope nothing else depended on it. The model had coined it on the spot from two familiar parts, and the result was instantly clear.
A few days later another agent said it would use “the
pickaxe” to find where a string had been introduced. That one
isn't coined. Git's own documentation calls the machinery behind
git log -S and -G the pickaxe: search the
history for commits that added or removed a string. It's real
jargon, just obscure enough that most people who use the flag have
never seen the name.
Those two words point at something worth separating out. Vocabulary is moving in two directions at once.
Direction one: our jargon into their vocabulary
Models learned to talk from text that people wrote, and a large share of the technical text on the internet is programmers talking to each other. So models picked up the profession's slang along with its syntax, and they use it more fluently than many programmers do. A partial list of what we see often:
- Yak shaving: a chain of prerequisite tasks between you and the one you meant to do.
- Bikeshedding: arguing over the trivial part because it's the part everyone understands.
- Rubber-ducking: explaining a problem aloud until you find the bug yourself.
- Cargo culting: copying a pattern's form without knowing why it works.
- Chesterton's fence: don't remove something until you know why it was put there.
- Footgun: a design that makes it easy to hurt yourself.
- Blast radius and surface area: how far a change reaches, how much is exposed.
- Seam, shim, escape hatch: places to plug in, bridge, or bypass the normal rules.
- Bit rot and drift: things quietly decaying, or diverging from what they should match.
None of these is new. What's new is how often a model reaches for them, and how exactly. “Chesterton's fence” in a code review is a precise instruction: find out why this check exists before deleting it. A model that uses it correctly has compressed a paragraph of advice into two words that any experienced reader will unpack.
Coining versus borrowing
“Grep-and-hope” is different in kind from “pickaxe.” Borrowing retrieves a word that already has a meaning. Coining assembles one from parts whose meanings the reader already has. The model is doing what people do with language (compounding, analogizing, naming a thing that needed a name), and the test of a good coinage is the same for both: you understand it the first time you hear it.
There's a small lesson in that. A coined term has no dictionary entry to point at. Its meaning is carried entirely by its parts and its context. When a model coins a word that lands, it's evidence that it's working with meaning, not just retrieving strings.
Direction two: their habits into our writing
The other direction is newer and stranger. Words and constructions that models overuse are showing up more often in human writing.
The best-measured case is “delve.” A study of more than fifteen million biomedical abstracts, published in Science Advances, found that the vocabulary of scientific writing shifted abruptly once language models became widely available. The shift wasn't toward new topics. It was toward style words: verbs and adjectives models favor, with “delve” showing the largest excess. The authors estimate that at least 13.5% of 2024 abstracts were processed with a language model.
Some of that is people using models to edit their writing. But the habits also spread by exposure: read enough model-written text and its rhythms start to sound normal. Other recognizable patterns:
- “It's not just X, it's Y.” A construction models reach for so often that people now mock it on sight.
- The em dash. Models use it heavily, and some writers who have always used it now edit it out of their own prose, because it has started to read as a tell. (This site's earlier posts use plenty. We were here first.)
- Borrowed technical metaphors. “I don't have the for this today.” “You're hallucinating.” “That meeting burned a lot of .” AI vocabulary, used about people.
- “Vibe coding.” Coined in early 2025 for building software by describing what you want and accepting what the model writes, it spread from a single post into general tech vocabulary within months.
Why this is more than a curiosity
A shared vocabulary is a shared set of concepts. When a model says “blast radius,” it isn't decorating a sentence; it's telling you which consideration it weighed. That makes jargon a window into what a model is attending to, and a way to steer it: ask for the Chesterton's-fence check by name and you get a different review than if you ask for “any issues.”
The risk runs the other way. Style words are cheap, and a text that sounds like careful reasoning because it uses careful-reasoning vocabulary may not contain any. “Delve” is harmless on its own; the problem is that it arrives as part of a register that signals thoroughness without guaranteeing it. That's an signal (the words you can count) standing in for an intensional one (whether the thinking happened).
So we find ourselves applying a simple rule to our own writing and to what our agents write: judge a phrase by what it lets the reader do. “Grep-and-hope” passes; it names a specific failure and tells you what to avoid. “Delve into the nuances” fails; it announces depth and supplies none.
A vocabulary in common
Language has always moved between groups that work closely together, and people now work with models all day. It would be stranger if the words stayed put. The useful response isn't to ban the model's favorite words or to adopt them uncritically. It's to notice which ones carry meaning and which only carry style, and to keep the first kind.