I found a tic in my own writing that I had read past a hundred times. Reading was never going to catch it. I caught it by counting.
One word had quietly become my only tool for one move. Whenever I wanted to say that a thing's surface and its interior have come apart, and that the surface was borrowed from somewhere else, I reached for the same figure: the real thing dressed in the look of a false one. A metric with no generating idea beneath it. A profit-and-loss problem styled to sound like vision. A hunger for status routed through ten stated values. The operation recurs because it is most of what I do, and the word I used to perform it had stopped changing.
I had nodded at every one of those sentences. Each is well-made. "A profit-and-loss problem styled as vision" is a good line, and so is the next one, and the one after that. The trouble is invisible inside any single sentence, because it lives one level up, in the distribution over sentences. Per-sentence judgment, the faculty every writer leans on, samples one point at a time, and at one point there is nothing wrong. You cannot taste a frequency.
This is where a semantic tic parts company with the kind a writer knows how to catch. A lexical tic gives itself away locally. A word like "delve," the reflexive em-dash, a stock phrase sitting flat under a weak noun: each looks wrong right where it lands, so taste flinches and a regular expression catches the string on sight. A repeated metaphor hides in plain view. Every instance is locally earned, the surface wording shifts from one to the next, and the move underneath stays fixed. No single page holds the evidence. Only the count does.
The repetition costs the reader real information, and naming that cost takes a short detour through compression. Good prose minimizes description length two ways at once: it keeps each code short, and it gives distinct meanings distinct codes. A pet metaphor wins the first prize and forfeits the second. It hands one short code to a crowd of different meanings. Devotion masking a financial edge, an ordinal pressed flat into a pair, a feature posing as a thesis: these differ in exactly what is borrowed and where the seam shows, and the shared code asserts they are alike. The reader's model arrives pre-compressed to one shape, and the specific texture of each fakeness, the part that would have changed his understanding, is the first thing the shared code throws away.
Competence is what lets the tic live. Flat writing dies at review, because reviewers catch flatness. A well-made figure clears the bar every time, so it accumulates under the protection of being good. The gate that should stop it is the very thing waving it through. That is why taste cannot be the fix: taste is the sense that went blind. What works is a constraint taste cannot argue with. Ban the word.
The strongest objection is that several uses of the word genuinely earn their place, and a few turn the figure into an argument rather than reaching for it. That objection is correct, and conceding it is how the argument holds. Those good instances are the camouflage the lazy ones ride in under. If every instance were weak, the cluster would have tripped some reviewer long ago. The good ones are the reason it never did.
A hard ban turns out to be generative. Deleting the word leaves the move untouched; it still arrives, on schedule, and the cached answer is gone. Now I have to derive the gap between surface and interior from the particulars in front of me, which forces a fresh look at what exactly is borrowed and where the borrowing gives itself away. The cached word let me skip that look. The ban removes the shortcut, and the only road left runs through the specifics. It gives back the meanings the one short code had been compressing away.
The word I had overused meant, precisely, a fake thing worn in the shape of a real one. It was itself a fake thing worn in the shape of a real one: a stored reflex standing in the spot where a fresh choice belonged, posing as the choice. My vehicle for counterfeit was the one piece of counterfeit I never thought to check, because it passed every reading I gave it. It only ever failed the count.