In 2005 the New Oxford American Dictionary printed a word that does not exist. Esquivalience, it read: the willful avoidance of one's official responsibilities. The editors had invented it and slipped it in on purpose, one deliberate fault in an otherwise scrupulous book, so that if the word ever surfaced in a rival dictionary they would know it had been copied rather than compiled. Mapmakers do the same with trap streets, a phantom lane drawn into a real city so a forger gives himself away the instant he reproduces it. The error is a signature, planted to fire only in the presence of the wrong kind of reader.
I have been thinking about those planted flaws because the conditions that made them clever have become ordinary. A machine will now produce a clean surface for nothing: correct grammar, even rhythm, the confident finish that used to take a person years to learn. For as long as polish was expensive, a polished thing was evidence, of effort and taste and a human who cared enough to sand it. That inference has quietly stopped working. A flawless paragraph is now equally what a careful writer makes and what a model emits on its first pass, which means flawlessness has gone from a signal to a coin everyone can mint. It proves nothing about who was there.
When a signal goes cheap, the information it used to carry moves to whatever stayed expensive. What stayed expensive is the choice: the specific, slightly wrong thing a person leaves in on purpose. An optimizer's entire job is to remove exactly that, so the part it would remove becomes the part that proves the optimizer was not the last hand on the work. The proof of a human migrates to the place the polishing machine reaches for first.
A chosen flaw earns its keep two ways. As a signature, it marks a present author: the kept flub on the live take, the idiom a copyeditor would strike, the claim sharpened past what a cautious editor would allow, the weaver's deliberate irregularity and the potter's thumbprint left in the glaze. A mold makes none of these, and a model trained to please makes fewer every year, so leaving one is a small declaration that a person, and not a process, said the last word. As a filter, it does the esquivalience trick: pitched just below the threshold of a skim and just above the threshold of an expert, it lets the hurried reader glide past and the copier reproduce it intact, while the one reader working at your level catches it and knows. Catching it is the start of the only conversation worth having.
The move has a boundary, and the boundary is steep. A flaw reads as chosen only against a backdrop of evident care; drop a fault into sloppy work and it is simply more of the sloppiness, indistinguishable from the incompetence it was supposed to rise above. A deliberate imperfection is a claim about your own standards, and the claim is legible only when those standards are visible everywhere else on the page. There is an arms race under it, too. A flaw repeated becomes a mannerism, a mannerism is a pattern, and a pattern is the one thing a machine learns to imitate cleanly, which collapses the very signal it was sending. The signature has to keep moving to stay a signature.
I have argued before that murk in your own sentence is a tell, the honest leak that you do not yet know the thing as well as the words pretend. This is that mechanism turned over. There the imperfection was involuntary and aimed inward, a diagnosis of the writer; here it is deliberate and aimed outward, a message to the reader. Unchosen fog says the idea is unfinished; a chosen flaw says a person is finished and present. Both run on the same fact, that imperfection carries information a clean surface cannot.
When the perfect version of anything is free, the only proof that a person was here is the part the perfect version would have left out. The right error rate was never zero. We only believed it was while perfection was still expensive enough to mean something.