The update nobody could read
I was writing a weekly status email to a client. What shipped, what is pending, where we stand, the usual. Every fact was in there, correct and accounted for. And when I read it back, I knew the client would never make it to the bottom. It was a wall. Accurate, complete, and dead on the page, the kind of thing you forward and quietly hope nobody asks you about, because answering a question would mean reading it too.
Nothing in it was wrong. That was the strange part. I could not circle a single sentence and call it a mistake, and the whole thing still failed at the one job it had, which was to keep a busy person reading long enough to understand where their project stood. It was mechanical. It had lost the thing that keeps anyone reading anything.
That sent me looking, and I found a Wikipedia page kept by the editors who clean up AI slop. It catalogues the tells they use to spot machine text. I turned it into a checklist, expecting a long list of unrelated faults. What I found was one fault, wearing a lot of different coats. I built it into a skill I now run on everything I write, and the status email was the first thing it fixed.
One disease, a hundred symptoms
The checklist looks varied. Puffed-up significance, a bag of overused words, little clauses bolted onto a sentence to fake analysis, the stock contrasts, the bolded bullet headers. Take them apart and nearly every one is the same move underneath. The model reaches for the most average phrasing available, and the average is always vaguer and grander than the thing you actually meant.
The page has one example I keep coming back to. The inventor of the first train coupler becomes a revolutionary titan of industry. Look at what happened there. A specific person who made a specific thing got deleted, and in his place went a phrase you could glue onto anyone who ever ran a company. The subject got less specific and more exaggerated in the same breath. That is the whole trick, and it is what every tell on the list is quietly doing. My status email was the same move at the scale of a paragraph: real progress, flattened into on-track language so smooth it carried no actual news.
The model does not write badly in a hundred ways. It writes badly in one way, in a hundred places.
It predicts the likely next word, and the likely word is the common one, and common phrasing is generic by construction. Left alone it drifts toward the centre of everything it has read, which turns out to be a kind of beige. The puffery and the stock vocabulary and the fake-analysis tails are one current pulling one direction: away from the specific thing you meant, toward the average of everything anyone has ever meant.
The specific thing was the point
This is where it stops being a style complaint. The specific fact is the writing. It is the only part worth reading. A revolutionary titan of industry tells you nothing. He invented the first train coupler tells you everything the first phrase was pretending to have. When the model smooths the second into the first, it is not polishing the sentence. It is deleting the information and keeping the shape.
And that is the part you cannot hand off, because the model does not know the specific true thing. It was not in the room. It does not have the number. It does not know which blocker actually moved this week or which one has been stuck for ten days. It can produce a confident average of a status update, and the average is wrong in the one way that matters, which is that it tells the client nothing they could act on. So the work left for a person is exactly the work the smoothing destroys: putting the specific back. The real number, the honest still blocked, the one line that tells them where they actually stand.
Every tell I chase down ends in the same place. A vague grand sentence sitting where a plain specific one should have been, because the plain specific one was harder and I had not written it yet.
Human is not a coat of paint
So I stopped treating sounds like AI as a surface I could sand off. You cannot thesaurus your way out of it. Swap robust for sturdy, keep the inflated empty sentence around it, and all you have done is hide the problem one layer down. The tells are symptoms. The disease is having nothing specific to say and dressing the gap in important-sounding cloth.
Which means the thing we call a human voice is mostly downstream of something plainer: having a real thing to say and saying it straight. The warmth, the texture, whatever makes a sentence feel like a person wrote it, that is the residue left behind when you refuse to let the specific thing get smoothed away. The AI gets me to a clean draft fast. Then the real work starts, dragging it back off the average one sentence at a time, until it says the true, specific, slightly awkward thing I would have had to think of either way. That is what the skill is. Not a filter that adds a human voice, a set of rules that stops the specifics from being smoothed out in the first place.
So the next time something you wrote reads perfectly and lands on nothing: did you ask it for something to say, or only for something that looks like saying?