Why AI Writing Sounds Like AI (and What to Cut First)

The tells are real, and they are not the problem

Anyone who has read a page of AI fiction can name the smell. Certain phrases. A particular evenness. A metaphor that arrives from the wrong room. The mistake is assuming that removing the smell makes the writing good. It makes it quieter, which is not the same thing.

A draft can contain zero clichés and still be dead on arrival, because what readers are reacting to is not a word list. It is the absence of a decision. Nobody chose anything in that paragraph, so nothing is at stake in it, and a reader can feel the difference long before they can explain it.

So treat the tells as symptoms. They are worth learning to see, because they point at the place where the draft stopped being written and started being produced.

Three patterns that give the draft away

The "not X, but Y" reflex

This is the most reliable tell in English AI prose, and it is not a bad sentence on its own. The problem is frequency. A model reaching for depth will define a thing by negation, because contrast reads as insight:

It wasn't grief. It was the shape grief leaves in a room.

Fine once. By the fourth time on a page you are reading a tic. Put a hard limit on it: one per thousand words, and only where the contrast is genuinely the argument. If the negation is decoration, delete the first half and keep the image.

Then check the neighbour. The same reflex shows up as "not just a story, but a reckoning," and as any sentence where the second clause restates the first with more syllables. Those can go entirely.

Metaphors from the wrong shelf

Models pull figurative language from whatever corpus is thickest, which is often technical. So a character's silence becomes a background process still running. A memory is cached. A feeling is a signal with noise.

None of that is wrong exactly. It is out of register. A metaphor should come from the world the character stands in, because that is where their vocabulary lives. A farmhand compares things to weather and machinery. A night nurse compares things to shifts and vitals. A model compares things to software, because it has read a great deal of software.

The test is simple. Ask who in the scene would say this, and whether they would say it that way. If the answer is "a technical writer," you have found the seam.

Every paragraph the same size

Read the draft out loud and listen for the shape. AI drafts tend to run three to five sentences per paragraph, each closing on a small tidy turn, every turn in the same place. The rhythm is so consistent that it starts to feel like a metronome instead of a voice.

Human drafting is lumpy. A four-word paragraph after a nine-line one. A section that ends mid-thought because the thought was done. Fixing this costs nothing and changes how the page feels more than any synonym swap: break one paragraph down to a single sentence, merge two others, and let one section run long.

Readers are better at this than writers assume

There is a common assumption that only technical people notice AI prose. In practice the skill is exposure, not expertise. Someone who has scrolled past a thousand generated images gets good at spotting them, and the same thing happens with paragraphs. They cannot always name the feature. They just stop trusting the page.

This is why the judgment of a single reader is weak evidence and the judgment of a crowd is strong. One person saying "this feels AI" is a hunch, and hunches are wrong often enough to be useless on their own. Forty people in a comment section arriving at the same conclusion is a different kind of signal, because they are carrying different samples of different models and comparing notes without meaning to.

For a writer, the practical version is: do not ask one friend whether the draft sounds generated. Publish a chapter where readers have no reason to be polite, and watch whether they stay. Retention is a blunter instrument than a detector, and much harder to argue with.

A detector score, for what it is worth, is one model's guess about another model's output. It is not a verdict about your prose, and it should never be the reason you change a sentence.

Why "humanizing" passes make it worse

The usual fix is a vocabulary exchange. Swap the flagged words, break a few sentences into fragments, add an anecdote. The result reads clean and completely forgettable, because every pass converges toward the same average. You have removed the fingerprints and kept the absence.

There is a second cost. Every scrub teaches you to write for the detector instead of the reader, and the detector is not the thing that decides whether anyone finishes chapter two.

What actually moves a draft is decision density: how often the page contains something only this author would have written. A specific noun. An opinion you would defend. A detail from a place you have actually been. You do not need one per sentence. One per two hundred words is enough to keep the page from feeling machine-made.

A twenty-minute pass

Run this after the draft is structurally finished, not before. Editing for voice on a broken scene is polishing a window in a house with no floor.

1. Count the negations. Search for "not" and "isn't." Keep the two that carry the argument. Rewrite the rest as plain assertion.
2. Circle every metaphor. For each one, name the world it came from. Any metaphor from the writer's desk rather than the character's world gets replaced or cut.
3. Read one page aloud. Mark where you run out of breath and where you speed up to get through. Long sentence, short sentence, and the paragraph that refuses to be tidy.
4. Delete the first sentence of each section. You will keep about half. The other half were throat-clearing, and the section starts better without them.
5. Add one thing you know and the model does not. A brand name from your hometown. The wrong name people use for a street. The way a specific hospital smells in the morning. This is the part no amount of prompting will supply.

Then stop. A draft that has been through five rounds of de-AI-ification reads like a draft that has been through five rounds of de-AI-ification.

The part that is not about sentences

Everything above is craft, and craft is the easy half. The harder half is that a model will produce a competent scene whether or not the scene is worth writing, and competence is persuasive. If you cannot tell a good paragraph from a plausible one, a better model mostly gets you more plausible paragraphs per hour.

That is a separate problem with a separate fix, and it is covered in why a better model will not rescue a bad draft. For now, the useful takeaway is narrower: when a draft sounds like AI, the words are the symptom. The missing decisions are the disease.

If you want to see the difference immediately, take the same premise to two places. Generate a short story with a plain instruction, then generate it again with one concrete constraint — a named town, a specific job, a detail that could only be true of one person. Read both out loud. The second one is not better because the model tried harder. It is better because you gave it something to be wrong about.

FAQ

Is there a list of AI words I should avoid?

There are lists, and they go stale within a model generation or two. Words like delve and tapestry were reliable tells a while ago and are mostly just ordinary words now. Patterns last longer than vocabulary: the negation reflex, the uniform paragraph, the metaphor from the wrong shelf. Learn patterns.

Can readers actually tell, or is this paranoia?

Some readers can, and they are usually the ones who read a lot in your genre. They will not necessarily say "AI." They will say the chapter felt flat, or that the character stopped wanting anything. Treat that feedback as a pointer to a real problem, because it almost always is one, whatever produced it.

Does this mean I should not use AI at all?

No. It means the sentence-level output is the least valuable thing a model gives you. Outlines, alternatives, arguments against your own plan, and a first pass you can react to are all worth more than prose you then have to disguise. If you are writing in a fixed form, the templates are a faster way to see what the form demands than asking a chat window to guess.