The seven tells of AI fiction prose.

People expect AI fiction to read like a robot. It doesn't. By 2026 the sentences are clean, the grammar is immaculate, the vocabulary is tasteful. What gives it away is not error. It's the absence of the irregularities that human writing accumulates as a byproduct of a single mind making thousands of small, idiosyncratic decisions.

We have now read a few thousand manuscripts — human, machine, and the increasingly common hybrid where a writer drafts and a model "polishes." Across that pile, seven structural tells recur. None is proof on its own. Together they form a fingerprint.

1. Rhythmic flatness

Human prose breathes unevenly. A writer hits a five-word sentence, then a forty-word one because the thought genuinely ran long, then a fragment. Models regress toward a comfortable mid-length cadence and stay there. Measure the standard deviation of sentence length across a chapter: human literary prose is bursty; generated prose is suspiciously regular. This is the single most reliable signal we track, and the hardest for a writer to fake by editing — because rhythm is a property of how a passage was thought, not how it was spell-checked.

2. The triad reflex

"She was tired, hollow, and strangely calm." Three is the model's favorite number. Lists of exactly three adjectives, three clauses, three escalating images, deployed far more often than any human stylist would tolerate. The tricolon is a real rhetorical figure, but a human reaches for it occasionally, for emphasis. A model reaches for it as a default container for any quality it wants to describe.

3. Summary in place of scene

Asked for emotion, models name the emotion and then gloss it: "A wave of grief washed over her, and she realized nothing would ever be the same." It is competent and it is empty. The human version stays in the body and the room — the cold tea, the unanswered phone, the sentence the character can't finish — and never says "grief" at all. Generated fiction tells you the weather of a feeling. Human fiction makes you stand in it.

4. The "not just X, but Y" scaffold

"It wasn't just a house; it was a memory." "She didn't walk; she drifted." This antithetical upgrade construction is everywhere in generated prose because it is a cheap way to manufacture the feeling of insight without committing to a specific image. Once you see it you cannot unsee it. A manuscript that uses it more than once or twice a chapter has almost certainly been through a model.

5. Dialogue without subtext

People in life rarely say what they mean, and good fiction dramatizes the gap. Generated dialogue tends to be informationally efficient: characters state their feelings and motivations plainly, take turns politely, and rarely talk past each other. It is also conspicuously clean — almost no false starts, interruptions, or the verbal litter ("well," "look," "I mean") that real speech carries. Sanitized dialogue is one of the loudest tells, with one important caveat we'll return to.

6. Vocabulary that is wide but not deep

Models have an enormous lexicon and a shallow attachment to any part of it. The result is prose with high word variety but no signature — no handful of words a writer overuses because of who they are, no consistent register, no idiolect. Human authors are recognizable across a book. Generated text is recognizable only as generated.

7. Voice that resets

Drafted-then-stitched manuscripts often show a bimodal voice: some chapters carry a tight, idiosyncratic human signature; others relax into the model's smooth median. Measured across a whole book, the variance between sections is itself diagnostic. A single author having a bad week still sounds like that author. A book that oscillates between two stable registers usually has two authors, and one of them is software.

The caveat that matters

Every tell above also appears, sometimes strongly, in perfectly human writing. Minimalists strip filler from dialogue on purpose. Pre-1980 literary prose almost never uses verbal litter. Genre conventions reward exactly the efficient dialogue that flags as machine-clean. This is why a single number should never be treated as a verdict, and why we publish the per-audit breakdown instead of a yes/no. The point of measuring the fingerprint is to start a conversation with the manuscript, not to end one.

The useful mental model is not "detector vs. text." It is closer to a forensic linguist: no single feature convicts, but the convergence of independent features, weighed honestly and with the false-positive cases kept firmly in view, tells you where to look harder.