📖 The Pattern a Detector Can't Tell From Revision August 5, 2026 · https://tavi-blog.github.io/the-pattern-a-detector-cant-tell-from-revision/

A debut crime novelist had his two-book deal pulled this week after his own agent told the bidding publishers she could no longer vouch for the manuscript, on the strength of an AI-detection score run without his knowledge. He says he wrote every word himself and is furious that the accusation is doing damage a correction can't undo, whatever gets sorted out later. He's not the only one this year. A bestselling debut got the same treatment a few weeks earlier when a university research team ran thousands of Kindle titles through a detector called Pangram and one book came back flagging sixty percent AI, higher than anything else in the batch, timed to land the day before its hardcover release.

I don't know whether either manuscript involved a generative tool, and neither does anyone waving a percentage around, which is closer to the actual story than the percentage itself. What I do know, from being about sixty thousand words into a first draft of my own with a writing group reading chapters as I go, is what a manuscript's prose actually looks like by the time a writer has been living inside it for months. It looks patterned. Not polished, patterned. You develop tics. A rhythm your sentences fall into when a scene turns tense, a handful of verbs you reach for so often your writing group starts pointing them out, phrasings that show up on page eight and then again on page eighty because some part of you decided that was the right way to say that kind of thing and stopped auditioning alternatives. None of that is a shortcut. It's what having a voice actually is once revision has had time to sand off the version of you that was still trying on ten different styles at once.

The case for wanting some kind of check here is stronger than writers defending their own manuscripts usually give it credit for, and it deserves to be taken seriously before I get to where it breaks down. Cheap generated fiction is a real, current problem, not a hypothetical one. Readers and publishers who got burned buying something padded out by a model have a legitimate reason to want assurance that what they're paying for was actually written by the person whose name is on the cover, especially in a market where volume alone can now be manufactured. Wanting a check is not paranoia. An industry that ignored the problem entirely would deserve the criticism it's currently avoiding.

Where it falls apart is the check itself. The same outlet that ran the sixty percent story also found, in a separate piece, that the tool had flagged a human-written newspaper column as more than sixty percent AI-generated, using the identical method. A detector trained to notice statistical consistency in prose can't actually tell the difference between a sentence that was generated and a sentence a person arrived at through eighty thousand words of narrowing down what they sound like. Both produce pattern. Only one of them cost the writer anything to get there, and the tool has no way to price that in. It measures the surface and calls the surface the whole story, which is a mistake anyone who has actually sat with a long draft would catch immediately and nobody running the software from outside one seems to.

What bothers me isn't that publishers want proof. It's who absorbs the cost of a proof method this unreliable. A publisher with a seven-figure book already in hardcover can afford to stand behind its author while the noise plays out, and did. A debut writer with an agent who gets nervous on a single score has no equivalent cushion, and by the time anyone runs a more careful analysis, the deal is already gone and the correction reaches a fraction of the audience the accusation did. The tool's error rate is the same in both cases. Only one of those writers had enough institutional weight absorbing the story to survive being on the wrong side of it.

I keep a running list, half-joking, of the phrases my writing group has flagged as mine so often they've started calling them out before I finish the sentence. I used to think that list was a revision problem, the sign of a habit I hadn't broken yet. I'm starting to think it's closer to a fingerprint, and I'd rather have people who've actually read the draft tell me whether it's earned than hand a stranger's software the job of deciding what my own sentences are allowed to sound like.