Book News

How to Tell If a Book Was Written by AI

Somewhere in the last couple of years, “wait, is this AI?” became a question readers ask about novels the same way they used to ask it about term papers and news articles. Sometimes the suspicion is warranted. Often it isn't. The honest answer is that telling the difference is harder than either the confident detector scores or the confident skeptics online want it to be.

This guide is our attempt to lay out what's actually known: what AI-detection tools do and where they fall apart, what to listen for in your own reading that's worth trusting as an instinct rather than proof, and why “AI-written” is doing a lot more work as an accusation than it can actually support on its own.

What AI detectors actually do (and why they get it wrong)

Tools built to catch AI-generated text, including one called Pangram that's become prominent in publishing circles, work by scoring how closely a piece of writing matches patterns the tool has learned to associate with machine-generated prose. The output is a probability, a confidence score. It is not a transcript of how the text was actually produced, and it is not a substitute for knowing what happened in a specific writer's process.

That gap between “scored as AI-like” and “was written by AI” is not theoretical. It played out publicly this year when a Stony Brook University study, using Pangram, flagged the bestselling novel Daggermouth by H.M. Wolfe at 60% AI-written. Wolfe denied it outright, and her publisher, Simon & Schuster, went further, stating that it does “not believe conclusions about an author's work should be drawn from AI-detection tools that have been shown to produce false positives.” The underlying study, notably, still hasn't been peer-reviewed. We wrote the full story of that case separately: A Bestseller Was Accused of Being AI-Written. Here's What's Actually Known.

We're not going to pretend to explain exactly how Pangram or any other detector works under the hood. No source we can verify lays out its internal mechanism in detail, so we won't invent one. What we can say, because publishers, researchers, and the tools' own limitations keep demonstrating it, is that false positives are a real, acknowledged risk, not a fringe complaint. A score is a flag worth taking seriously. It is not a verdict.

The reading-level tells (treat these as instincts, not proof)

Detectors aside, readers have developed their own sense of what “reads AI” to them. Some of that instinct is genuinely useful. All of it should be held loosely, because readers get this wrong constantly in both directions — flagging clean, careful human prose as suspicious, and missing genuinely AI-assisted passages that happen to read as distinctive.

A few of the tells worth paying attention to, with the caveat attached to every single one of them:

Specificity versus generic detail. Human writing tends to reach for oddly specific, sometimes unnecessary concrete detail — the exact brand of a childhood cereal, the specific ache in a specific joint. Generated or heavily smoothed prose tends to default to averaged-out, plausible-sounding description that could belong to almost any scene. This is a real pattern. It is also true that some human writers, especially working fast in a commercial genre, write in exactly that averaged-out register on purpose, because it's what the genre rewards.

Emotional flatness or sameness across chapters. When a character's emotional register doesn't vary much scene to scene, or every conflict resolves with a similar shape and pacing, it can read as a text that was generated in similar-sized chunks rather than lived through as a whole. But plenty of human authors, particularly under deadline pressure or writing long series, produce exactly this kind of sameness without any AI involved at all.

Structural tidiness. A manuscript where every chapter resolves cleanly, every setup pays off on schedule, and nothing feels like the productive mess of a person actually wrestling with a draft can feel machine-smoothed. It can also just mean the author is a very disciplined outliner, or that a strong editor took a heavy hand to the manuscript, which is a normal and often invisible part of traditional publishing.

Repeated or unusual phrasing. This is closer to what the Stony Brook researchers actually measured in the Daggermouth study: multi-word phrases, five words or more, that appear in an unusual order rarely found in human writing but more often in AI-suspected text. It's a more rigorous version of the “this phrase again?” feeling careful readers sometimes get. As a casual reader, you almost certainly can't run this kind of frequency analysis yourself. Treat it as the researchers do: one signal among several, not a standalone verdict.

None of these tells, alone or together, are forensic proof of anything. They're reading instincts. A good instinct is worth noticing. It is not the same thing as evidence you could put someone's name next to in public.

Ghostwritten, AI-assisted, AI-written, and unfairly accused are four different things

Part of what makes this conversation so muddled is that several genuinely different situations get flattened into one accusation. It's worth keeping them separate:

Ghostwriting is a human writer, not the credited author, producing some or all of the text, typically under contract and with the credited author's involvement in shaping the story. It predates generative AI by decades and isn't, on its own, an AI story at all.

AI-assisted work means an author used AI tools somewhere in their process — brainstorming, outlining, line-editing suggestions — while the prose and the underlying creative decisions remain substantially the author's own. Where exactly the line sits here is genuinely being worked out across the industry in real time, and reasonable people disagree about how much assistance is disclosable versus unremarkable.

AI-written means the text was substantially generated by a model with minimal human authorship, credited to a human anyway. This is the accusation with real teeth, and it's also the hardest one to prove from the outside with any single tool or score.

Unfairly accused is an author whose natural prose style, editing process, or simple bad luck trips a detector's pattern-matching, producing a score that looks damning without being true. This is the category Daggermouth currently sits in, unresolved: accused, denied, defended by her publisher, and not settled by the one number that's been circulating.

What publishers are doing about it

As AI-detection tools get more common and more publicly cited, publishers are increasingly being asked to respond to accusations they didn't initiate. In the Daggermouth case, Simon & Schuster's response followed a pattern worth noting: it didn't simply defend the author in the abstract. It pointed to its own editorial and production process as evidence the book went through normal channels, and it directly challenged the reliability of the detection tool itself rather than treating the tool's output as settled fact. That combination — standing behind the process and questioning the measurement — is likely to become a more familiar publisher response as more of these stories surface.

Key takeaway from How to Tell If a Book Was Written by AI: “Somewhere in the last couple of years, “wait, is this AI?” became a question readers ask about novels the same way they used to ask it about term papers and news articles.”

Watch the companion video

We made a video covering this same territory, including the Daggermouth case as the live example. You can find it on our YouTube channel.

The honest bottom line

A detector score is a data point, not a confession. A prose tell is an instinct, not a citation. And a non-peer-reviewed study is a starting point for scrutiny, not the end of it. If you're trying to figure out whether a book you're reading was written by AI, the most honest answer is usually that you can develop a reasonable suspicion and you cannot, from the outside, reach certainty. That's worth remembering the next time a number like “60% AI-written” shows up attached to a real person's name.