The AI Detector Says My Poem Is AI: Creative Writing's Flag Problem

You wrote the poem at two in the morning, or on a train, or in the margin of a receipt while the feeling was still hot. You cut a line you loved because it was showing off. You moved a stanza three times. You said the word salt and then decided it should be ash, and then went back to salt. And then, because someone told you to, or because you were curious, or because a teacher required it, you pasted the finished thing into an AI detector. And the little meter swung, and a number came back, and the number said your poem was 92 percent likely to be machine-generated.

It lands like an accusation. Not because you think a piece of software has feelings or judgment, but because the poem was the truest thing you made all month, and now a tool is telling you it reads like the output of a language model that has never been sad, never lost anyone, never stood in a kitchen at midnight. There is a particular vertigo in being told your art is fake by a machine, and it is worth saying plainly, before anything else: the machine is wrong, and I can show you why it is wrong, and why creative writing specifically is the genre these tools fail hardest.

Detectors were built for a different kind of writing

The first thing to understand is that AI detectors were never designed with poetry in mind. They were built to look at expository prose: essays, articles, reports, the kind of five-paragraph, thesis-and-support writing that fills classrooms and content mills. Everything about how a detector reasons assumes that shape. It expects sentences that march along in ordinary syntax. It expects paragraphs. It expects a certain steady, workmanlike rhythm where each sentence is a reasonable length and each one hands off to the next without too much strangeness. Feed a detector a school essay about the causes of the French Revolution, and it is at least operating in the world it was trained to judge.

A poem is not that world. A poem is a deliberate assault on ordinary syntax. It compresses. It breaks lines where prose would never break. It repeats on purpose. It leans on meter and sound and white space to carry meaning that prose would spell out. When a detector meets a poem, it is like a wine critic being handed a cup of espresso and asked to score its tannins. The instrument is calibrated for something else, so whatever number it produces is close to noise. The problem is that the number does not look like noise. It looks precise. It looks confident. And that false confidence is the whole trap.

To see why creative writing trips these tools so reliably, you have to understand the two or three things detectors actually measure. They do not read for meaning. They cannot tell whether a line is beautiful or true. What they do is measure statistical texture, and the words for that texture are worth knowing: perplexity and burstiness. We have a fuller walkthrough of these in our explainer on perplexity, burstiness, and watermarks, but the short version matters here because it explains the whole misfire.

Why crafted language reads as machine language

Perplexity is, roughly, a measure of surprise. A detector runs your text through a language model and asks, at each word, how predictable was that word given the ones before it? If the text is full of the choices a model itself would have made, perplexity is low, and low perplexity is exactly what detectors treat as a signal of AI authorship. The reasoning goes: machines produce the statistically likely next word, so smooth, predictable, low-surprise text is suspicious.

Now think about what a poem does. A good poem is not random; it is the opposite of random. It is language that has been sanded and fitted until every word feels inevitable. The poet spent an hour choosing between two syllables so that the line would land with a click, so that the sound would carry the sense, so that nothing felt loose or accidental. That is craft. And craft, measured by a machine that equates surprise with humanity, reads as low entropy — as smooth, fitted, inevitable — which is to say it reads as artificial. The detector cannot tell the difference between the smoothness of a sentence that a model spat out on autopilot and the smoothness of a line that a human being polished for a week. Both are low-surprise. To the tool, they look identical.

This is the cruelest irony of the whole business. The better your poem is — the more you revised, the more you cut, the more you made each word carry its full weight — the more likely it is to be flagged. Rough, messy, first-draft prose full of odd tangents and clumsy transitions is, statistically, "burstier," more surprising, more human-looking to the tool. Polished, deliberate, compressed art is not. You are being penalized for exactly the thing that makes the writing good. A poem scoring "AI" is not evidence that your poem came from a machine. It is evidence that you succeeded at making language feel intentional, and that the tool has no way to distinguish intention from imitation.

Burstiness makes it worse. Burstiness measures how much sentence length and rhythm vary across a passage. Human prose tends to swing: a long, winding sentence, then a short one. Machine prose tends to be more even. Detectors look for that unevenness as a sign of humanity. But poetry, especially formal poetry, is built on regularity. A sonnet holds a steady meter on purpose. A villanelle repeats whole lines by design. A poem in tight quatrains keeps its lines close to the same length because the form demands it. To a burstiness detector, that intentional regularity looks like the flat, even output of a machine. You did the hard thing — you held a form — and the tool read your discipline as evidence of automation. If you want the mechanical detail of what these tools scan for, our breakdown of what AI detectors actually look for lays it out signal by signal.

Poetry, fiction, and song lyrics are not flagged the same way

It helps to separate the genres, because they fail the detector for slightly different reasons, and knowing which is yours can steady you.

Poetry is the most vulnerable of all, for every reason above and one more: it is short. Most detectors are unreliable on short passages and openly admit it, because they need a certain volume of text to gather any statistical signal at all. A fourteen-line poem simply does not give the tool enough to work with, so it fills the gap with a guess and dresses the guess in a percentage. Compression, unusual line breaks, deliberate repetition, tight meter, and brevity all stack against poetry at once. It is almost the perfect storm of things a detector cannot parse.

Fiction is flagged less often but still gets caught, especially literary fiction and anything with a strong, controlled prose style. A writer who has developed a clean, spare voice — short declarative sentences, careful rhythm, not a wasted word — is producing exactly the low-perplexity, low-burstiness texture that trips the tool. Minimalism reads as machine-like. So does polished dialogue, because dialogue tends to be short and even. Genre fiction with a lot of convention in it — the familiar beats of romance or thriller pacing — can also read as predictable, because the detector confuses "following a genre's rhythm" with "producing the likely next word." Again: the more assured the craft, the higher the risk.

Song lyrics may be the single worst case. Lyrics repeat by design — the chorus comes back, the hook comes back, whole lines return. Repetition is the point. But repetition is one of the loudest "AI" signals a detector knows, because looping and echoing text is a classic failure mode of older language models. So a lyric that repeats its chorus four times, exactly as every song in human history has done, gets read as suspiciously mechanical. The form's oldest, most human feature — the return, the refrain, the thing that makes a song a song — is the thing the tool distrusts most.

The famous-text problem

There is a specific, almost comic failure worth knowing about, because it is the fastest way to prove to yourself that these tools are not measuring what they claim. Take a famous poem — Dickinson, Frost, Whitman, a Shakespeare sonnet, the opening of a canonical novel — and paste it into a detector. There is a good chance it will come back flagged as AI-generated.

The reason is straightforward once you see it. Detectors judge text partly by how well a language model predicts it. Famous literature has been read, quoted, and reproduced so many times across the internet that it sits deep inside the training data of every large model. The model has effectively memorized it. So when the detector asks "how predictable is this text to the model," the answer for a beloved, endlessly-quoted poem is: extremely predictable — the model has seen these exact lines thousands of times. That extreme predictability registers as low perplexity, which registers as "AI." Frost did not use ChatGPT. The tool is not detecting authorship. It is detecting familiarity, and then mislabeling familiarity as automation.

Sit with that for a moment, because it dismantles the whole premise. If a detector calls Emily Dickinson a machine, the detector is not a machine-detector at all. It is a familiarity-and-smoothness detector wearing a costume. And your poem, if it happens to be smooth and clean and deliberate, gets caught in the same net as Dickinson — not because you copied her, but because you and she both did the human thing of making language feel inevitable. That is company, not guilt.

Being told your art is fake

I want to stop and name the emotional part, because the technical explanation, however airtight, does not by itself dissolve the sting. There is something uniquely deflating about a false positive on creative work that a false positive on a lab report does not carry. A poem is not an assignment you completed. It is, at least a little, a piece of you that you decided to show. When you submit an essay, you are performing a task. When you submit a poem, you are confessing something. So when the verdict comes back "not human," it does not read as a technical error. It reads as a judgment on the confession — as if the machine peered at the most honest thing you made and said, this is not real, this did not come from a person, there is no one in here.

That feeling is real, and it deserves to be taken seriously rather than argued away. But it is also, and this matters, entirely misplaced — because the tool did not read your poem. It could not. It cannot feel the grief in the third stanza or hear the way the final line turns. It ran a statistical texture-check and returned a number, and the number knows nothing about you. The vertigo you feel is the gap between what you put into the poem and what the tool is capable of perceiving, which is almost nothing. You are grieving a judgment that was never actually made, because no one, and nothing, was there to make it. A detector cannot deny your authorship any more than a smoke alarm can review your cooking.

What this means for students and workshops

If you are a creative writing student, or you teach one, the stakes sharpen, because now a false positive is not just a bruised feeling — it is potentially an accusation of cheating attached to your name and your grade. This is where the argument has to move from "the tool is imprecise" to "the tool is invalid for this use, full stop."

Here is the core of it. AI detectors carry a false-positive rate — a rate at which they flag genuinely human writing as machine-made — and that rate is not zero even on the ordinary prose they were built for. On creative writing, which violates every assumption the tool relies on, the effective false-positive rate is far higher and essentially unknowable. That means a detector result on a poem or a story is not evidence of anything. It cannot support an academic-integrity finding, because it does not distinguish a false flag from a true one, and on creative work it produces false flags constantly. We walk through how these errors happen and why they are so common in our piece on why AI detectors produce false positives, and the situation is directly parallel to what happens when Turnitin flags an essay a student wrote themselves — same broken instrument, higher stakes on creative work.

For a workshop or a program, the sane policy is to stop running creative writing through detectors at all. There is no version of it that yields useful information. A low score proves nothing, because the tool is unreliable. A high score proves nothing, because the tool flags Dickinson. Using it can only manufacture false accusations against your most careful students — the ones who revise hardest, whose polished work reads as "low entropy" — while teaching everyone in the room that craft is dangerous and roughness is safe. That is precisely the opposite of what a writing program exists to do. The presence of the tool actively corrodes the thing you are trying to build.

How to respond when it happens to you

Suppose it has already happened — the flag exists, and maybe someone with authority over your grade is taking it seriously. Here is how to stand your ground without spiraling.

• Keep your drafts and your process. The single strongest response to a false accusation is evidence of how the work was made. Version history in a document, dated notebook pages, the messy earlier drafts, the line you cut and the line you added — this is a record no detector can produce and no accuser can wave away. For creative work especially, the trail of revision is the proof of authorship, and it is far more persuasive than any counter-score. Save your drafts as a habit, not just when trouble comes.
• Reframe the score as evidence about the tool, not about you. Do not argue the number down. Argue that the number is meaningless for this genre. Point out that the same tool flags canonical, pre-AI poems as machine-made. Point out that the tool measures smoothness and predictability, and that polish is exactly what a revised poem is supposed to have. Make the conversation about the instrument's validity, not your innocence, because on the merits the instrument loses.
• Ask what the detector actually claims to measure. Most vendors, in their own documentation, disclaim reliability on short texts and warn against using scores as sole evidence of misconduct. Those disclaimers exist for a reason. Bringing the tool's own caveats into the room is often enough to end the matter.
• Do not let a machine relitigate your relationship to your own work. This is the quiet one, and it is the most important. You know how the poem was made. You were there. A statistical texture-check run by software that cannot read does not get a vote on whether you are the author of your own art.

The score is a mirror pointed at the tool

Step back far enough and the whole episode inverts. You went to the detector to learn something about your poem, and what you actually learned was something about the detector. A poem "scoring AI" is not a fact about the poem's origin. It is a demonstration of the tool's limits — a small, precise map of everything these instruments cannot see. They cannot see meaning. They cannot see craft, except to mistake it for automation. They cannot see the difference between memorized Dickinson and a model's autopilot. They cannot see you, standing in the kitchen at midnight, choosing between salt and ash.

What they can see is texture, and texture is a terrible proxy for authorship in a form whose entire purpose is to shape texture deliberately. When a tool built to expect paragraphs meets a thing built to break them, the tool fails, and the failure tells you the tool was pointed at the wrong target — not that the target was fraudulent. The high score is a mirror. It reflects the detector's blind spots back at itself, and you happened to be standing in the way of the reflection.

Keep writing the difficult, deliberate, human thing

So here is what I would leave you with, if you are the poet at two in the morning who got the bad number. Nothing about your poem changed when the meter swung. The grief in the third stanza is still there. The turn in the final line still turns. The word you chose is still the right word, chosen by you, for reasons a piece of software will never be equipped to weigh. The flag measured how smooth and fitted your language had become, and smooth and fitted is another way of saying finished — it is a backhanded compliment from a machine that does not know it paid you one.

If anything, treat the flag as a strange kind of receipt. It says: this reads as though every word were placed on purpose. That is what you were trying to do. You did it well enough that a tool trained on the ordinary and the average could not tell your care apart from a machine's indifference — which says everything about the poverty of the tool and nothing about the poverty of the poem. Detectors are junior instruments built for junior prose, and they will keep improving, and they will still never be the right tool for pointing at a poem and pronouncing it real or fake. That was never a question a percentage could answer. Go finish the next one. The kitchen is still there, and so are you, and the poem is waiting to be made by the only kind of maker who has ever made one.