AI Detectors Don't Work, and Pretending They Do Hurts Writers

A student gets accused of cheating because a website spat out "98% AI-generated" on an essay she wrote by hand over three nights. A freelancer loses a client who ran their draft through a checker and didn't like the number. An author gets a one-star review claiming the book is "obviously AI" because somebody pasted a sample into a tool that confidently lied.

These aren't edge cases. They're the predictable output of a technology that does not work and is being treated as if it does. AI detectors are unreliable in a specific, dangerous way, and the harm of pretending otherwise lands hardest on people who did nothing wrong.

Why detection can't work the way people think

Start with the basic problem. An AI detector tries to guess whether text was produced by a language model. But language models are trained to produce text that looks like human writing. That's the entire design goal. So you're asking a tool to reliably distinguish a thing from its own deliberate imitation of that thing.

Most detectors lean on signals like "perplexity" and "burstiness," roughly, how predictable and how varied the sentences are. The theory is that AI writes smoothly and predictably while humans are messier. There are two fatal cracks in that theory:

Clear human writing looks predictable. When a person writes plainly and correctly, simple words, clean grammar, low surprise, they trip the exact signals detectors read as "machine." Good, clear prose is a false positive waiting to happen.

AI can be made to look messy. Anyone trying to evade detection just asks the model to vary sentence length and throw in some quirks. The signal collapses. So the tool fails in both directions at once: it flags careful humans and waves through anyone actually gaming it.

A detector that catches the honest and misses the cheaters is worse than no detector. It's a random number generator wearing a lab coat.

The false positives aren't random. They hit specific people.

This is the part that should end the conversation, and it's the part the vendors are quietest about. The false positives cluster, and they cluster on people who write differently from the average native, neurotypical, formally-educated writer.

Writers using English as a second language get flagged far more often. Their prose tends toward simpler structures and more limited vocabulary, exactly the "low perplexity" pattern detectors read as artificial. A Stanford research group found that detectors flagged a large share of essays by non-native English speakers as AI-generated while clearing native speakers. The tool isn't detecting AI. It's detecting "doesn't write like a native," and punishing it.

Neurodivergent writers get caught too. Plenty of autistic and otherwise neurodivergent people write in patterns, consistent structure, precise repetition, even cadence, that happen to match what detectors call machine-like. They wrote every word and the tool calls them liars.

Anyone who writes plainly is at risk. Edit hard for clarity, strip the clutter, and you've made your prose more "predictable" by the detector's logic. The better the clean writing, the higher the false-positive odds. That's the whole mechanism, inverted into an accusation.

So the burden of this broken technology doesn't fall evenly. It falls on second-language writers, neurodivergent writers, and disciplined editors, the people with the least power to push back when an institution waves a percentage at them and calls it proof.

"But the number was 99%"

The numbers are theater. A detector reporting "99% AI" is not measuring a 99% chance the text is AI. It's outputting a confidence score from a model that is itself often wrong, with no calibration you can trust and no way for you to audit it. The vendors know this, which is why the fine print quietly disclaims using the output for high-stakes decisions, right under the marketing that invites exactly that.

And the false positive rate, even when it sounds small, is catastrophic at scale. A tool that's wrong 1% of the time, run across a school of thousands of honest essays, falsely accuses dozens of innocent students every term. Each one is a real person facing a real consequence over a coin flip dressed as evidence. "Mostly accurate" is not good enough when the cost of a miss is someone's record, job, or reputation.

What to do instead

If detection is broken, the answer isn't to find a better detector. There isn't one, and chasing it just launders the same harm. The answer is to stop treating "did a machine touch this" as the question worth policing, and to build trust the old, durable way.

For educators and editors: judge the work and the process, not a black-box score. Talk to the writer about their argument. Ask them to walk through a revision. Look at drafts over time. A person who wrote the thing can discuss it; a score can't tell you anything you can act on fairly.

For writers under suspicion: keep your receipts. Draft history, version snapshots, notes, the messy trail of your actual process. The strongest defense against a false accusation is showing the work as it grew. A manuscript with a real revision history is a far better witness than your protest that the detector is wrong.

For everyone: drop the binary. "AI or not" was always the wrong frame. The honest questions are about how a tool was used and whether the writer is being straight about it, which is exactly why we make the case for authors disclosing AI use proportionally instead of relying on anyone to detect it. Transparency you can trust. Detection you can't.

This is also where the coach model of AI sidesteps the whole mess. When the tool's job is to help you brainstorm, find plot holes, and pressure-test your own draft, your fingerprints are all over the work because you did the deciding. You're not laundering a generated draft and hoping a detector misses it. The honest use never needed a detector to clear it, which is the point of what Polyz does.

The detectors don't work. Believing they do isn't neutral. It manufactures accusations against the writers least able to defend themselves, and it lets the actual bad actors stroll through untouched. Stop outsourcing your judgment to a number from a tool that can't deliver what it promises. Read the work. Talk to the writer. Keep the receipts.

Write in a way that leaves an honest trail and you'll never need a machine to vouch for you. Start a free trial.