Can AI detectors be wrong?

Yes. Every detector — including ours — makes mistakes in both directions. The useful question is not whether errors happen, but how, how often the conditions for them arise, and what a responsible person does with a result anyway.

Two ways to be wrong

A false positive is a human-written text flagged as AI-generated. A false negative is AI-generated text that passes as human. Both are built into the method. A detector does not observe how a text was produced; it measures patterns in the finished words — how predictable they are, how evenly that predictability is spread — and compares them against what it has learned human and machine writing tend to look like. Pattern overlap between the two is real and permanent, so the boundary between them can never be a wall. Any tool that claims certainty is telling you more about its marketing than its method.

Why human writing gets flagged

False positives are not random. They cluster where human writing is forced into the same shape as machine writing. Genre is the biggest factor: lab reports, legal boilerplate, technical summaries, and application essays follow templates, and a template suppresses exactly the variation detectors look for. Heavy editing has a similar effect — a draft revised many times toward “correct” prose converges on the most probable phrasing, which is the model's home territory. Short texts make everything worse, because a few sentences simply don't carry enough signal to distinguish habit from generation, which is why responsible checkers refuse to judge very short passages at all.

The non-native speaker problem

One group carries a well-documented, unfair share of the risk: people writing in a second language. A smaller working vocabulary and a reliance on learned, standard constructions produce text that is — through no fault of the writer — more predictable, in the statistical sense detectors measure. Research on widely used detectors has found exactly this bias against non-native English writers, while native speakers' work passes far more easily. The practical consequence should be stated plainly: a flag on a non-native writer's text is weaker evidence than the same flag on a native writer's, and anyone in a position of judgement — a teacher, an editor, an employer — has a duty to know that before they act on a score.

Why AI writing gets through

False negatives are just as systematic. Detection reads the surface statistics of a text, so anything that disturbs those statistics disturbs the reading: paraphrasing a generated draft, translating it and back, mixing generated passages with genuinely human ones, or simply prompting a model toward a rougher, more idiosyncratic style. A mixed text — part human, part machine — is the hardest case of all, which is why honest reporting distinguishes “mostly human with possible assistance” from a clean bill of health. A detector that says “human” is really saying “no strong machine pattern found here,” and those are not the same sentence.

What a result should — and should not — do

A detection result is an indicator. It can justify a closer look, a request for drafts, a conversation. It cannot, on its own, justify a punishment, because the person it falls on cannot prove a negative — there is no test that certifies “a human wrote this,” and demanding one reverses the burden of proof. Institutions that use detection well treat it as the start of an evidence-gathering process that includes drafts, version history, the author's ability to discuss their work, and subject knowledge of the genre's conventions. If you have been flagged for work you wrote yourself, ask for exactly that process, and bring your drafts: they are the strongest answer there is. And if you are the one holding the report — remember that the cost of a false accusation lands on a person, while the cost of a missed case lands on a policy. Weigh them accordingly.