Why AI detectors do not work, and what to do instead

A detector offers exactly what an overwhelmed department wants: paste the essay in, get a percentage, act on it. The trouble is that the percentage does not mean what it appears to mean, and treating it as evidence produces the single worst outcome available to a school — a diligent student accused of cheating.

The failure is structural, not a product defect

Detectors work by measuring how predictable a text is. Machine-generated writing tends to choose likely words in likely orders, so unusually smooth, unsurprising prose scores as machine-written.

The problem is that plenty of human writing is also smooth and unsurprising, and it is a specific kind of human writing: careful, conventional, formal. Which is to say, the writing produced by a student who has been taught to write clearly, who is working in a second language and sticking to safe constructions, or who is following a structure the school gave them.

A detector systematically suspects the students who have most carefully done what they were told.

This bias is not something a better model removes, because it is the signal itself. Meanwhile a student who asks a model to write with more variation, or who rewrites a few sentences by hand, defeats the detector without effort. So it points the wrong way in both directions at once.

What a percentage actually is

"87% AI" reads like a probability that the student cheated. It is not. It is a score describing the text's statistical texture, on a scale the vendor chose, with a threshold the vendor picked. It carries no information about who wrote it.

Test this before trusting any tool: run several pieces of work you watched a student produce. Schools that do this usually stop using the detector, because the outputs do not survive contact with cases where the answer is known.

The cost of one wrong accusation

Consider what it takes to defend yourself. The student has no evidence of their own innocence beyond insisting, and the school is holding a number. It falls on a child to disprove a machine.

That damages the student's trust in the school permanently, and the story travels — to their friends, to their parents, to the wider community. One such case costs more than any number of undetected AI essays.

What actually works

Assess the process, not only the artefact

Ask for the outline, the draft with its revisions, and the sources. A student who did the work has these and a student who did not cannot fabricate them convincingly. This also happens to be better pedagogy than grading a finished product alone.

Make some work unfakeable by design

Handwritten in-class writing, a short viva on a submitted essay, or a task tied to something specific to your classroom — a discussion you had, a text you annotated together. Two minutes of "talk me through your second paragraph" separates the student who wrote it from the student who did not, with no technology at all.

Require disclosure and make it safe

One line: which tool, which part, what they changed. If honest disclosure is punished, students conceal, and you are back to guessing. Treat the line as compliance.

Say what each task is for

Most AI-in-homework anxiety comes from tasks whose purpose was never stated. If a task exists to practise an unaided skill, say so and assess it in conditions that make that real. If it does not, AI use may not matter much.

If you already have a detector

You do not need to throw it away, but change its status. It may raise a question; it may never be the answer. Nothing should follow from a score alone except a conversation, and the conversation — not the number — is what you act on.