AI detectors do not work, and what to do instead

Every few weeks a tool appears that claims to tell you whether a pupil's writing was produced by a language model. Schools buy them because the alternative — reading thirty pieces of work and forming a judgement — is expensive and uncomfortable.

They do not work, and the way they fail is worse than not working.

The failure is not symmetric

A detector that is wrong 5% of the time sounds tolerable until you ask which 5%.

Detectors work by measuring how predictable text is. Fluent, plain, unadorned writing scores as machine-like, because that is what a model produces. So the pupils most likely to be flagged are:

• pupils writing in a second language, whose sentences are shorter and more regular
• pupils who have been explicitly taught to write plainly, which is most of them
• pupils with a limited vocabulary, who reuse the words they are sure of

The pupils least likely to be flagged are confident writers with idiosyncratic style. An accusation lands hardest on the child least equipped to argue their way out of it, and there is no appeal process because the evidence is a percentage from a tool nobody can inspect.

There is also no ceiling on the countermeasure. A pupil who wants to defeat a detector asks the model to write less predictably. It takes one sentence.

What to do instead

None of this is an argument for ignoring the question. It is an argument for a process that does not rest on a number.

Ask for the working, not the work

Set tasks that produce a trail: a plan, a draft with changes visible, a paragraph written in the room. The trail is the assessment and the final piece is the artefact. This is more work to set up once and less work to mark forever, because a missing trail is visible at a glance.

Make attribution routine and unremarkable

Ask every pupil to write one line at the end about what they used and how. Every pupil, every piece, including "nothing".

The reason this works is that it removes the incentive to hide. A declaration that is required of everybody is not a confession. Our course spends a whole plate on this, and the outcome we look for is a child doing it without being asked.

Ask them about it

Two minutes, one question about a choice they made in the writing. "Why did you start with the storm?" A pupil who wrote it can answer instantly. A pupil who did not, usually cannot, and the conversation is a conversation rather than an accusation with a number attached.

This is the oldest technique in teaching and it remains the most reliable one.

The thing worth saying out loud to a class

Most pupils are not trying to cheat. They are trying to finish something they found hard, using the tool that was on the screen.

A class that has been told, plainly, what the tool is bad at and where the line is will mostly stay the right side of it. A class that has only been told there is a detector will spend its energy on the detector.