AI detectors don’t work

What the testing found

The most cited evaluation is Weber-Wulff et al., published in the International Journal for Educational Integrity in 2023. It put fourteen detection tools, Turnitin among them, through a controlled test. Not one reached 80% accuracy. Only five got past 70%. Six of the fourteen produced false positives — human writing labelled as machine-written.

The authors’ own summary is that the tools are “neither accurate nor reliable”. That is a direct quote from a peer-reviewed paper about the entire product category, not a vendor comparison or a blog post.

Worth knowing about their bias direction, because it is the opposite of what most people assume: the tools leaned toward calling text human. They miss AI writing more often than they invent it. The false positives are the rarer error, which is exactly why they are so damaging when they land — nobody is expecting one.

Who gets wrongly flagged

In 2023 a Stanford team ran seven widely used GPT detectors over 91 TOEFL essays, all written by humans, all by students whose first language is not English. The detectors wrongly flagged 61.3% of them on average. Nearly one in five was flagged by every detector in the test. Over 97% were flagged by at least one.

The same detectors were close to perfect on essays written by native-speaking American eighth-graders. The difference is not that one group cheated. It is that detectors work by measuring how predictable the writing is, and someone writing carefully in a second language produces exactly the flat, regular prose the tools associate with a machine.

That mechanism catches other students too. Anyone who writes in short, plain sentences produces low-variability text. Researchers have warned that neurodivergent students are more likely to be hit by false positives, and Bloomberg documented the case of an autistic student flagged and given a zero. Neither of those is a large empirical study, so treat it as a warning rather than a measurement — but it is a warning from people who have looked closely.

What Turnitin says about Turnitin

This is the part worth reading twice, because it comes from the vendor rather than a critic. Turnitin publishes a false-positive rate of “under 1%” for documents containing more than 20% AI writing. At the level of individual sentences, the figure it gives is around 4% — roughly one highlighted sentence in twenty-five may be human-written.

And in its own FAQ: the percentage on the AI writing indicator “should not be used as the sole basis for action”. The company selling the detector is telling schools not to act on the number alone. That instruction is frequently the first casualty once the score is in front of somebody.

One more piece of context. OpenAI built a classifier to detect its own model’s output and withdrew it in July 2023, citing low accuracy. It caught about a quarter of AI-written text and falsely flagged nearly one in ten human samples. The company with the most training data and the strongest commercial reason to solve this could not.

Institutions that have switched detection off

Curtin is the one that matters most to a WA school, because it is the university a good share of your Year 12s will walk into. From the start of this year they are marking without it. Waterloo went further and said why: their own IT testing found the product flagging human-written text as “100% generated by AI”, more than once.

None of these institutions decided that AI use stopped being a problem. They decided the detector was not evidence.

What works instead

1. Treat a score as a reason to look, never as a finding A flag is a prompt to open a conversation. It is not a result, it cannot be shown to a parent as proof, and on the vendor’s own instruction it cannot be the sole basis for action.
2. Design the task so the process is visible Drafts, planning notes, an in-class writing stage, a short verbal follow-up on the argument. Work that shows its own development is far harder to outsource and far easier to defend if it is ever questioned.
3. Say what is allowed, in writing, per task Most students who use AI on work where it was not wanted were never told clearly. A one-line statement on the assessment sheet — what is permitted, what must be declared — removes most of the ambiguity before it becomes a case.
4. Ask students to declare use rather than hide it A disclosure line turns AI use into something discussable. Detection turns it into something to conceal, which makes the concealment better rather than the use rarer.
5. Never open a misconduct process on a percentage alone The Australian Catholic University processed thousands of flags in a year and later dismissed many of them, according to ABC reporting. The cost of that lands on students, on staff time, and on how much your families trust the school.

Why we publish this

A vague answer helps nobody, so this one is written down where a Head of Learning Area can check it before relying on a detector.

The figures above link to their sources below. The Australian Catholic University details come from ABC News reporting, which we haven’t linked here. We’ll update this page if any of it changes.