Assessment Design

Stop Using AI Detectors: What Works Instead

An AI detector gives you a number you can't explain, a conversation you dread, and a student who may or may not have cheated. Here's the position to take instead, and the evidence that protects you when the meeting gets difficult.

New here? Put one of your own assignments through the free Assignment Audit. In about five minutes it finds the weakest point and names the first change to make.

Why the Detector Fails You

The conversation in schools since 2023 has largely centred on detection: how do we catch students using AI, and what happens to them once we do? The answer, by now, is that we can't catch them reliably, we're often not permitted to act on a detector score alone, and catching and punishing shouldn't be the goal. AI writing is good enough that no tool detects it dependably, and false accusations built on detection software have already done real harm to students across North America.

There's also a professional risk that's easy to underestimate. The moment you put a percentage in front of a parent, the meeting becomes a debate about software you didn't create and can't defend. You'll lose that debate, because the parent is right about the detection software.

You don't have to take my word for it. In August 2026, MIT's committee on AI in teaching and learning recommended against relying on AI detectors. It warned that they can mistake the writing of non-native English speakers and neurodivergent students for AI, and noted that MIT's own disciplinary committee doesn't treat detector output alone as sufficient evidence. There's more on what that report means for English in a separate piece, and the full report is on MIT's website.

The AI watermarks that arrived in 2026 don't change this. A watermark can suggest that one company's AI was involved, but it can't show who did the thinking, and most teachers can't check for one. What watermarks can and can't tell you is covered in its own piece.

The Better Question

The useful question isn't "did this student use AI?" It's "can this student demonstrate their own thinking in front of me?" That reframe changes the design of the assessment rather than the intensity of the policing, and it's the part of the problem a classroom teacher can most directly control.

AI can produce the essay. It can't produce the 15 minutes of timed writing in your classroom, the dated drafts, the peer feedback the student responded to, or the sentence they can explain and defend in front of you.

What to Collect Instead

Three kinds of evidence do the work a detector can't.

1
ProcessCollected in stages and dated, so that a finished piece has visible drafts.
2
In-class CognitionAny thinking done in front of you: a timed write, an exit conversation, and/or an oral defence.
3
SpecificityThe student's own material, the reference to Tuesday's discussion, or the moment their choices changed and why.

None of this requires software and none of it requires an accusation. It requires the assignment to ask for things an AI chatbot can't supply.

Not sure whether your assignment collects any of it? The free Assignment Audit is eight questions and about five minutes. It names the level your task is actually asking for and the one change worth making first. Take the Audit.

The Conversation, When It Comes

If a meeting does happen, the move that works is to take the detector off the table in your first sentence. You agree that detection software isn't definitive. You say that it's not the basis of the decision. Then you put the draft stages on the table, if the student produced them, and talk about the gap between the process observed in class and the work that was handed in. The parent arrived prepared to fight about the detector. You've declined the fight and moved to ground where the evidence is yours. I learned that the hard way, in a meeting I describe in What Critical AI Literacy Looks Like in Classrooms.

Where to Start

Run an ordinary assignment through the Audit first, so you know what your baseline actually is. Then rebuild the pillar it scores worst on, starting with the task where a detector would have tempted you. One assignment. One change. Starting tomorrow.

Confirm by clicking the link in the email that follows, and the files arrive straight after. An occasional email after that, usually every week or two. Unsubscribe in one click. A personal address works better than a board one, because school filters often block the confirmation link.

Questions Teachers Ask

Should I use an AI detector at all?

As evidence in a meeting with a student or a parent, they're a liability, because you'll be asked to explain a number you can't really explain. Don't use them.

What do I say when a parent raises false positives?

Agree with them in your first sentence. You're not there to defend the software. Say the detector score isn't the basis of the decision and move the conversation to the gap between the process observed in class and the draft submitted.

What if I have no process evidence?

Then the conversation isn't winnable, and the honest move is to treat this one as a design problem rather than an integrity problem. Change the assignment, collect process from the next one, and you won't be in this position again.

Does this mean students can use AI freely?

No. It means the control is structural rather than forensic. You decide what level of AI use a task permits, you say so in advance, and you build the task so that a student who ignores you can't produce the evidence the task requires. Don't spend hours after the fact trying to prove something an effectively designed AI-Resistant assignment shows you plainly.

Try It on Your Own Assignment

Eight questions, five minutes, and the one change worth making first.

Audit an Assignment Read the Framework

Sources checked, October 2026.

Developed by George O'Toole: over 35 years in Secondary English, Head of Department, Educational Technology Coach, Faculty of Education instructor. Meet George