AI detection for teachers
Why a detector score cannot settle a case, what evidence can, and how to have the conversation with a student.
EdexiaUpdated 15 October 2026 · 3 min readYou read a piece and something is off. The voice isn't theirs, the argument is too tidy, the vocabulary jumped a year overnight. The reflex is to paste it into a detector and let the number decide. The number can't decide it, and when it is wrong, a student who did nothing wrong ends up accused.
What a detector actually measures
A text detector only sees the finished writing. Most estimate how predictable the words are, and how evenly that predictability runs through the piece. Language models write smooth, predictable text, so smooth, predictable text scores as "likely AI".
That is a guess about style, not a record of how the work was made. It goes wrong in both directions.
- False alarm
Clear, formulaic writing looks like AI
Students taught to write to a tight structure, and students writing in a second language, often write predictably. Researchers at Stanford found popular detectors flagged essays by non-native English writers as AI far more often than essays by native speakers (Liang et al., 2023).
- Missed
Edited AI text looks human
A student who rewrites, paraphrases or adds a few deliberate errors will often pass. Detectors are weakest against the students who try hardest to get past them.
The companies that build these tools say the same. OpenAI withdrew its own AI text classifier in 2023 because of its low rate of accuracy, and Turnitin tells institutions its AI indicator should not be the sole basis for action against a student.
Why a flag is hard to read
Most of your students wrote their own work. That is exactly what makes a flag hard to read. Set the numbers below however you like.
- Used AI, flagged: 8
- Wrote it themselves, flagged anyway: 5
- Used AI, not flagged: 2
- Wrote it themselves, not flagged: 85
The detector flags 13 students. 5 of them wrote the work themselves, about 38% of the flags. The score looks the same for all of them.
An illustration. The rates are hypothetical, not the figures for any real detector.
Push the catch rate all the way up and the wrong flags are still there. They come from the honest majority, and nothing in the score tells you which flagged student is which. A score can only ever be a reason to look closer.
What evidence holds up
A detector asks whether the text looks like AI wrote it. The question that matters is how the text was written, and that has a much clearer answer if you can see the process. Tick what you would have in a real case.
Not a case yet. Everything here describes the finished text or the screen, not how the student wrote it.
No single signal decides a case. Several kinds of evidence pointing the same way, plus a conversation with the student, is what a fair process rests on.
How to have the conversation
The aim is to find out what happened and teach from it, not to win.
- Look at the process first. Draft history, a replay, plans and notes tell you more than the final text.
- Ask them to talk you through it. Why this argument, why this quote, what they changed. Students who wrote a piece can usually explain it.
- Ask what tools they used. Plenty of students use AI in ways they think are allowed. If the line wasn't clear, this is the moment to teach it.
- Use your school's academic integrity process for anything formal, especially work that counts toward a senior certificate. Each board has its own rules on authentication.
- Write down what you saw and decided, in case it is reviewed later.
Make the honest path the easy one
Catching AI use after the fact is the expensive end of the problem. The cheap end is the task itself.
- Make drafting part of the task. Collect plans and drafts, give feedback on them, and assess the growth as well as the product.
- Keep some writing in class or handwritten. It gives you a baseline of each student's own voice.
- Give students help that doesn't write for them. If the only help at 10pm is a chatbot that writes the essay, plenty will use it. A tutor that asks questions and points them back to the text keeps the thinking theirs.
How Edexia helps
Edexia is where the class writes, so the process is visible by default. You can replay how each piece was written, see where text was pasted and when a student left the tab, and see an AI-likelihood signal next to it. Each one is an input to your judgement, never a verdict.
Students get a tutor in the same space. It asks questions, explains and points them back to the text, and it can't write the response for them. You see the whole conversation, so AI use stops happening in a hidden tab and becomes something you can teach into.
Common questions
Can AI detectors prove a student used AI?
No. A detector score is a statistical guess about the finished text, and it is wrong in both directions. Treat it as a reason to look closer, never as proof. How the work was written, and a conversation with the student, tell you far more.
What is the best AI detection tool for teachers?
The best evidence comes from seeing how the work was produced, not from scoring the final text. Tools that record drafting, pasting and revision (a writing replay) give you something you can show a student and ask about. Use a text detector only as one input among several.
Are AI detectors biased against EAL/D students?
There is published evidence that they can be. A 2023 Stanford study found popular detectors flagged essays by non-native English writers as AI far more often than essays by native speakers. Take extra care before acting on a score for an EAL/D student.
What should I do if I suspect a student used AI?
Look at the process first, then ask the student to talk you through the piece and what tools they used. Use your school's academic integrity process for anything formal, and write down what you saw and decided.
Does Edexia detect AI?
Edexia shows you how a piece was written: a replay of the writing session, paste and tab-change flags, and an AI-likelihood signal. It gives you evidence to judge, not an automatic verdict.