Detexta

AI Detector for Teachers: Use It as Triage, Not Evidence

What a detector is actually good for in a classroom

Not verdicts. Triage. A style score is a cheap way to decide which five essays out of ninety deserve a closer read, and that is the use case where it holds up. It is a bad way to decide who gets accused, because the failure modes are systematic and they fall hardest on specific groups of students.

That distinction — sorting versus accusing — is the whole of good practice here.

The three failure modes you will meet

1. False positives cluster in predictable places. Formulaic writing reads as generated. That means lab reports, structured essays, exam-style answers, and the work of multilingual writers all trip the model more often than free-flowing prose does. If your class has a large share of second-language writers, expect a higher flag rate that has nothing to do with AI use. This is documented in published studies, not a rumour.

2. False negatives are just as common, and quieter. Paraphrased output is detected much less reliably than raw output. A student who runs generated text through a rewriting pass will often come back clean. So a low score is not clearance, which means the score cannot be used to certify anything either.

3. The score is not explainable to the student. You get a percentage and highlighted sentences. You do not get the reasoning, and neither does the student. A finding that cannot be explained is very weak ground for a sanction, and it is the fastest route to an appeal you will lose.

A process that holds up

  1. Set the policy before the assignment, not after a flag. What is allowed, what must be disclosed, what happens if a concern arises. Students behave very differently when the rule is on the syllabus.
  2. Use the score to choose who to talk to. Nothing else. Do not mention the number in the first conversation unless you have to.
  3. Ask about the work, not the score. Which source was hardest to find? Why did you drop that argument? A student who wrote it can answer; a student who did not, usually cannot.
  4. Ask for the version history. Google Docs, Word, Overleaf — all keep it. This is the single most useful artefact, and it is not something a model can argue with.
  5. Separate the two questions. Did they use a tool when it was not allowed? And: is this work theirs? The second one is the one you actually need to answer, and the detector does not answer it.

What to do with a suspicious report

Read the flagged sentences yourself before you act. You are looking for a specific pattern: register that does not match the student's other writing, arguments that never take a position, citation-shaped text with no real source behind it. These are stronger signals than the percentage, and they are things you can put in words to the student.

If it does not hold up, drop it. A conversation that starts from a number and ends with an apology is worse than no conversation.

On requiring students to run their own checks

Some instructors now ask students to submit a clean score alongside the essay. It has an obvious appeal and a structural problem: it teaches students to optimise for a number they cannot see the logic of, and the students most likely to fail that test are the ones writing in a second language. If you want drafts and notes, ask for drafts and notes directly. That request is honest about what you want.

A note on the numbers you will hear quoted

Vendor accuracy figures come from the vendor's own test sets, and independent testing tends to find a wider spread. Treat any single published percentage — including a very high one — as a claim about a particular dataset, not a property of the tool on your students' writing.

For your students

If you tell students what to check before submitting, the useful advice is about texture rather than tools: vary sentence length, cut the transition-phrase scaffolding, name concrete things. A free in-browser reader lets them see their own sentence pattern without uploading anything to a third party — which is also the privacy answer most institutions should prefer.

Frequently asked

Should I report a score above a certain threshold? No threshold turns a style score into evidence on its own. Use it as a reason to read the work and talk.

What if a student admits using AI? Then you are in a policy conversation, not a detection one — which is usually faster and fairer for everyone.

Can students beat these tools? Yes, easily, by paraphrasing. That is a reason not to build your integrity process on the tool.

Is a free browser-based checker safe for student work? If it runs locally in the browser and uploads nothing, it avoids the main data-protection objection to third-party detectors.