AI detector for teachers: what it can and cannot prove
An honest guide to AI writing detection in the classroom. What the scores mean, where they go wrong, and how to use them without getting it wrong yourself.
The position to start from
No AI detector available to teachers today, including this one, can prove that a student used AI. The honest framing is that detectors produce a probability score based on the patterns the model can see. Useful as a starting point. Not useful as a verdict.
Universities that have treated detector scores as evidence in academic misconduct cases have already had to reverse course. The published research on detector accuracy is consistent on this. The score is a signal. The conversation is the evidence.
What an AI detector actually does
An AI writing detector looks for statistical and structural patterns that tend to show up more often in model-generated text. Things like sentence length consistency, predictable vocabulary distribution, low burstiness, and phrasing that fits common training patterns.
When several of these signals stack up, the detector raises its risk score. When they do not, it lowers the score. That is the whole mechanism. There is no hidden lookup of “things ChatGPT wrote”, because that data does not exist in a usable form.
The implication for teachers is the part that gets lost. A high score means the writing resembles model output along measurable dimensions. It does not mean the writing was produced by a model. A student writing in a clean, formal, structured style can look statistically similar to ChatGPT, especially in essays where structure is taught and rewarded.
Where AI detection goes wrong
Non-native English speakers get flagged more often
Several published studies have found that detectors flag essays by non-native English speakers at materially higher rates than native speakers. The reason is straightforward. Less idiomatic phrasing and more uniform sentence structure both look like model output to a pattern detector. A teacher relying on the score alone risks penalising students for how they were taught to write.
Highly structured student writing gets flagged
Year 11 and sixth-form students who have been drilled on PEEL paragraphs, topic sentences, and connective phrases produce writing that statistically resembles model output. That is the same coaching that gets them marks. Penalising them based on a detector score punishes the structure you taught them.
Lightly edited AI text fools most detectors
A student who runs model output through a paraphraser, swaps a few words, or rewrites the opening line can drop the detection score significantly. This is where detectors are weakest, and it is also the most common way AI is actually used. The clean-input case is the easy case. The real cases are messier.
Short text is unreliable
Detectors need a reasonable sample to work with. A 100-word answer does not give the statistical model enough to make a confident call. Many detectors will still produce a score for short text, but that score should be treated with much wider error bars.
How to use detection well
The teachers getting the most out of detection use it as a triage tool, not a verdict tool. A high score routes a piece of work into a conversation. The conversation is where the real evidence comes from.
- Compare the flagged piece against the student's previous in-class writing. A sudden jump in style, vocabulary, or argument depth is more informative than any score.
- Ask the student to talk through their argument. Genuine authors can explain their choices. AI-assisted work tends to fall apart under follow-up questions.
- Look at the flagged passages specifically, not just the overall percentage. Knowing which sections triggered detection helps you ask better questions.
- Set expectations up front. Tell students how you screen, what scores mean, and what happens when a piece is flagged. Predictability protects everyone.
The aim is not to catch students out. The aim is to keep assessment honest while staying fair to the students who did the work themselves.
Where Is It AI? fits
Is It AI? is built around this framing. It runs multiple detectors in parallel, shows which passages triggered detection, and explains what the model saw. Student work is never stored. Plans start at £9.99 per month for individual teachers, with department options for schools.
It is not a cheat-catcher. It is a screening tool that gives you a fair starting point for a conversation.
Comparing tools? See 13 AI detectors side by side.