Are AI detectors accurate enough to punish a student?

No. They return a probability based on style, and the score changes with editing, language background and text length. A single score is not evidence that a specific student used AI.

Why do detectors flag non-native English writers?

Simpler, more regular sentence structures are common in second-language writing and also common in machine text, so classifiers trained on both tend to score that style higher.

Can a student beat a detector?

Often, by producing a rougher draft or rewriting output. That is the point: the tools measure style, so style changes move the score without changing who wrote the work.

Note / Education

What school AI detectors actually measure

Home/Notes/AI detectors

By the VOXA team · notes on detection research and school policy

4 November 20264 min readAnalysis

The short version

School AI detectors do not detect AI. They score text against patterns and report a probability, and that score moves when a human edits the text, when a student writes in a second language, or when the essay is short. The research literature has found both false positives and false negatives at rates that should make any disciplinary decision built on a single score unsafe.

What the tool actually does

Every detector on the market is a statistical classifier. It was trained on text labelled as human or machine, and it learned features of that training set: sentence length variance, word frequency, punctuation habits, the predictability of each next word. At run time it compares the essay to those patterns and returns a number.

That number is not evidence of authorship. It is evidence that the text resembles the machine-written examples more than the human-written ones in that particular training set. Change the training set and the same essay scores differently.

Where it breaks

Second-language writers. Students writing in English as a second language use simpler, more regular sentence structures. The same regularity appears in machine text, which is why this group has been flagged at higher rates in published testing.
Short answers. A 150-word paragraph gives the classifier very little to work with. Short submissions produce the most confident wrong answers.
Editing. Run an AI draft through a human rewrite and the score usually drops, without the content changing. The detector is reacting to style, not to provenance.
Formulaic writing. Lab reports, legal memos and five-paragraph essays are taught as patterns. Learning the pattern makes your writing look more machine-like.
False negatives. A student who prompts a model for a rough, uneven, error-containing draft can score as human. The tools are as weak in that direction, and that direction gets no headlines.

What schools are actually buying

A number and a dashboard. The vendor reports a confidence score, the school treats it as a finding, and the burden lands on the student to prove a negative, which is close to impossible. Several universities have stopped publishing detector results for exactly that reason, and others now require a human review before any accusation.

The honest framing for a policy is: a detector can start a conversation, not end one. It can prompt a teacher to ask about the process, look at drafts, or set an in-class version of the task. It cannot establish that a specific student cheated.

What to do instead

Assess the process, not just the product. Drafts with timestamps, in-class writing, oral defence of a submitted argument, and version history in a document all carry more signal than a probability score. None of them are fooled by a rewrite.

If a school does use a detector anyway, three rules keep it fair: never act on a single score, always show the student the evidence, and always offer a path to demonstrate the work, such as a short oral explanation. Those three rules would prevent most of the cases that end up in the news.

The wider point is that detection is a symptom of assessment design. When the only artefact is a finished essay produced at home, every tool in the chain is guessing. When the process is visible, the guessing stops mattering.