New AI Rules in Schools for 2027: What Counts as Cheating Now

The single most useful thing to know is that AI detectors are unreliable, and that they are unreliable in a specific and unfair direction. A study of seven widely used detectors found they wrongly flagged 61 per cent of essays written by non-native English speakers as AI generated.

On essays by native speaking US eighth graders, the same detectors were near perfectly accurate. That asymmetry is the thing a parent most needs to understand about how these accusations happen.

Seven widely used AI detectors, tested against student essays
What the study testedResult
Essays by non-native English writers, average false positive rate61.22 per cent wrongly labelled AI generated
Essays flagged by all seven detectors18 of 91 non-native essays, 19.78 per cent
Essays flagged by at least one detector89 of 91 non-native essays, 97.80 per cent
Essays by native English writers, US 8th gradersNear perfect accuracy
Non-native essays rephrased to sound more nativeFalse positive rate fell to 11.77 per cent
Native essays simplified to sound less nativeMisclassification rose from 5.19 to 56.65 per cent

Liang, Yuksekgonul, Mao, Wu and Zou (2023). The authors caution against using these tools in evaluative or educational settings.

Why the detectors fail this way

The study’s explanation is that detectors lean on perplexity, a measure of how predictable the writing is. Writing with a more limited vocabulary and simpler sentence construction is more predictable, and that describes competent non-native writing as much as it describes machine output.

The demonstration is elegant and damning. When the non-native essays were rephrased to sound more like a native speaker, the false positive rate fell from 61.22 per cent to 11.77 per cent. When native speakers’ essays were simplified to sound more like non-native writing, misclassification rose from 5.19 per cent to 56.65 per cent.

So the detectors are substantially measuring linguistic sophistication rather than authorship. The authors conclude by cautioning against their use in evaluative or educational settings, which is about as direct as research papers get.

What this means if your child is accused

Ask two questions, calmly and in writing. What evidence exists besides the detector score. And what false positive rate the school attaches to the tool it used.

Most schools cannot answer the second, which is itself the point. A tool with a published 61 per cent false positive rate for a large group of students is not evidence of anything on its own.

Then ask for the process to rest on things that actually indicate authorship: draft history, document version history, the student’s ability to discuss their own work, and a comparison with their previous writing. Those are harder to fake and harder to misread.

How to raise something with a teacher without it becoming a confrontation applies here more than anywhere, because an accusation conversation goes badly very easily.

Five positions a school might take, and what gets argued about
Common policy positionWhat it usually permitsWhere disputes arise
ProhibitedNothing, for any part of the workSpellcheck, grammar tools and translation sit in a grey area
Permitted with disclosureUse, if the student says howWhat counts as needing disclosure
Permitted for process, not productBrainstorming and planning, not draftingWhere planning ends and drafting begins
Permitted for specified tasksWhatever the assignment statesInconsistency between teachers
Encouraged with instructionUse as a taught skillAssessment design, since the old essay no longer works

Get the rule in writing, per subject

The most common way students get into trouble is not dishonesty. It is applying one teacher’s rule in another teacher’s class.

Policies genuinely differ, and the same school can have a department that treats AI as a taught skill and another that prohibits it outright. A student who has been encouraged to use it for planning in one subject will reasonably assume the same applies elsewhere.

So the useful request is not what the school’s AI policy is. It is what the rule is for this assignment, in this subject, in writing. Grammar and translation tools in particular sit in an unstated grey area in most policies and are worth asking about specifically.

Where we think schools are getting this wrong

Buying detection rather than redesigning assessment. A detector promises a cheap solution to an expensive problem and, on this evidence, does not deliver one while creating a new category of injustice.

The schools handling it best that we have seen have gone the other way: more work visible as it develops, more in class writing, more assessment that involves talking about the work. That is more effort and it does not generate false accusations.

We would also say the students worst served by detector based enforcement are precisely those who most need school to be fair: multilingual students, students with smaller academic vocabularies, students whose writing is plain because they are still building it. The tool penalises exactly them.

Source

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns (Cell Press); preprint arXiv:2304.02819. Seven detectors tested against 91 TOEFL essays by non-native writers and 88 US 8th grade essays by native writers.

How to find what has actually changed in your state covers tracking policy, the full set of score charts covers the assessments, readyscores.com/ is the place to start with a report, and our methodology sets out our sourcing.

Common questions

What counts as cheating with AI at school?

It depends entirely on your school’s policy, and policies differ substantially. The practical answer is to get the rule in writing per subject, because the most common cause of trouble is a student applying one teacher’s rule in another teacher’s class.

Are AI detectors reliable?

No. A 2023 study of seven widely used detectors found they wrongly flagged 61 per cent of essays by non-native English writers as AI generated, while being near perfectly accurate on essays by native speaking US 8th graders. The authors advise against using them in educational settings.

My child was accused based on a detector. What can I do?

Ask what evidence exists besides the detector score, and ask for the detector’s false positive rate. Request the process be based on drafts, version history and a conversation with the student rather than a tool the research says is unreliable.

Is using AI to check grammar cheating?

In most policies no, but it is frequently unstated, which is why it causes disputes. If the policy does not say, ask the teacher in writing and keep the answer.

How should a school handle this?

Clear rules stated per assignment, assessment designed so the work is visible as it develops, and no reliance on detector scores as evidence. The disputes almost always come from vague rules rather than from dishonest students.

Similar Posts