Your child has been accused of using AI

Before anything else: do not ask your child whether they cheated

It is the natural first sentence and the one that costs you the most. A direct accusation produces a denial almost regardless of what happened, because the stakes of admitting it have just been set at the maximum. You then have a denial you cannot evaluate — and you have taught them that the first answer is the one to defend.

Ask about the process instead. "Walk me through how you did it — what did you start with?" A child who wrote it can tell you where they got stuck, which bit they changed, what they nearly wrote instead. A child who did not usually cannot, and the conversation gets there without either of you saying the word. You are collecting the same thing the school should be asking for.

Keep the clock in mind too: most schools give a window to respond, and it is usually shorter than the feeling. Get the deadline in writing early, then spend the evening on evidence rather than on the argument.

If nobody has accused anyone and you are the one who is unsure

What the detector score is actually worth

Less than the email implies, and this is measured rather than argued. Weber-Wulff and colleagues tested fourteen tools — twelve public ones plus Turnitin and PlagiarismCheck — and concluded in print that the available detectors are "neither accurate nor reliable". Elkhatat and colleagues, testing separately, found the same passage can come back with different verdicts from different tools.

The bias also runs in a direction most people get backwards. Weber-Wulff’s tools leaned towards calling text human-written, so their headline failure is missing AI text rather than inventing accusations. A high score is rarer than you might fear — which is exactly why a school treats one as damning when it appears.

The false-accusation risk comes from a second finding. Liang and colleagues reported that detectors "consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified." The mechanism is the part to understand: these tools score plain, short, predictable language as machine-written. That describes a child’s writing, a child writing in their second language, and any child recently taught to write in a clean five-sentence structure. Schoolwork looks like the thing the detector is trained to flag.

One caveat before you take this to a teacher: none of these studies tested essays by children. They tested adult academic writing. What transfers is the mechanism, not a rate — so the sentence to use is "these tools are documented as unreliable", never "there is a known false-positive rate for children", which nobody has measured.

The first hour, in order

1. 1. Get the accusation in writing, with the deadline One short, non-defensive email: thank them, ask what specifically prompted the concern, what the process is, and by when you need to respond. You concede nothing by being easy to deal with, and the written record matters more than the tone of the first phone call.
2. 2. Ask what was noticed, not what the tool scored The useful question is "what in the work looked different from their usual writing?" A teacher who can point at a register, a vocabulary, a structure they do not recognise is telling you something a tool cannot. A teacher whose only answer is a percentage has a weaker case than they think, and most will say so when asked plainly.
3. 3. Find the drafting history before it expires The most valuable thing you can do tonight, and it has a shelf life. Google Docs keeps version history under File, then Version history; Word keeps its own in Microsoft 365. A document that grew over three evenings, with sentences written and rewritten, is near-impossible to fake and far better evidence than any argument about detectors.
4. 4. Write the reply around the process, not around the technology Offer the version history, and offer for your child to talk through the piece. Mention the reliability finding once, with the source — not as the centre of your case. A reply that is entirely an attack on the detector reads as a parent with nothing else; one that leads with the drafting record reads as a parent who checked.
5. 5. Decide who speaks, and let your child be in the room For a primary or middle-school child, you speak and they are present. Rehearse the one question they will certainly be asked — "tell me about this part" — on a paragraph from the middle. Not a script; just not the first time they have explained their own work out loud.

Four things it might be, and what each one calls for

The accusation is one event with several possible causes, and they do not all end in the same place. Work out which one you are in before deciding how hard to push — two of these four are genuinely your child’s mistake, and handling those as an injustice is how a small thing becomes a pattern.

The part that outlasts this assignment

Whichever of the four it was, your child has now learned something about how writing gets judged, and you get a say in what. The lesson you do not want is "write worse and you will not get flagged" — the rational response to a detector, and one children work out fast.

The better lesson is that the record of how you worked is worth keeping. Drafting in one document rather than pasting from five, leaving the messy first version in the history, being able to say why a paragraph sits where it does — habits that happen to be a defence, and also just how competent work gets done.

The honest other half: this is happening because AI can produce a passable school essay in nine seconds. The answer for a ten-year-old is not abstinence, which has never worked for any technology, but a specific rule — ask it to help you think, never to produce the thing you hand in — plus somewhere to use AI where making something is the whole point.

Where the line sits between help and doing it for them, subject by subject · How to tell whether they learned it or collected it

The research under this page

Two independent evaluations of detection tools plus the bias study, with what each does and does not support. Note that the first two were run on adult academic writing — no published study has measured detector accuracy on the writing of a nine-year-old, which is itself worth knowing before anyone quotes a number at you.

1. Weber-Wulff, Anohina-Naumeca, Bjelobaba, Foltýnek, Guerrero-Dib, Popoola, Šigut & Waddington, 2023 — Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, article 26. (Fourteen tools — twelve public plus Turnitin and PlagiarismCheck — concluded to be "neither accurate nor reliable", with a main bias towards classifying output as human-written, and performance significantly worsened by content obfuscation.) Used for the central claim that the tools are not reliable enough to decide on, in the authors’ own words, across fourteen tools including Turnitin. Note its error direction: the tools missed AI text more readily than they flagged human text. What it does not show: Tested on adult academic text, and its headline error direction is the opposite of the one a worried parent fears: the tools MISS AI text more readily than they flag human text. It supports "do not act on the score"; it does not by itself establish the false-accusation rate, which is what liang2023 is for.
2. Liang, Yuksekgonul, Mao, Wu & Zou, 2023 — GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. ("Our findings reveal that these detectors consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified.") Used for the false-positive mechanism — detectors systematically misread plain, limited language as machine-written. Cited for the mechanism only; its samples were adult writers, so no rate from it applies to a child. What it does not show: The samples tested were essays by adult non-native English writers and by US students, not by seven-to-twelve-year-olds. The mechanism it identifies — detectors score plain, limited, formulaic language as machine-written — is what transfers to a child’s writing; the specific rates do not, and we do not quote them.
3. Elkhatat, Elsaid & Almeer, 2023 — Evaluating the efficacy of AI content detection tools in differentiating between human and AI-generated text. International Journal for Educational Integrity, 19, article 17. Used only for consistency: different tools return different verdicts on the same passage. A small evaluation, so it corroborates rather than independently establishes. What it does not show: A small evaluation on a limited set of passages, so it is corroboration rather than a second independent measurement. Cited only for the consistency point — that the same passage can get different verdicts from different tools.
4. Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, 2024 — Generative AI without guardrails can harm learning: evidence from high school mathematics. (Nearly 1,000 students; an unguarded GPT-4 assistant raised practice scores but LOWERED later unaided exam scores by about 17%, an effect essentially eliminated by a hint-giving tutor version.) Used in the last section for why the answer is a rule about how AI is used rather than a ban. Unguarded AI help lowered later unaided scores; a hint-giving version removed the harm. High-school maths, not primary-school writing. What it does not show: High-school maths in Turkey, not elementary school, and the "guardrails" arm showed no POSITIVE exam effect either — it removed the harm. Anyone citing this as proof that AI tutoring raises attainment is citing the opposite of what it found.

Questions parents ask

Can AI detectors be wrong?

Yes, and the people who test them say so in print. Weber-Wulff and colleagues evaluated fourteen tools, including Turnitin, and concluded they are "neither accurate nor reliable". Liang and colleagues found detectors consistently misclassify plain, non-native English writing as AI-generated. Neither tested children’s writing, so use them to argue the score cannot settle it — not to claim an error rate for a fourth grader.

What should I say to the teacher?

Ask what they noticed in the work itself rather than what the tool reported, and offer the drafting history. Keep the reliability point to one sentence with its source. Most teachers already suspect the tools are shaky, and a parent leading with evidence of process is far easier to find a sensible route with.

How can I prove my child wrote it themselves?

Version history is the strongest thing you have — Google Docs keeps it under File, then Version history, and Microsoft 365 keeps its own. A document that accumulated over several sittings with real revision is very hard to fake. Second-best is your child explaining the piece aloud, including what they nearly wrote instead.

What if my child did use AI?

Disclose it before it is discovered, and separate the two conversations. The school one is about the rule and is usually survivable, especially when volunteered. The one at home is about what made handing in someone else’s work look like the best option that night — time, difficulty, fear of the mark — and that is the conversation that changes anything.

Somewhere the process is the point

The defence against all of this is a child who can explain their own work. Luchi has kids make things and then say what they made and why, and writes you a short note afterwards in their own words — so the record of how they got there exists before anyone asks for it.

Start your first 7 days for $5

Is my kid cheating with AI?

Part of the series: How to end the daily homework battle

Related guides in this series

• Teaching kids to start homework without reminders — Independence is a prompt you remove on purpose, not a trait you wait for. A five-rung ladder from "you say it daily" to "you only ask afterwards", with how long each rung takes.
• Seven practical homework focus strategies — Seven things that help once they have started, in the order worth trying them — plus an honest table of which rest on solid evidence and which do not.
• What parents should know about AI homework helpers — Unguarded AI help measurably lowered later unaided test scores; guardrails removed the harm. What is known, what is not, and five questions to ask any product.
• AI homework help for kids — Where the line sits between help and doing it for them, subject by subject — plus a five-step routine for one hard question and five questions to ask any homework product.
• Is my kid cheating with AI? — Why AI detectors are not reliable enough to accuse a child on, the ninety-second check that is, and what to do about each of the four things you are likely to find.
• The after-school hour, without the nagging — The fight is about the forty minutes before homework. A five-block plan, the four traps that turn a routine back into nagging, and the research that makes it work.
• Is your child actually learning from AI? — Fluent output is not learning. The three signs of real learning, the three signs the tool did the thinking, and a five-point checklist a parent can run on any AI app tonight.
• When reading is hard: what AI helps with, and what it hides — One sentence from a child can mean three different problems, and they want opposite responses. A ten-minute check that tells you which, what read-aloud support actually measures, and the dull thing with the best evidence.
• My child just asks the AI for the answer — Not laziness — a help-seeking failure documented in tutoring software twenty years before chatbots. The one change with a measured effect, and the commit-first rule you can start tonight.
• Let the AI mark the draft, not write it — The feedback research says it has to be nonevaluative, supportive, timely and specific — and a chatbot’s default reply fails three of the four. The sentences that fix that, and why the child does the typing.
• A science fair project with AI, without the AI doing it — Teacher-led science beat student-led by about .40 in effect size across 37 studies — so hands-off and hand-it-over fail the same way. Which two of six parts a chatbot helps with, and the four questions that prove the project is theirs.
• Studying for a test: what goes on tonight’s list — Eleven topics, three evenings. Asking what they feel shaky about is the one input that has been measured and found uncorrelated with what they actually know — so sort it in eight minutes instead, and do not cross anything off after one right answer.