The 3-minute homework check that beats any AI detector

A parent and child at a kitchen table doing the three-minute homework check: a sheet of paper shows 2x + 6 = 20 with 20 crossed out to 26, and the redo worked through to x = 10; the laptop is closed to one side — Three minutes, one changed number, and you know what stuck — no detector score required.

You run the homework through a detector. It says “likely AI-generated.” Now what?

You have a percentage, not a conversation. You can’t show it to a child and expect learning to follow, and you can’t defend it either — the same tools that flag polished AI writing also flag human writing, and a lightly edited AI answer sails straight through. A detector judges the finished text. It never saw how the work was made.

So stop judging the product. Check the process instead. Three minutes, no app, no accusation. All you need is the homework, a pen, and one small change you make yourself.

The check, minute by minute

Minute	Do this	You’re looking for
1	Ask to see the mess	Rough work, a crossed-out attempt, a first try
2	Change one thing, redo one step	The same skill, applied cold, without a screen
3	Ask one “why this step?”	A reason for a single decision, not a full explanation

Minute 1 — Show me the mess

Say: “Show me where you got stuck, or your first try.”

Real thinking leaves a trail: a crossed-out line, a restarted calculation, a note in the margin, a draft sentence that got rewritten. A pasted answer leaves none. If the work is typed, the trail can be version history, an earlier draft, or notes — same idea, different place.

No trail is not proof of cheating. Some children work neatly, some tidy up before handing work in. Treat a missing trail as a reason to do minutes 2 and 3, not as a verdict.

Minute 2 — Change one thing, redo one step

This is the heart of the check, and it takes about sixty seconds. Pick one small part of the finished work, change it, and ask your child to redo just that part — on paper, screen closed.

Worked example, maths. The homework says: solve 2x + 6 = 20. The answer is x = 7. You change one number: “Same question, but make it 2x + 6 = 26. Just that step.” The redo is 2x = 20, so x = 10. A child who did the original work gets there, maybe slowly. A child who only received the answer has nowhere to start — and that tells you more than any detector score.

Worked example, an essay. The homework argues that school should start later because teenagers need sleep. You change one premise: “Keep your first paragraph, but argue the opposite for one paragraph — school should start earlier. Give me one reason, in your own sentence.” Again, you are not asking for a second essay. You are asking for one decision, made fresh.

If your child can redo the changed step, the learning is there, whatever any detector says. If they can’t, you have found the exact place the understanding stops — which is exactly where help should start.

Minute 3 — One “why this step?”

Point at a single step in the original work and ask: “Why this step here, and not another one?”

This is not “explain the whole topic to me.” It is one decision: why divide here, why this piece of evidence first, why this word. A child who worked through the problem can usually answer for at least one step, even roughly. A child repeating a received answer often can’t answer for any step at all — the answer arrived whole, so no decision was ever made.

Read the result like this

What you saw	What it means	What to do next
Trail + changed step done + one “why” answered	The work is theirs, detector or no detector	Tell your child you can see it’s their work, then drop it and carry on with the evening
Trail, but the changed step fails	They started it and got lost partway	Redo just that step together tonight, then re-check tomorrow
No trail and the changed step fails	The answer was almost certainly received, not worked out	No lecture, no detector printout: pick one problem and do it from scratch together, on paper

Notice what none of these rows says: “punish” or “prove it.” The check is a diagnostic, not a court case. Its job is to find the gap, in three minutes, on a normal evening — not to win an argument about AI.

When the check doesn’t fit, bend it — don’t drop it

No rough work because everything is typed? Use version history or the previous draft as the trail, and still do the changed step on paper.
A subject with no numbers? Change a premise, a date, or a character’s choice instead of a number. The mechanism is identical: one fresh decision.
A younger child? Skip minute 3. Trail plus one changed step is enough, and keep the changed step tiny.
A child who gets upset? Say the frame out loud before you start: “I’m not checking whether you cheated. I’m checking which part is solid and which part needs another go.” If the evening is already tense, do it tomorrow — a check done calmly beats a check done tonight.
Only one minute? Do minute 2 alone. The changed step is the single most informative minute of the three.

The script, ready to use tonight

“Show me your first try or where you got stuck. Now — same question, but I’m changing one number [or one premise]. Do just that step, no screen. And tell me why you did that step, not the next one.”

That is the whole routine. Copy it, use it as written, and it works for any subject, any age that does homework, and any school — no detector subscription, no account, no software.

Why this beats a detector, in one line

A detector asks: does this text look like AI? This check asks: can this child do one changed step without it? The first question can be gamed by editing and can wrongly accuse a human. The second question can’t be faked in sixty seconds — and when the answer is no, it hands you the exact next step instead of a percentage to argue about.

Detectors sell certainty. A three-minute check gives you something better: a true picture of what stuck, and a clear, calm next move when it hasn’t.