AI homework help that does not do the homework

A show of hands in any staff meeting settles the question quickly. Students have AI homework help. They had it before the policy existed, they have it on their phones whether or not the school issues devices, and the ones who are struggling most are often the ones using it most heavily. The useful conversation is not whether to allow it. It is what shape of help you are prepared to defend at a parent meeting, and how you keep any evidence of learning once the work moves out of the room.

The three kinds of homework help, and why only one of them is a problem

Lumping everything under "AI homework help" makes the policy conversation impossible, because the same phrase covers three completely different transactions.

1. Answer delivery. The student photographs the question and receives the finished answer, sometimes with working attached. This is what most people picture, and it is the only one of the three that genuinely destroys the assignment.
2. Explanation on demand. The student asks why a step works, or asks for the same idea said a different way. This is a textbook with a search box. It has existed in some form for decades and no policy has ever tried to ban it.
3. Coached practice. The student attempts the work, gets stuck, and receives the next question rather than the next step. This is what a good tutor does, and until recently it was available only to families who could pay for one.

Most school AI policies are written as though only the first exists. That is why they read as blanket bans, why they are unenforceable, and why they collapse the moment a diligent student points out that they used it to check their own reasoning.

What actually breaks when answer delivery becomes normal

It is worth being precise about the damage, because "cheating" is too blunt a word to plan around.

The teacher loses the diagnostic signal

Homework has never mainly been about practice volume. Its real function is diagnostic: twenty-eight pieces of work tell a teacher where the class is before the next lesson is planned. When a meaningful share of that work is generated, the signal does not just get noisier. It gets inverted, because the students most likely to hand over a perfect worked solution are the ones furthest behind. The teacher plans tomorrow from data that points the wrong way.

The student loses the productive struggle

The learning happens in the gap between not knowing and knowing, and that gap is uncomfortable by design. An instant answer closes it before anything happens. Students are not being lazy when they take the answer; they are doing what every human does with an available shortcut under time pressure at 10pm.

The school loses the ability to trust submissions

This is the failure that ends with a ban. Once a teacher cannot trust any submission, every piece of work has to be re-verified in class, which costs more time than the homework saved. At that point banning the tool looks rational even though the ban is unenforceable, and the school ends up with the worst outcome: no AI help for the students who need it, and continued unsupervised AI use by the students who were always going to.

The policy question is about default behaviour, not about access

You cannot control whether a fourteen-year-old opens a chat app at home. You can control what the school's own tool does by default when a student gets stuck, and you can build the routine around that.

A tool whose default is a hint changes the economics of the shortcut. If the school-provided option asks "what is twenty percent of forty-eight?" and the student has to answer it, then the fastest path to a finished assignment still runs through the student's own head. If the school-provided option hands over the solution, it is simply a faster version of the thing you were worried about.

The practical test for any homework AI: after a session, can the teacher tell what the student could and could not do? If the honest answer is no, the tool has removed information rather than added it.

Six rules that survive contact with a real school

These are deliberately narrow. Broad principles about "responsible use" do not help a teacher at 8am on a Tuesday.

1. Name the assignment type on the assignment. Three labels, printed on every task: AI-free, AI-assisted, AI-required. Students stop guessing and staff stop improvising. Most schools find that the AI-free category shrinks to the small number of tasks where it genuinely matters.
2. Set hints as the default for practice work. Answers stay available for the moments a student is genuinely stuck and there is no adult to ask, but the student has to choose to switch, and that choice is visible.
3. Require the reasoning, not the answer. If the artifact you collect is the explanation of the method, generated answers stop being useful shortcuts and the assignment becomes self-enforcing.
4. Keep one short in-class checkpoint. Five minutes, one problem, no devices, once a week. It is not surveillance; it is the calibration point that tells you whether the homework signal is real.
5. Do not build the policy on detection. AI detectors produce false positives, they disproportionately flag students writing in a second language, and an accusation you cannot substantiate costs far more than the misconduct. Design assignments that do not need a detector.
6. Write the privacy line into the policy. If the school recommends a tool, the school has taken a position on where student work goes. Say which tool, say what happens to the work, and say it in the same document. Our note on ChatGPT and student privacy covers where that usually goes wrong.

Homework that produces evidence instead of destroying it

The most useful reframing we have seen: stop treating at-home AI as a leak to be plugged and start treating the session as an instrument. A tutoring exchange contains far more information about a student's thinking than the final answer ever did. The problem has only ever been that nobody was capturing it.

That is the shape of the fix. If the student's attempt, the point where they stalled, the hint that unstuck them and the skills involved are all recorded, then homework produces a richer diagnostic than the graded page it replaced. The teacher opens a per-skill record instead of a stack of answers, and the class arrives with the misconceptions already grouped.

What that looks like in practice

A middle-school class works ten percentage problems at home. Nineteen students finish without a hint. Six stall in the same place: they apply the second percentage to the original number rather than the new one. Three stall earlier, on converting a percentage to a decimal at all.

The next morning the teacher has two groups rather than a class-wide reteach, an activity for each, and a sentence per student that can go into a progress report unedited. Nobody entered a score. Nobody scheduled an assessment. The evidence is a by-product of the homework the students already did, which is exactly the mechanism described on how it works.

What to ask a vendor before you recommend anything

• What is the default when a student is stuck? If the honest answer is "it gives the answer", it is a homework completion tool, whatever the marketing says.
• What does the teacher receive afterwards? A transcript is not progress data. Ask to see the actual teacher-facing artifact.
• Does it work over our material? A tool locked to its own content library gives you evidence in somebody else's scope and sequence.
• Is student work used to train models? Ask for the answer in writing. It is the question your legal review will ask, and it is the one that kills pilots late.
• What happens to the data if we leave? Export and deletion should be a sentence, not a support ticket.
• What do you not have? A vendor that cannot name a single missing certification or capability is not being straight with you. Ours are listed on the student privacy page.

The equity argument, stated plainly

Private tutoring has always been the cleanest predictor of who gets unstuck and who quietly gives up. Families who can pay for an hour of a good tutor buy exactly the thing described above: someone who will not hand over the answer, who diagnoses the misconception, and who tells the parent what to work on next.

An AI tutor that does the same thing at a per-student cost of pennies is one of the few genuinely levelling applications of this technology. An AI tool that hands over answers is the opposite: it widens the gap while appearing to close it, because the students with the most support at home are the ones who use it as an explanation engine, and the students with the least use it as an answer engine and fall further behind.

That is the whole reason the default matters more than the policy. See how the diagnosis works in math, or read our classroom policy template for the exact wording schools have found enforceable.

A ten-minute version of this article

If you have to brief a department tomorrow, this is the whole thing:

• AI homework help is already in use. Debating access is a waste of the meeting.
• Answer delivery is the only genuinely destructive mode, and it is destructive because it inverts the diagnostic signal, not because it is "cheating".
• Set hints as the default on any tool the school recommends, and make the label on each assignment do the policy work.
• Collect reasoning rather than answers, keep one short device-free checkpoint, and do not build on detection.
• Choose a tool that hands the teacher a per-skill record, so homework generates better evidence than it did before AI existed.
• Get the training-use and deletion answers in writing before the pilot, not during procurement.