How to choose an AI tutor that teaches instead of doing homework

TLDR: There is no “best AI for studying” without naming the job. A tutor that asks questions and gives hints, an equation scanner and a flashcard generator may all carry an education label, but they are not substitutes. Choose the job first, filter age and access next, and only then compare three products with a test that ends outside the application.

The homework is right and the test is not

A Tuesday in October, the maths grade slips. Somebody in the class group chat drops an app name, it gets installed that night, and the improvement is immediate: homework starts coming back correct, the nine o’clock argument stops, the notebook looks perfect. Three weeks later the test comes back with the same grade as before, or worse.

The tool did not fail. It did exactly what it was asked to do, which was finish the exercises. What nobody decided at install time was whether the result being bought was the completed homework or the understood concept, and those are two different products sold under one word. This guide is about deciding that before paying, and about verifying it with a test the app’s own demo cannot pass on your behalf.

1. Separate the job before looking at brands

The request “an app to help my child study” contains at least five separate markets. Understanding a concept calls for a guided tutor. Photographing a calculation calls for a maths solver. Speaking English calls for voice, correction and conversational continuity. Turning a PDF into revision calls for fidelity to the source. Learning to read calls for a child-specific experience that listens and waits.

When those jobs are mixed, the comparison rewards the tool with the most features or the strongest brand rather than the one solving the problem. Khanmigo and ChatGPT Study mode compete for guided explanation, while Photomath enters when the starting point is a photographed problem. Quizlet replaces neither: it turns personal material into cards and mock exams.

Write the observable result in one sentence. “Understand fractions and solve a new problem independently” is a specification. “Study better” is not.

Worth saying before any money moves: if what is on the table is a set of notes that already exists and needs turning into revision, that job has its own method and needs no tutor at all. It lives in AI study materials.

2. Filter by age, account and country

Age is not footer detail. OpenAI says ChatGPT is not meant for children under 13, without conditioning that on supervision, and that in an education context at that age an adult must conduct the interaction. Ages 13 to 18 require parental consent, and its terms add whatever minimum age your country requires on top. Khanmigo has learner and family access, but that subscription is sold only in the United States, with US residence and a US billing address. Microsoft’s Study and Learn agent requires age 13+ and a Microsoft 365 Education licence supplied by an institution.

That changes the ranking before quality is scored. A family outside the United States cannot treat Khanmigo as a purchasable recommendation, though the teacher route and the school or district route do exist outside it. A teenager without an education account cannot receive Study and Learn. A young child should not end up in a general assistant as if age conditions did not exist.

For school material, check the data path too. Consumer and institution-managed accounts may have different terms. If your school or district keeps an approved-tools list, that list outranks any household preference, and it is the cheapest question to ask before subscribing rather than after. Do not upload reports, names, identifiable work or information about a child’s difficulties to a service that the family or institution has not approved for that context.

3. Run a test that finishes outside the app

Demos mislead because almost every product can produce a convincing answer. The useful test measures transfer:

1. Choose a concept the learner does not yet master.
2. Explicitly ask the tool not to reveal the result and to use questions and hints.
3. Introduce a real mistake and see whether it detects it and changes explanation.
4. Close the application.
5. Give a different problem using the same concept.

If the learner only completes the example the AI solved in front of them, you measured answer quality. If they solve the new problem and explain why, you measured learning.

Step 3 fails more products than any other and is the one almost nobody runs. A good tutor, faced with a wrong answer, does not repeat the same explanation more slowly: it changes route, asks where the learner lost the thread, or drops to a smaller example.

A product that meets an error by restating itself is reading from a script, and by the third session the learner has stopped reading back.

For a study-from-notes app, change the test: put one small exception in the PDF, generate five questions and check every answer against the original page. For languages, speak for ten minutes in the same situation and see whether correction helps without breaking the conversation. For children’s reading, observe how long it waits before intervening and measure two weeks, not first-day excitement.

The four exceptions that always show up

The test above sorts products well. These four cases slip around it, and none of them is fixed by switching apps:

The learner who works out how to work the tutor. By week two they have found that “I already tried and I can’t get it” or pasting the question with “just tell me if this is right” turns a guided tutor into a solver. No parental setting catches it, because there is nothing to block. It shows up outside the app, on the new problem, which is exactly what step 5 exists for.

The right method on the wrong syllabus. The explanation is good and does not match the classroom: a different long-division layout, different notation, a different order of topics. In primary that causes more confusion than help, because the learner has no basis yet for knowing that two valid routes exist. Test it on a topic already covered before using it on a new one.

Two children, one account. The session carries over the level, the language and the explanations of whoever used it yesterday, and the younger one gets answers pitched at the older. Separate accounts, or where the product does not allow that, a fresh session every time, which is worse but works.

Six weeks with no change. If use has been consistent and the observable result has not moved, the problem is neither the product nor the effort. That is the point to raise it with the school, and the support workflow starts in differentiated materials rather than in a fourth app.

What to measure in the first weeks

Three numbers counted at home, with no tooling and without trusting the app’s progress dashboard, which measures usage rather than learning:

Who decides what, and what no app decides

The account, the country and what gets uploaded are an adult’s decision, always, and that decision comes before any quality test. Where school material is involved the institution outranks the household, and asking costs one email.

None of these tools is used to diagnose dyslexia, dyscalculia or another learning difficulty. They detect that a problem was answered wrongly, not why, and that gap is exactly what separates a consumer product from a professional assessment. Persistent problems call for human assessment.

From the other side of the desk, if what you want is how to set work so that AI cannot do it for the student, that is teaching craft and it lives in student assignments in the age of AI.

The short decision

The current options and their limits are grouped under AI for education and the best learning tools.

Frequently asked questions

When is paying for an AI tutor not worth it?

When the problem is not explanation. If what is broken is that there is no fixed study time, that the material is incomplete, or that the subject was abandoned two months ago, no subscription repairs it and it creates the feeling that something is being done. The free tier of nearly every product here is enough to find out whether explanation was the problem. If nothing changes on the free tier, the paid one will not change it either.

Does an AI tutor replace a private tutor?

For explaining a concept at eleven at night, yes, and far more cheaply. For holding the habit together and noticing something is wrong before the grades notice, no. A household paying mostly for the second thing is not paying for explanation.

Should the teacher know it is being used?

It helps. If the learner turns up using a different method for the same operation, the person who can head off the confusion is the one teaching the class. Schools also set their own rules about what may be used on graded work, and that comes before anything the family decides.

Is it useful for cramming a test the week before?

For revision on material that already exists, yes, and a flashcard app beats a tutor there. For understanding a topic from zero that was never covered, the week before is late, and the result will be recognising the question without being able to answer it.