When should an AI tutor stop helping?

If an AI tutor can rescue a child from every difficult step, when should it choose not to?

I think this is the harder question behind the school AI debate. A tutor that leaves a learner stuck is not much use. But a tutor that does all the difficult thinking can also fail at its job, even while producing a page of correct answers.

The distinction becomes visible in a high-school mathematics experiment in Turkey, conducted in autumn 2023 and published in PNAS in June 2025. Students given a general-purpose AI interface performed 48% better during assisted practice, but 17% worse than the control group when subsequently tested without it. Those are relative grade differences, not percentage-point changes in pass rates.^[1]

The result asks us to separate two products. One helps a child finish the problem in front of them. The other leaves the child better able to approach the next problem.

Sometimes they will be the same product. We should not assume they always are.

Help with what, exactly?

Imagine a learner facing an unfamiliar algebra problem. This is an illustration, not a classroom observation.

The learner might be unable to remember a rule. They might remember the rule but not recognise where to use it. Or they might know the method and simply have made an arithmetic mistake.

A complete worked solution could remove all three difficulties from the page. But it would not necessarily reveal which difficulty the learner had. That is why “the tool answered correctly” does not settle whether the interaction taught the right thing.

One response is to tell the AI never to give answers. There is evidence for more carefully designed help: the same Turkish experiment included a tutor using teacher-informed hints and restrictions on giving away solutions. It largely mitigated the loss on the unassisted test, but did not establish a positive test-score effect.^[1]

That is an important boundary. Avoiding the harm is not yet proof of improved learning.

Nor would I take the result as a reason to leave a beginner struggling indefinitely. The US Institute of Education Sciences' research guide recommends alternating worked examples with problems learners solve themselves.^[2] The principle predates current generative AI. Seeing how a problem is solved and attempting one are not necessarily competing approaches; they can be successive parts of learning.

The problem, then, is not the existence of an explanation. It is an interaction that never asks the learner to do anything with it.

An AI could demonstrate a method, ask the child to explain one step, and then offer a related problem without completing it. If the child cannot begin, it could restore a narrower piece of help. That is a proposed teaching sequence, not a tested claim about a particular product.

What makes the sequence interesting is that assistance has a purpose and a point at which it recedes. The aim is not to minimise help from the first minute. It is to stop providing the same help once the learner may be ready to work without it.

Withdrawal is not abandonment

There is an obvious objection: what if the child experiences the withdrawal as failure and disengages?

That possibility is why I would not turn this into a rigid timer or an arbitrary limit on questions. “You have used your three hints” tells us nothing about whether the child understands. The decision needs to respond to the work, not merely the number of interactions.

There is also counterevidence to any blanket claim that AI assistance undermines learning. A teacher-supported, six-week programme in Edo State, Nigeria reported an English gain of 0.24 standard deviations. The researchers attribute the result to the full programme, not the chatbot alone.^[3] It does not identify the ideal moment to withdraw help, but it shows why the question has to remain about the design of assistance rather than assistance itself.

For an Indian school considering an AI tutor, I would want to see one concept followed through this process. Not a demonstration in which the AI answers a sequence of questions brilliantly, but one in which the learner gradually takes over the reasoning.

That is more specific than asking whether the platform measures outcomes. It changes what we want the product to do during the lesson.

A useful demonstration would show the learner attempting something comparable without the preceding prompts, then explain what the system does if they cannot. The independent attempt need not be a high-stakes examination. It is a way to discover whether an explanation has become usable knowledge.

I would still distinguish that short-term evidence from lasting understanding. The studies here do not settle long-term effects in Indian classrooms, and they do not establish one withdrawal schedule for every learner or subject.^[1][3]

But they do support a narrower purchasing principle: a learning tool should be able to explain how it hands the work back to the learner. More interactions may be useful along the way; they are not the destination.

The tutor's job is not to be needed at every step. It is to help the child take a step that previously required it.