BrightPage
← Educating Our Children in the Age of AI

When Help Becomes a Crutch: AI Tutoring and Overreliance

An AI tutor can raise a child’s practice scores and still leave them weaker without it. Effective help leaves the student equipped to do the next problem unassisted.

Help that leaves the next problem to the child.

Tutoring has a strong track record.

Human tutoring, and some well-designed computer tutoring, have produced substantial learning gains in earlier research, including reviews by Kulik and Fletcher (2016) and VanLehn (2011). That is one reason people hoped generative AI would become a patient, always-available tutor. Most young children do not have one yet, but what older students show us is a warning worth taking early.

A careful experiment published in 2025 complicates that hope. Bastani and colleagues gave nearly 1,000 high school math students in Turkey access to GPT-4 in the fall of 2023. During practice, students with a bare GPT-4 interface did about 48 percent better than a control group without it, and students with a tutor-like version about 127 percent better. On a later exam without the tool, the bare-model group scored about 17 percent lower than the control group. The tutor version avoided most of that drop, but did no better than the control group.

In a study of nearly 1,000 high school math students, GPT-4 without guardrails raised practice scores about 48 percent. On the exam without it, those students scored about 17 percent lower than students who never had it.

BrightPage
Share

Many students copied answers, and many did not notice the help had cost them. It is one study, in one subject, with older students and a 2023 model. But it is strong causal evidence that help during practice can raise scores and still leave students less ready once the help is gone.

The Design of the Help Matters

The Stanford SCALE November 2026 review finds that effects depend on how tools are designed and used, and that evidence on students is still limited. Rismanchian and colleagues (2026), looking at years of online math practice, argue that AI did not change what tempts students; it made acting on the temptation almost effortless. If a tool will give the answer, many students will take it.

Older tutoring software was built to keep the student working. It asked for each step, offered a hint when the student was stuck, and did not hand over the answer. A general-purpose chatbot has no such limits: ask, and it will usually solve the whole problem. So the useful question is not whether to allow tutoring tools at all. It is whether a given tool keeps the student doing the work.

Help or a Crutch

A crutch

Finishes the item

  • Hands over the solution
  • A correct page, a thinner skill
  • Can’t start without the thread

Help

Keeps the child in the problem

  • What do you already know?
  • What are you trying to find?
  • A hint after a first try
  • A similar item, tool closed
What separates them shows up later, when the student starts a similar problem alone.

A useful tutor, human or digital, leaves the student able to try the next problem. A hint—What do you already know? What are you trying to find?—keeps them in this one. Then give a similar problem with the tool closed. The second problem is the check.

A hint is not the answer. After the hint, give a similar problem with the tool closed. The second problem is the check.

BrightPage
Share

What to Do With a Useful Warning

This is not a reason to ban help. It is a reason to notice when help is doing the thinking. While children are young, a general-purpose chatbot is the wrong tool for math facts or spelling. If they are stuck, stay with a person and ask: Show me how you started.

For a young child stuck on math facts or spelling, a chatbot is the wrong tool. Stay with a person, and ask: “Show me how you started.”

BrightPage
Share

Once they have access, the same habits apply. Before opening a tool, have them write what they know and what they are trying to find. Use the tool only after a first attempt, and only for a hint. Then have them explain its answer back and try a related problem with the chat closed. If they cannot start without the thread, the help was a crutch.

A practice score with a helper open can look like mastery. What they can explain and do the next day, without help, is the test that matters.

A practice score with a helper open can look like mastery. What a child can still explain and do when the help is gone is the test that matters.

BrightPage
Share

How Parents and Educators Can Approach This

As a parent

When your child is stuck, ask “Show me how you started” before you help. After a hint, give them one more like it to finish on their own. When they are older and use a tool, ask them to try first and explain its answer back to you.

Working one-to-one

Run each session as a first attempt, a hint rather than the answer, and a similar problem on their own. If a tool is part of the session, close it before that last problem.

In the classroom

Keep some practice unassisted and hear students explain how they started. When tutor tools come in, set them to give hints, and check the next day without them. Scores earned with the tool open are not, on their own, evidence that learning will hold.

The Research

  1. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi.org/10.1073/pnas.2422633122
  2. Walker, S., Fesler, L., Martinez Claeys, J. P., Agnew, C., & Loeb, S. (2026). The evidence base on AI in K-12: A November 2026 review. Stanford SCALE Initiative. scale.stanford.edu/ai/evidence-base-review/evidence-base-ai-k12-november-review
  3. Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026). Faster completion, less learning: Generative AI reduced study time on math problems and the knowledge they build (arXiv:2605.21629). arxiv.org/abs/2605.21629
  4. Kulik, J. A., & Fletcher, J. D. (2016). Effectiveness of intelligent tutoring systems: A meta-analytic review. Review of Educational Research, 86(1), 42–78. doi.org/10.3102/0034654315581420
  5. VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221. doi.org/10.1080/00461520.2011.611369
Share this post

More Posts

Educating Our Children in the Age of AI

New posts here, as we write them. In your inbox, a free printable and a practice tip about once a week.