When AI Becomes a Crutch: The Evidence

The most important piece of evidence about AI and children's learning is also one of the most sobering, and every parent considering AI tools should know it. In a large, carefully designed study, students who used an AI tutor performed dramatically better while they had it — and then significantly worse than their peers once it was taken away. The AI had not taught them; it had become a crutch. Understanding this study, and what it does and does not mean, is essential to using AI without falling into its central trap.

The study

Researchers ran a field experiment with nearly 1,000 students in a Turkish high school, giving some access to GPT-4 during maths practice sessions (Bastani et al., "Generative AI Without Guardrails Can Harm Learning," PNAS, 2025). The results came in two parts, and both matter.

With the AI, students soared. Access to the AI tutor improved practice performance enormously — by 48% to 127% depending on the version. On the face of it, a spectacular win for AI tutoring.

Without it, they crashed. But when the AI was removed and students faced an exam on their own, those who had used the unrestricted version scored 17% worse than students who had never had AI at all. They had leaned on the machine to get answers instead of building the skills themselves, and when the machine was gone, so was their apparent ability.

The headline is stark: the AI made students look like they were learning while actually preventing them from learning. It was a crutch that masked, and then worsened, their real ability.

Why this happens

The mechanism is human, not technological. When an easy source of answers is available, the path of least resistance is to use it — to get the problem solved rather than to struggle through and learn. Struggle is where learning happens; an AI that removes the struggle removes the learning. This is the same trap as rote learning and, more broadly, any shortcut that delivers the result without the understanding: it feels productive and leaves the child no more capable than before.

Children are especially vulnerable because they cannot always tell the difference between "I got the answer" and "I can now do this myself." The AI's help feels like competence. It is not.

The crucial second finding

Here is the part that too many summaries omit, and it changes everything. The same study tested a safeguarded version of the AI tutor — one designed to guide students toward solutions rather than hand them over. That version largely avoided the harm. Students who used it did not suffer the same collapse when it was removed.

So the study's real lesson is not "AI harms learning." It is: unrestricted AI harms learning; well-designed AI need not. The difference between a crutch and a genuine aid is entirely in the design — whether the tool makes the child do the thinking or does the thinking for them. This is the line we drew in When AI Re-Teaching Genuinely Helps, and it is the whole subject of AI Guardrails.

How to avoid the crutch trap

For a parent, the practical implications are clear:

Beware tools that just give answers. A general chatbot that will solve any homework problem on request is precisely the unrestricted version that caused harm. The child gets the answer and learns nothing.

Favour tools that guide. A well-built learning tool refuses to simply hand over solutions. It asks the child to attempt the work, re-teaches the specific gap, and makes the child apply the understanding themselves — building independence.

Watch for the tell-tale gap. If your child performs well with a tool but poorly on unaided tests or class work, that gap is the crutch effect in action. Real learning shows up in unaided performance.

Keep the struggle. Protect the productive struggle that learning requires. A child who never has to work through difficulty never builds the ability to. The right amount of help removes the obstacle, not the effort.

The balanced conclusion

It would be a mistake to read this study as a reason to ban AI, and an equal mistake to ignore it. The honest reading is that AI is a powerful tool that can help learning or sabotage it, depending entirely on how it is built and used. Used as an answer machine, it makes children dependent and worse off. Used as a guide that builds independence, it can genuinely help — the same conclusion the researchers themselves reached.

What to do with this

• Know the finding: unrestricted AI produced big gains that reversed into a 17% deficit once removed.
• Understand the cause: AI that gives answers removes the struggle that learning requires.
• Hold onto the hopeful half: a safeguarded version largely avoided the harm — design decides.
• Choose tools that guide and build independence, and watch your child's unaided performance.

The crutch is real, and the evidence is strong. But so is the escape from it. Choose AI that makes your child more capable on their own — and measure it by exactly that.

SmoothSeas is built to guide, never to hand over answers — so your child's ability grows with the tool and lasts without it. Start free.