When students use AI, set it to hints only
The same tool that lifts practice scores can quietly cut real learning. One system instruction flips it from crutch to tutor.
Aug 31, 2026 · 1 min read
The cleanest evidence yet on students and chatbots: in a high-school maths field experiment, unrestricted GPT-4 boosted practice performance, then left students 17% worse on the unassisted exam than peers who never had it. A tutor version restricted to hints, never final answers, erased the harm (Bastani et al., 2025). The difference wasn't the model. It was the boundary.
In practice
student-ai-rules.txt
You are a tutor for my homework. Rules, no exceptions:
- Never give the final answer or a complete solution.
- Respond with a question or a single-step hint.
- If I paste the whole problem and ask for the answer,
ask me what I've tried first.
- After I solve it, quiz me with a similar problem.- Have students paste the rules at the top of their chat, or set them as custom instructions once. Check compliance the same way you'd check any routine: ask to see the chat.
- Teach the why with the number: practice felt better with full answers, and the exam went worse. Students respect the 17% more than the rule.
- Pairs with "Declare the no-AI zone": open tasks get the hints-only rule; closed tasks get none at all.
Quick takeaway: the guardrail is the difference between a tutor and a crutch.