The most consequential design fact in AI education fits in one sentence: a helpful assistant and a good tutor optimise for opposite things. An unrestricted chatbot is built to complete your task, answer immediately and minimise friction, which is exactly right for drafting an email and exactly wrong for learning, because learning is produced by the friction. A tutor should optimise for productive effort, error visibility, progressive disclosure and, above all, your subsequent performance when the tutor is gone. This is not philosophy; it is now measured, and the measurement is the reason "just use ChatGPT to study" is worse advice than it sounds.
The experiment that measured the difference
The PNAS field experiment with nearly 1,000 mathematics pupils ran the comparison directly. With AI available, a generic GPT lifted performance 48% and a purpose-built guarded tutor 127%. With AI removed, the generic-GPT group performed 17% worse than students who never had AI at all, while the guarded-tutor group avoided that harm, though without beating controls on the later unassisted exam. Same underlying model class, opposite unassisted outcomes, and the only variable was design: what the system would and would not do for the learner. The full evidence context sits in the hub at /blog/do-ai-tutors-improve-medical-education-evidence; this article is about the mechanism and what to do with it.
The crutch mechanism, named precisely
Why does frictionless help harm later performance? Three cognitive shortcuts compound. Copying: when the full answer is available, working the problem becomes optional, and many learners, honestly, stop working it. Premature answer exposure: seeing the solution before attempting one eliminates the generation effect, the well-established boost that comes from producing an answer, even a wrong one, before feedback. And reduced cognitive generation across the session: each assisted item trains the habit of consulting rather than retrieving, so the practice session quietly stops being practice. None of this requires laziness; it is what any efficient organism does when a lower-effort path to the same immediate outcome exists. The tutor's job is to close the path.
The guardrails that matter for medical learning
What the structured tutor did differently translates into five rules any medical learner can enforce. Require an initial answer: nothing is asked of the AI until you have committed to an option. Ask for reasoning: state why, in a sentence, before checking. Reveal progressively: request a hint, not the answer; a second hint before the explanation. Separate formative support from scored assessment: AI help during practice, never during anything that measures you. And retest without assistance: return to the topic days later, unaided, because that unassisted attempt is the only honest measurement of whether learning occurred. The same rules, engineered into software rather than willpower, are what distinguish a tutor-shaped product from a chat window bolted to a question bank, and the full self-enforced method is at /blog/answer-first-ai-second-clinical-learning.
What guardrails cannot solve
Honesty about the limits keeps the argument credible. Guardrails govern the interaction pattern; they do not correct incorrect source content, rescue a badly written question, calibrate a model's misplaced confidence, or substitute for clinical supervision, and a perfectly Socratic dialogue about a wrong answer is still a wrong answer taught well. Those failure modes belong to different layers, source grounding, question governance, uncertainty handling, and the novice-facing risk they create is serious enough to have its own evidence and its own article: /blog/novice-paradox-ai-confidence-medical-students. Guardrails are necessary; they are not the whole machine.
A safer prompt pattern for general-purpose AI
For learners who will use ChatGPT-class tools anyway, which is most learners, the pattern that imports the guardrails: "I will attempt this question first. Do not give me the answer or eliminate options. After I commit, ask me one question about my reasoning, then give me a hint if I ask, and only explain fully when I say 'reveal'. Afterwards, quiz me on the underlying concept with a new question." It is clumsy compared with a system that enforces the sequence automatically, willpower against a product optimised for frictionless answers is an unfair fight, which is precisely the case for integrated tutors, but it is far better than the default, and it costs one pasted paragraph.
Frequently asked questions
Is ChatGPT bad for studying, then?
Unguarded, for practice, the best evidence says it can be actively harmful to later independent performance; guarded, by product or by prompt, it becomes useful; the variable is the interaction design, not the model.
Do these findings from school mathematics transfer to medicine?
The mechanism, generation, effort, premature exposure, is domain-general cognitive science; effect sizes will vary, and the medical randomised evidence is assessed separately at /blog/what-20-randomised-trials-generative-ai-medical-education.
What should I look for in a product that calls itself a tutor?
Whether it makes you go first: attempt-first flow, progressive hints, reasoning elicitation and unassisted retesting, visible in the product, not the marketing.