Dr. Fady Hannah-Shmouni’s recent KevinMD essay, “Continuing medical education needs a tutor, not a lecture,” makes a case I share without reservation. Slide decks and ten-question quizzes survive because they are easy to accredit, not because they necessarily produce durable learning. An AI tutor that presents a case, waits for the clinician’s answer, and probes the reasoning behind it is far closer to how expertise actually develops.
His guardrails are the right ones: grounding, physician oversight, transparency, privacy, equity, and honest measurement. He is right that reasoning matters more than recall alone. I would add one easily overlooked guardrail.
The tutor must protect the learner’s effort, not only the accuracy of the content.
A perfectly grounded tutor can still teach badly if it makes thinking optional. Guarding against that requires four commitments.
Real physician voices, twice a week
Free, and one click to unsubscribe.
Keep the difficulty that does the work
Learning science offers a counterintuitive finding: Conditions that make learning feel smooth can make it shallow, while conditions that demand effort can make it last. Researchers call these desirable difficulties: retrieving from memory instead of rereading, spacing practice over time, interleaving problem types, and generating an answer before seeing one.
An adaptive tutor notices struggle and helps; it notices hesitation and hints. Each intervention can remove effort that builds durable knowledge. Fluent explanations also create an illusion of competence. A clinician who follows a beautifully reasoned walkthrough can mistake recognizing good reasoning for producing it.
Think of a toddler learning to walk. She wobbles, stumbles, and falls, and every attempt challenges her nervous system to adapt. A parent who carried her across the room each time she faltered would get her there faster, but she would have fewer opportunities to learn. Crossing the room was never the goal; learning to walk was. An AI tutor that rescues a clinician at the first sign of struggle risks carrying her across the room.
Difficulty is not inherently beneficial. Productive struggle requires appropriate challenge and guidance, not frustration.
The design question is not how fast the tutor can resolve confusion. It is when to intervene and how much help to provide. Dr. Hannah-Shmouni’s example of returning to a missed concept a week later is exactly right. Spacing should be a foundational design principle, not an optional feature.
Effort before augmentation
The simplest form of this guardrail is a sequencing rule. The clinician commits first, in her own words: a differential, a next test, a management plan, and her reasoning. Only then does the AI respond.
Order matters because AI can amplify whatever the human brings to it. A clinician who reasons first and then consults the tutor has her thinking sharpened, challenged, and corrected. A clinician who consults first may receive an excellent answer and a feeling of understanding without having generated the reasoning herself. Same tool, same content, but potentially opposite effects on learning.
This is also the habit we need clinicians to carry into practice, where AI decision support is already arriving. CME that trains “Think, then check” builds the reflex we want at the bedside. CME that trains “Ask, then accept” risks building the opposite one.
Make the learning process verifiable
Dr. Hannah-Shmouni asks AI tutors to show the evidence behind every teaching point. The same standard should apply to the learner. If credit is awarded for reasoning, the reasoning process should be visible.
An agentic tutor is unusually well placed to provide this. It can record what the learner committed to before any help, which hints she requested, how her answer changed after feedback, and whether a corrected misconception stayed corrected weeks later. That record can help educators distinguish independent reasoning from heavily scaffolded answers.
Such a record complements independent assessment rather than replacing it, and hesitation or requests for help should never be read as deficiencies. Done transparently, with privacy protections and learner control, it need not be surveillance; it can be the learning equivalent of showing your work.
Measure independence from AI
Dr. Hannah-Shmouni says we should study these tools the way we study a new drug. I would sharpen that by naming a critical endpoint. The outcome that matters is not only performance with the tutor. It is performance without it.
A clinician who scores well while AI scaffolds every step has shown that the combined system works. She has not yet shown that she learned.
Programs should test unaided reasoning on unfamiliar cases after a delay, assessing retention and transfer. Can the clinician retrieve knowledge, construct a differential, recognize uncertainty, and defend a decision without assistance?
Assisted performance matters too, but it differs from independent competence. Both deserve measurement.
If the gains vanish when the tutor is removed, we have built an effective assistance system, not a more capable clinician. If they persist, we have something worth accrediting.
The question to ask
None of this argues for slowing down. It argues for building AI tutors that leave clinicians more capable on their own, not more dependent on the tool.
Dr. Hannah-Shmouni closes by urging clinicians to ask hard questions of AI-driven education programs. I would add one:
When the tutor is turned off, what can the learner do that she could not do before? If a program cannot answer that, it has simply carried the clinician across the room.
Crossing the room was never the goal.
Disclosure: I am a co-founder and the chief learning officer of an education company that develops learning tools for health professions students, and I operate a separate, independent learning-systems company. I also helped originate an AI tutoring platform; I hold no equity in it and receive no compensation from it.
John Sullivan is a physician, neuroscientist, and associate professor who serves as chair of basic sciences and research for Touro University’s physician assistant programs. With a background in neurology and neuroscience, his work focuses on learning science, clinical reasoning, and the integration of artificial intelligence into health professions education.
He develops innovative educational approaches and AI-powered learning tools designed to strengthen critical thinking, promote independent learning, and prepare future clinicians for an increasingly technology-driven health care environment. He is also a published commentator on the implications of artificial intelligence for education, medicine, and human cognition, with work appearing in the Los Angeles Times.
