Medical students and GP trainees should still learn to take a history, form an interpretation and explain a plan independently. They should also practise reviewing AI-generated material: recognising missing information, challenging a plausible differential and deciding what requires reassessment. AMIE's recent study makes that second task more concrete; it does not make the first obsolete.
The Lancet feasibility study published on 8 October 2026 evaluated supervised pre-visit conversations with real patients. It did not test medical curricula, measure long-term deskilling or evaluate iatroX. The teaching approach below is a proposed response to a changing workflow, not an intervention proven by the AMIE results.
Break the consultation into learnable tasks
An AI-assisted consultation can move several activities into the pre-visit period: eliciting a chronology, asking follow-up questions, proposing an interpretation and presenting a summary. In the AMIE pathway, that material was then available to the clinician conducting the usual appointment. The March 2026 methods describe the division of work.
For teaching, the important question is what the learner is expected to contribute at each stage. Reading a summary is not identical to gathering the information that makes it possible. Identifying an incorrect recommendation is not identical to generating an appropriate plan from an incomplete starting point.
An educator can therefore assess information gathering, interpretation, explanation and decision-making separately. This does not require every trainee to repeat every automated task in routine care. It requires selected opportunities to demonstrate the underlying abilities, so that polished assistance does not obscure what the learner understands.
Establish an unaided baseline before adding assistance
Begin with a fictional case and withhold the AI-generated summary. Ask the learner for a problem representation, the next question they would ask and the reason it matters. Then introduce further information and ask what changed.
The aim is not to reward a long list of remembered diagnoses. A concise, justified next step can reveal more understanding than an exhaustive differential. The learner should distinguish a likely explanation from a serious possibility that requires attention, and identify what cannot yet be concluded.
Necessary reference access should be specified rather than treated as cheating. An exercise assessing knowledge recall and one assessing clinical application can reasonably permit different resources. The essential point is to know what the learner produced before the assisted version was visible.
A reusable teaching session
The following session is an editorial teaching design. It can be run with a human-written simulated AI output; it does not require an actual AMIE account or access to the research system.
| Stage | Learner task | What the educator examines |
|---|---|---|
| Independent attempt | Gather the next useful information and propose an initial interpretation | Whether the learner can explain the decision without a prepared answer |
| Assisted review | Read a supplied AI-style summary and differential | Which useful additions and unsupported assumptions they notice |
| Challenge | Identify a missing question or contradiction | Whether the challenge would change the clinical reasoning |
| Revision | Produce a revised explanation and plan | What was accepted, rejected or changed, and why |
| Reflection and transfer | Review the error, then attempt a different case later | Whether understanding transfers beyond the original wording |
A completed worksheet should record a small number of consequential decisions, not every possible criticism. Otherwise, learners may discover that the exercise rewards scepticism for its own sake rather than appropriate use of assistance.
A fictional headache case with a misleadingly complete summary
In this constructed teaching case, the learner receives a history of recurrent headaches affecting work. The supplied AI-style summary labels the presentation as an established benign pattern and records no important change. The underlying dialogue, however, contains a statement that the current episode feels different; it was not explored.
The task is not to diagnose from this paragraph. The learner should identify the discrepancy, explain why the change matters and specify what must be clarified before accepting the prepared interpretation. Examination and further assessment may be required; a text exercise cannot perform them.
Next, give the learner an alternative summary that correctly preserves the uncertainty and highlights the patient's concern. Ask what can now be accepted without repetition. This prevents the lesson becoming "always distrust AI". Good supervision includes recognising useful work as well as identifying defects.
The educator can then change the presentation while preserving the reasoning problem. The later exercise might concern abdominal symptoms with a new feature that was lost during summarisation. Success is recognising the unsupported assumption in the new setting, not remembering the previous case's answer.
Teach communication, not just correction
A learner may detect a flawed summary yet struggle to explain the revised position to a patient. Ask them to say what is known, what remains uncertain and why further assessment changes the plan. Avoid treating a technical critique of the model as an adequate patient explanation.
The GMC's decision-making and consent guidance, checked on 10 October 2026, provides a professional foundation for dialogue tailored to the patient's concerns. It does not supply an examination score for a fictional exercise, but it helps keep the teaching focused on understanding rather than performance of a script.
Feedback should therefore address the relationship between reasoning and communication. Did the learner explain why the AI's initial interpretation was provisional? Did they avoid implying that every disagreement means the technology is useless? Could the patient identify the current next step?
Where question practice, Tutor and simulations fit
This article is published by iatroX, and its own learning formats are included here as options rather than as products tested in the AMIE study. In iatroX's October 2026 product description, the Socratic Tutor starts from an attempted question and uses targeted follow-ups to explore the learner's misconception. That is a relevant design for the point at which a learner needs to explain why an attractive answer was wrong.
Question-bank practice can supply a bounded decision before assistance appears. The published September 2026 simulation design adds voice or text practice, a pause-and-coach mode, uninterrupted examination practice and transcript-linked feedback. These are complementary learning methods; they do not replace bedside examination, procedural teaching or trained human observation. Clinician-reviewed cases are not the same as examining-body endorsement.
Under iatroX's UK offer for October 2026, Ask-iatroX and free question access remain free, without a trial expiry or verification gate. Ongoing paid learning access is £99 paid upfront for a year, equivalent to £8.25 a month billed annually, or £29 a month. The subscription includes question banks, Socratic Tutor, the study planner, iatroX Simulations and CPD tools together; simulations and CPD are not separate add-ons. Three monthly payments total £87 and four total £116, so annual payment becomes cheaper from the fourth month.
The educational value to assess is whether those methods address the learner's current professional goal. Access to unrelated examinations is not, by itself, a reason to subscribe.
Keep the assessment claim modest
An improved practice score does not establish clinical competence, examination readiness or protection against deskilling. A stronger evaluation would examine later performance on unfamiliar cases, the quality of explanations and the ability to identify consequential errors without the same assistance.
The proposed curriculum should therefore protect both independent performance and intelligent collaboration. The learner should be able to use a good summary efficiently, challenge an inadequate one and recognise when the task requires direct clinical assessment rather than further discussion with software.
Frequently asked questions
Does AMIE show that students no longer need to learn history taking?
No. The study evaluated a supervised care workflow, not the consequences of removing history-taking education from a curriculum.
Has AMIE been shown to cause clinical deskilling?
Not by this feasibility study. Long-term changes in independent competence would require a different evaluation with appropriate baseline and follow-up assessments.
Does using iatroX certify someone to supervise clinical AI?
No. Question practice, Tutor and simulations are educational formats, not certification that a clinician or another AI product is safe for a particular deployment.