Training remains the responsibility of educational and clinical teams, even when software performs more of the routine workflow. The challenge is to preserve the parts of ordinary work through which learners develop judgement, while removing genuinely unnecessary administration. That requires deliberate educational design, not a choice between keeping every manual task and allowing trainees to approve outputs they cannot explain.
Find the learning hidden inside the routine
Writing a summary can involve identifying the important facts, distinguishing an observation from an interpretation and explaining why a plan follows. It can also involve repetitive formatting that adds little educational value. Those activities should not be treated as interchangeable merely because they occur in the same document.
A training programme should identify which cognitive work the learner needs to perform and which mechanical work can reasonably be delegated. The aim is to preserve the learning objective, not the inefficiency of the old process.
The GMC's Outcomes for graduates, checked on 10 October 2026, sets out knowledge, skills and behaviours that new UK medical graduates must demonstrate. A polished AI-assisted output is not, by itself, evidence that the learner can demonstrate them.
Distinguish losing a skill from never developing it
An experienced clinician becoming less practised at a task is different from a trainee having fewer opportunities to learn it. Both deserve investigation, but neither should be assumed to occur inevitably whenever AI is introduced.
Automation could release time for better supervision and more patient interaction. It could also remove the moments when a learner normally discovers a gap, asks for help or receives corrective feedback. The outcome depends on how the work and teaching are redesigned.
A useful evaluation therefore measures the learner's development, not just the quality of the documents produced during the placement. Ask what they can explain independently and how they perform when the next case differs from the one rehearsed.
A fictional first-draft exercise
Imagine a trainee reviewing a fictional case before the AI-generated summary is revealed. The supervisor asks for a short problem representation, the principal uncertainty and the fact that would change the plan.
The trainee then compares their account with the generated version. They identify useful omissions from their own summary, challenge an unsupported inference in the AI output and explain the final revision. The learning comes from the comparison and explanation, not from declaring either author the winner.
Later, the trainee encounters a new fictional case with the same underlying reasoning problem but different surface details. They attempt it without the earlier answer visible. This tests whether the discussion helped them transfer understanding rather than merely remember a polished formulation.
This is a proposed teaching design, not a proven intervention for preventing deskilling. Its usefulness should be evaluated against the programme's actual learning goals.
Protect selected unaided work, not all work
Unaided assessment can reveal what the learner knows and where assistance is masking uncertainty. It should be used in appropriate educational settings, without withholding clinically necessary support from real patients.
The protected task can be brief and focused. A learner might formulate an initial differential, explain a management priority or identify the information missing from a summary. They do not need to manually reproduce every administrative step to demonstrate understanding.
The supervision plan should also state when to seek help. Independent reasoning does not mean isolated practice or reluctance to use references. The goal is to know why assistance is needed and how to assess what it provides.
Make supervision about decisions rather than polish
A supervisor reading an excellent note should ask what the trainee understood, accepted and changed. Useful questions include: Which claim needed checking? What alternative did you consider? Which missing fact would make this plan unsuitable? What did the AI suggest that you rejected, and why?
The GMC's standards on colleagues and safety, checked on 10 October 2026, support appropriate supervision and help for colleagues. Applying those principles to AI-assisted training requires visibility of the learner's reasoning, not just the final output.
Feedback should be specific enough to guide another attempt. "Be more confident" or "check the AI" is less actionable than identifying the unsupported assumption and asking the learner to apply the distinction to a new case.
Teach the work of challenging an assistant
AI supervision is not a single generic skill. It includes recognising missing information, distinguishing source support from applicability, checking action status and knowing when the system is outside scope.
A curriculum can use different fictional exercises for each. One case might contain a genuine citation supporting the wrong population. Another might describe a referral as complete when acceptance is unknown. A third might present a confident explanation that ignores a contradictory fact.
Do not reward disagreement for its own sake. The learner should also recognise when the tool has identified something important that they missed. Effective supervision means calibrated acceptance and challenge, not reflexive scepticism or automatic approval.
Use supplementary practice for a defined gap
As described in October 2026, iatroX offers question-bank practice, a Socratic Tutor linked to attempted questions and educational simulations. Its learning-design explanation describes complementary ways to practise knowledge and reasoning. These formats can supplement supervised experience; they do not replace bedside examination, procedural teaching or assessment by the training programme.
A learner who repeatedly overlooks uncertainty could attempt a relevant question, explain their reasoning, review feedback and practise a contrasting scenario. The important outcome is a better justified decision on a later case, not merely more platform activity.
The October 2026 subscription described by iatroX is £99 paid upfront for a year, equivalent to £8.25 a month billed annually, or £29 monthly. It includes the paid question banks, Socratic Tutor, study planner, simulations and CPD tools together; Ask-iatroX and free question access remain free without a trial expiry or verification gate. At those published October 2026 prices, three monthly payments total £87 and four total £116, so the annual option becomes cheaper from the fourth month. The relevant value is several learning methods for one professional goal, not access to unrelated examinations, and purchasing a bundle does not establish that a learner needs every included method.
Assess what the redesigned placement actually teaches
A proposed review should compare exposure, feedback and demonstrated performance before and after workflow change. Consider whether trainees still encounter ordinary cases, whether supervisors can identify misconceptions and whether learning transfers to unfamiliar situations.
Do not interpret every change as an AI effect. Staffing, case mix and supervision time may also change. Record those factors and avoid turning platform scores into a substitute for professional assessment.
The educational bargain should be explicit: remove repetitive work where that helps, but replace lost learning opportunities deliberately. Tomorrow's clinician needs both the ability to use assistance and the knowledge to recognise when the assistance is insufficient.
Frequently asked questions
Does automating routine work inevitably prevent trainees from learning?
No. It can either remove learning opportunities or create time for better teaching, depending on how the workflow and supervision are redesigned.
Should trainees avoid AI until they are fully qualified?
Not necessarily. Appropriate educational use can include independent attempts, critical comparison and supervised discussion, with clear boundaries around patient care and assessment.
Can question banks and simulations replace clinical placements?
No. They can provide supplementary deliberate practice, but they do not replace supervised patient contact, physical examination, procedures or workplace assessment.
Practise clinical reasoning with question-linked Tutor support →