Doctors would still need supervised experience of ordinary presentations, deliberate practice with uncertainty and assessment of both assisted and independent reasoning. If AI handles more routine cases, training cannot assume that reviewing its exceptions will teach everything required to manage those exceptions. This is an educational design challenge, not evidence that automation inevitably reduces competence.
Nolla's emergence provides a reason to ask the question, but the workforce scenario below is not a claim that Nolla has caused deskilling. It considers what educators would need to preserve if condition-specific AI services changed the mix of encounters reaching trainees and supervising clinicians.
The objective is not to make doctors repeat every task software can perform. It is to retain the knowledge and practical judgement needed to recognise an unsuitable plan, explain why it is unsuitable and take over effectively.
The routine case contains several different learning tasks
An ordinary consultation may involve gathering information, deciding which details matter, considering alternatives, explaining uncertainty and agreeing a plan. Some of its administrative work may be repetitive without being educational. Other parts may reveal how a clinician distinguishes a typical presentation from one that only initially appears typical.
A training programme should separate these elements before deciding what to automate. Requiring a trainee to repeatedly format a document is not the same as asking them to justify a working assessment. Removing the first need not remove the second.
The proposed risk is that learners receive an increasingly polished interpretation without sufficient opportunity to construct one. A supervisor then sees a reasonable output but cannot tell whether the trainee identified the decisive information independently or simply accepted the software's selection.
There is also a positive possibility. Automation might release time for observation, patient discussion and feedback. Whether it does so depends on how that time is used. Neither beneficial learning nor lost learning should be assumed from the presence of AI alone.
Recognising an error is not the same as generating a plan
A learner reviewing an answer has a starting structure. A learner facing an uncertain presentation must first decide what structure is appropriate. These are related but different tasks.
An educational assessment should therefore include more than the ability to spot a conspicuous error in an AI-generated plan. It should ask the learner to identify missing information, explain why a particular uncertainty matters and decide whether remote assessment remains suitable.
It should also avoid turning supervision into indiscriminate scepticism. Rejecting every suggestion does not demonstrate good judgement. A capable reviewer should be able to accept a well-supported recommendation, correct a limited problem or stop the pathway when its assumptions no longer hold.
The useful competence is calibrated intervention: understanding enough to know what needs changing and why. That is more specific than being comfortable with AI or having completed an onboarding tutorial.
A fictional three-stage case for an educational session
The following is a proposed teaching case, not a patient record, a Nolla incident or a claim that an identical station exists in the iatroX catalogue.
In the first stage, a fictional adult reports a new facial skin problem through a remote service. The learner receives a brief history and an image description, but no generated diagnosis. They must state their provisional interpretation, identify missing information and explain what would make this pathway unsuitable.
The exercise is not asking them to prescribe from an incomplete vignette. It is asking whether they recognise the incompleteness before a polished answer makes it less noticeable.
In the second stage, the learner sees a deliberately incomplete AI-generated plan. It confidently describes the concern as routine and proposes continuing within the service. Its explanation is fluent, but it has not established previous treatment, the full medication history or whether the problem extends beyond the visible area.
The learner should identify the unsupported assumptions without inventing findings that are not supplied. They should distinguish what is reasonable in the draft from what cannot yet be justified. The trainer can then ask which unanswered question would most alter the next step.
In the third stage, new information appears: the patient reports symptoms beyond the original concern and a recent change in another treatment. The learner must revisit the assessment and explain why the original pathway may no longer fit.
There is no requirement to reach a particular diagnosis from this limited scenario. The performance being assessed is whether the learner responds appropriately to changed information rather than defending the initial interpretation.
Make AI supervision a set of observable skills
A proposed assessment rubric could examine whether the learner distinguishes missing from negative information, checks the scope of the service and notices when a physical assessment may be needed. It could also assess whether uncertainty is communicated honestly and whether an escalation includes the information another clinician needs.
The rubric should include the ability to explain an appropriate plan, not just criticise a flawed one. Otherwise training risks rewarding objections without testing whether the learner can offer a workable alternative.
A separate recovery exercise could interrupt a partially completed workflow. The learner would need to establish what has already happened, what remains pending and whether repeating an action could create a problem. This tests practical understanding rather than prompt-writing fluency.
These are proposed competencies to evaluate. They are not a validated certification framework, and a strong simulation performance should not be equated automatically with independent workplace competence.
Test assisted and independent performance without compromising care
An educational sequence could begin with an independent attempt, introduce an assisted version, require an explanation of any disagreement and revisit a related case later without the original answer visible. The later case should change an important feature so that memorising the previous explanation is insufficient.
This design would let educators ask whether the learner transfers understanding rather than simply producing a better answer while assistance is present. It would also allow the possibility that well-designed assistance improves learning.
Unaided testing belongs in appropriate educational settings. It should not mean withholding necessary support from real patients or expecting clinicians to work without ordinary reference resources. The comparison must match the actual capability being assessed.
Confidence should be recorded separately from performance. A learner who feels reassured by an explanation may or may not be better able to identify when it stops applying. Feedback should address that distinction without treating low confidence as a desirable endpoint in itself.
Continuing development should follow the changing role
The RCGP's mandatory training and CPD guidance, published on 8 April 2026, emphasises maintaining competence across a GP's whole scope of work. A move into supervising AI-supported pathways would create learning needs within that scope, not remove the need for development.
A clinician's learning plan might therefore include evaluating incomplete histories, reviewing externally initiated care and managing a failed handover. The plan should reflect the work actually undertaken rather than a generic course labelled AI literacy.
Evidence should record what the clinician did, what they understood differently and what they intend to change. An automatically generated reflection is not evidence of learning until the clinician has reviewed and personalised it.
How iatroX can contribute without replacing clinical training
As described on 5 October 2026, iatroX question practice combines adaptive sequencing and spaced repetition, while the Socratic Tutor starts from an attempted question and explores the learner's reasoning. These are relevant ways to practise identifying distinctions, not proof that a subscription prevents deskilling.
The iatroX Simulations product, launched in September 2026, includes voice and text practice, coaching and uninterrupted examination modes, and transcript-linked feedback. Clinician-reviewed cases are not examining-body endorsement, and digital practice does not replace supervised bedside, procedural or interpersonal learning.
Under iatroX's subscription terms on 5 October 2026, question banks, Tutor, study planner, Simulations and CPD tools are included together for £99 paid upfront annually, equivalent to £8.25 per month billed annually, or £29 monthly. Simulations and CPD are not separate add-ons; three monthly payments total £87 and four total £116, so annual becomes cheaper from the fourth month. Ask-iatroX and free question access remain free without a trial expiry or verification gate.
The value is a set of learning methods for a relevant professional goal, not access to unrelated examinations. CPD records document reviewed learning; they should not be described as automatically accredited CME or certification to supervise a particular AI service.
iatroX's companion article on evaluating the whole AI care pathway considers the parallel research question: how to measure the system's immediate performance without confusing it with the competence its users retain.
Frequently asked questions
Does automating routine care necessarily deskill doctors?
No. The effect would depend on what work changes, what learning opportunities remain and how competence develops over time.
What should doctors practise to supervise AI-supported care?
Practise identifying missing information, testing assumptions, recognising when the pathway no longer fits and communicating an appropriate revised plan. Include both independent assessment and review of assisted work.
Can question banks and simulations replace clinical experience?
No. They can contribute structured practice and feedback, but they do not replace supervised patient care or demonstrate workplace competence on their own.
Practise clinical reasoning with the iatroX Socratic Tutor →