You do not need to know every answer before using a reference tool, but you do need a way to recognise the limits of your verification. For a junior doctor, that means separating what can be checked directly, what requires further learning and what needs supervision. Understanding an explanation is not the same as independently establishing that it is correct.
Familiarity is an insufficient test
Imagine a fictional trainee practising a dizziness case. The supplied history is incomplete, and an AI answer offers a tidy diagnosis with a convincing explanation. The trainee recognises the terminology and thinks the reasoning sounds right. Yet they cannot identify which additional information would make that interpretation unsafe to accept.
The difficulty is not necessarily a lack of effort. The trainee is trying to evaluate a conclusion while still learning the distinctions on which it depends. The explanation may supply exactly the language that makes the conclusion feel familiar, without revealing what has not been established.
A better learning question is: "Which fact would I need to verify before this explanation becomes applicable?" This moves attention from whether the paragraph sounds professional to whether its underlying propositions are supported. It also allows the learner to say, usefully and precisely, what they do not yet know.
Divide verification into manageable tasks
A novice may be unable to judge a complete management plan but still be able to perform important checks. Does the cited guidance concern the relevant population? Is the explanation answering the question that was asked? Has it turned an unknown finding into an assumed negative? Does the proposed action depend on information absent from the case?
These checks do not establish overall safety. Their value is that they make the remaining uncertainty visible. A learner who identifies an unresolved dependency can seek targeted help rather than either accepting everything or starting an unfocused search from the beginning.
For clinical practice, the supervisory relationship remains essential. The GMC's guidance on delegation and referral, reviewed on 10 October 2026, addresses appropriate instructions, competence and the ability to obtain support. It does not turn access to an AI answer into evidence that a trainee can independently perform a delegated task.
Try an attempt, a source and a contrasting case
Here is a proposed learning sequence for the fictional dizziness exercise. First, the learner gives an initial answer and identifies the uncertainty rather than hiding it. They might say that their interpretation depends on the timing of symptoms, the circumstances in which they occur and information not yet supplied.
Next, they inspect an appropriate original reference. Instead of collecting an entire chapter, they identify the passage that addresses the disputed distinction. They explain why it supports, limits or contradicts the AI response. When the source does not resolve the issue, that becomes the question for a tutor or supervisor.
Finally, the educator changes a consequential feature of the case. The learner must decide whether the original explanation still applies and justify the answer without copying the earlier wording. This tests transfer: whether the learner can use the distinction in a new situation rather than merely repeat a persuasive paragraph.
A later revisit can test retention. Neither immediate agreement with feedback nor a rising practice score should be treated as proof of readiness for unsupervised clinical responsibility.
Educational AI design can matter, but evidence must stay in context
A PNAS study published in 2025 examined generative AI assistance in high-school mathematics. Access to an unrestricted assistant improved supported performance but was associated with worse subsequent unaided performance than the control condition; a learning-oriented tutor design substantially mitigated that problem.
This was not a study of medical trainees, clinical supervision or iatroX. It supports investigating how assistance is designed, rather than assuming that better answers during a session automatically mean better independent learning afterwards. It does not establish that a particular clinical Tutor prevents deskilling.
For medical education, the relevant research question is more specific: does the learning sequence improve later performance on unfamiliar cases, including recognition of uncertainty and appropriate escalation? That requires direct evaluation with suitable learners and assessments.
Make supervision expose the learner's reasoning
A supervisor reviewing only a final, AI-assisted answer may struggle to distinguish the trainee's understanding from the tool's contribution. A more informative discussion asks what the trainee initially thought, which proposition changed and what remains unresolved.
That need not become a lengthy oral examination after every task. A focused question such as "What would make you choose a different plan?" can reveal whether the learner understands the decision boundary. Another is "Which part of this answer have you not been able to verify?"
Supervisors should also examine correct answers reached for weak reasons. A trainee may select the right option by recognising a phrase but misunderstand why the alternatives are inappropriate. Conversely, a defensible explanation may reveal an ambiguity in the teaching material rather than a learner error.
The proposed aim is useful feedback, not catching someone out for using assistance. Training should make permitted use explicit while preserving opportunities to demonstrate understanding independently.
Where a question-linked Tutor can fit
According to iatroX's product description in October 2026, its Socratic Tutor opens on the attempted question, asks targeted follow-ups and explores the learner's misconception rather than simply supplying the answer. Question practice and a subsequent reasoning discussion can therefore form part of the attempt-and-review sequence described here.
This is an article published by iatroX and includes its own tools. The format is a product design choice, not evidence that every explanation is correct or that the service replaces a clinical supervisor. Use fictional educational cases, and inspect the underlying sources when an answer affects understanding of a clinical decision.
For clarity, the iatroX offer supplied for October 2026 includes genuinely free Ask-iatroX and free question access, without a trial expiry or verification gate. Ongoing Tutor use sits within the paid bundle with question banks, the study planner, simulations and CPD tools: £99 paid upfront for a year, equivalent to £8.25 a month billed annually, or £29 a month. Simulations and CPD are included, not separate add-ons. At those October 2026 prices, three monthly payments total £87 and four total £116, so the annual payment is cheaper from the fourth month.
The reason to consider such a bundle is a coherent learning process for the relevant professional goal, not the number of unrelated examinations available. A trainee needing bedside examination feedback still needs supervised practical teaching. Someone struggling with a specific reasoning distinction may benefit from targeted questions, source checking and another independent attempt before buying more general content.
Frequently asked questions
Can a junior doctor use AI when they do not already know the answer?
Yes, but uncertainty about the answer should determine the checks and supervision required. A plausible explanation is not sufficient verification for a consequential clinical decision.
Is asking another AI a substitute for asking a supervisor?
No. Another model can share the same missing information or misconception, whereas appropriate supervision can examine the patient context, the trainee's competence and the decision that must actually be made.
Does good performance with a Tutor prove independent competence?
No. Independent application to new cases, appropriate supervision and relevant assessments are needed to establish what the learner can do without the same assistance.