Giving a clinician access to an answer changes what they can do with assistance. It does not, by itself, establish what they can explain, remember or apply later without that assistance. A global medical AI rollout can improve access to information while leaving the educational question unanswered: what capability is the clinician developing?
This distinction should shape the learning responsibilities that accompany wider deployment. The aim is not to make every clinical search into a lesson or to withhold useful answers during patient care. It is to design an optional, deliberate route from a question encountered in practice to durable professional learning.
The pathway below is iatroX's educational proposal. It is not a report that the Anthropic and OpenEvidence initiative, or iatroX itself, has demonstrated improved long-term clinical outcomes.
Immediate assistance and learning require different measures
A supported task asks whether the clinician and tool can reach an appropriately checked result together. An educational task asks whether the clinician's knowledge, reasoning or self-assessment changes beyond that moment.
These outcomes can diverge. A person can select the correct answer while looking at an explanation but struggle to reproduce the reasoning later. Conversely, a demanding practice session may feel slower while exposing a misconception that was invisible during reading. Neither observation should be assumed from satisfaction ratings alone.
A learning evaluation therefore needs to specify the capability being developed. Recognising a fact, explaining a mechanism, choosing between plausible options and identifying when guidance does not apply are related but different achievements.
Counting questions submitted does not answer those questions. Nor does a record of time spent with an application. Those measures describe activity. They become educational evidence only when connected to an appropriate assessment of what changed.
What the spaced repetition evidence actually supports
Maye and Hurley's systematic review and meta-analysis, first published on 28 January 2026, included 14 studies in the review and 13 in the meta-analysis. Across 21,415 learners, it reported an effect favouring spaced repetition on objective knowledge tests: a standardised mean difference of 0.78, with a 95% confidence interval from 0.56 to 0.99.
That standardised effect is not a 78% improvement or an examination pass rate. The interventions included flashcards and spaced questions, and the authors called for further work on optimal delivery and longer-term performance. The review supports a learning principle; it does not validate a particular AI platform, simulation product or professional-outcome claim.
For product design, the implication is to make later retrieval possible and relevant. For evaluation, it is to measure whether the chosen implementation helps the intended learners rather than borrow a published effect size from a different intervention.
Do not create a false contrast with general-purpose AI
Guided learning is not exclusive to specialist medical platforms. OpenAI's Study Mode announcement of 29 July 2025 describes an approach using questions, hints and structured explanation rather than simply presenting a finished answer.
The relevant comparison is therefore not "general AI answers, medical AI teaches". It is whether a particular learning experience elicits an attempt, identifies the learner's difficulty, uses appropriate sources, gives useful feedback and revisits the concept at the right time for the objective.
A specialist product may make that sequence easier by connecting it to an examination, an attempted question or a practice case. A general tool may also support valuable guided discussion. Neither should receive an educational-outcome claim merely because its interface looks Socratic.
This article is published by iatroX and includes iatroX in that comparison. Its educational proposition should be judged by the same distinction between intended design and demonstrated learning.
Start with a learning question, not a copied consultation
Consider a fictional trainee who realises after clinic that they cannot explain why a familiar recommendation changes when a key eligibility condition is absent. The learning need is a conditional distinction, not a transcript of the patient encounter.
The trainee can write a de-identified educational question: "Which feature makes this recommendation applicable, and what would change the reasoning if that feature were missing?" Patient-identifiable information is unnecessary for that purpose.
This reformulation matters because the clinical task and the educational task are different. During time-critical care, the priority is appropriate assessment, supervision and checked information. Later, the trainee can explore the misconception without delaying the patient pathway or turning a clinical decision into an unsupervised experiment.
A useful learning system should make the transition easy while keeping the purposes separate. It should not imply that completing a reflection retroactively validates a decision made during care.
A learning pathway built around the unresolved distinction
The first stage is an unaided attempt. Before opening an explanation, the learner states what they think matters and how confident they are. This creates something specific to examine. "I do not know" is also informative when it accurately identifies a gap.
The second stage is source comparison. The learner inspects the relevant guideline or other authoritative material and distinguishes the actual recommendation from the AI's paraphrase. They identify the population, conditions and exclusions, rather than collecting a citation without reading the part that matters.
The third stage is targeted explanation. A tutor can ask why the learner preferred one option, what evidence would contradict it and which detail changes the interpretation. The goal is to expose a misconception, not simply generate a longer explanation of the correct answer.
The fourth stage is a changed case. Keep the underlying concept but alter the feature that controls applicability. A learner who memorised the earlier answer may repeat it; a learner who understood the distinction should be able to explain why the reasoning changes.
The fifth stage is later retrieval. Return to the same concept without showing the earlier conversation. The learner should have to reconstruct the important distinction, not recognise the familiar wording of a previous answer.
The final stage is a personally reviewed learning record. It should capture the gap, what the learner checked, what changed in their understanding and what still needs work. A generated draft can support this process, but the clinician remains responsible for whether it represents their learning.
What should be measured?
A useful evaluation begins with the outcome rather than with whichever application metric is easiest to export. The following is a proposed assessment framework, not a set of observed iatroX results.
| Educational aim | Suitable assessment question | An insufficient substitute |
|---|---|---|
| Delayed retention | Can the learner explain the key concept later without assistance? | Number of explanations opened |
| Transfer | Can the learner apply the principle to a materially different case? | Repeating the original item correctly |
| Applicability judgement | Can the learner identify when a recommendation does not fit? | Recalling the recommendation's wording |
| Confidence calibration | Does confidence distinguish stronger from weaker answers? | Reporting that the application feels reassuring |
| Recognition of limits | Does the learner identify missing information or a need for help? | Producing an answer to every case |
| Useful reflection | Does the record accurately describe the learning and remaining gap? | Number of automatically generated records |
The design should also account for prior knowledge. A highly experienced clinician and an early trainee may need different tasks, even when both use the same reference feature. A comparison should avoid crediting an application for differences that were already present before the learning activity.
An evaluation that would support stronger claims
An educational study could compare a structured pathway with access to the same evidence and explanation without the additional practice sequence. That would help distinguish the contribution of the learning design from the benefit of simply obtaining the source.
The assessment should include an unaided baseline, a delayed outcome and new cases that test the same principle without repeating the original item. It should record use of other study resources and make the assessment process as independent of the product as feasible.
Not every question needs a numerical mark. Locally relevant educators can evaluate the quality of an explanation, recognition of a critical omission and appropriate acknowledgement of uncertainty. The scoring method should be specified before reviewing outcomes.
A study of knowledge or reasoning remains a study of knowledge or reasoning. Claims about changes in patient care would require a suitable clinical evaluation, with the additional safeguards and design that involves. No such outcome is inferred here from the spaced repetition review or from a product feature list.
What iatroX contributes to this proposed pathway
Per iatroX product information, September 2026, Socratic Tutor opens on an attempted question and asks targeted follow-ups intended to identify the learner's misconception. The question banks cover more than forty examinations and include adaptive sequencing and spaced repetition. The study planner incorporates the examination date, daily study time and quiz performance.
The same September 2026 specification describes eighteen examination-specific simulation tracks and 1,114 clinician-reviewed cases at launch. Voice and text practice, pause-and-coach and uninterrupted examination modes, full mock circuits, transcript-linked feedback, Tutor-led remediation and a prescribed next case are intended to connect practice with further learning. One complete simulation is free, with progress shared across web and native apps.
Those features do not predict an examination pass, replace bedside or procedural practice, or establish independent clinical validation. Clinician review of cases is also distinct from endorsement by an examining body.
For learning records, iatroX's September 2026 CPD specification includes profession and practice-context selection, a draft reviewed and personalised by the learner, PDF export and direct FourteenFish export for linked accounts. Completed evidence remains accessible after a subscription ends. A professional learning record is not automatically formally accredited CME.
Pay for a useful learning process, not unrelated examinations
As published by iatroX in September 2026, Ask-iatroX and free question access are genuinely free, without a trial expiry or verification gate. The combined learning subscription costs £99 upfront for a year, equivalent to £8.25 per month billed annually, or £29 per month. Question banks, Socratic Tutor, the planner, simulations and CPD tools are included together.
The decision should follow the learner's goal. Someone who only needs a checked reference answer may not need a paid learning service. Someone preparing for one relevant examination may value several methods that address the same weaknesses. Access to many unrelated examinations is not the central educational argument.
For a learner already using a general-purpose guided tutor effectively, a specialist subscription needs to justify the extra value through its organisation, feedback or practice sequence. For a learner who struggles to connect errors with later revision, that connected workflow may be the reason to consider it.
Wider access should include an educational invitation, not an educational claim
A global medical AI service can make it easier to encounter explanations. The stronger opportunity is to help clinicians turn selected questions into a repeatable learning process without distracting from patient care.
That invitation should be voluntary, proportionate and designed around the learner's context. It should encourage source checking, an unaided attempt and reflection on limits. It should also make space for supervisors, peers and local educators rather than treating the AI interaction as the entire educational environment.
Better training is not demonstrated by access alone. It is demonstrated when clinicians can do something worthwhile afterwards that they could not reliably do before, and when the evidence is clear about what that improvement does and does not mean.
Frequently asked questions
Does using a medical AI tool automatically improve clinical knowledge?
No: successful assisted use does not establish durable learning. Appropriate evidence would assess retention, transfer or another defined capability beyond the immediate interaction.
Does the 2026 spaced repetition review validate iatroX?
No: the review supports spaced repetition as a learning approach across its included interventions, not a particular iatroX product. Product-specific claims require product-specific evaluation.
Are iatroX simulations and CPD separate paid add-ons?
Per iatroX product information, September 2026, they are included with question banks, Socratic Tutor and the study planner in the £99 upfront annual or £29 monthly subscription. A CPD record remains distinct from automatically accredited CME.
Turn a question attempt into guided learning with iatroX Tutor →