Doctors who use AI in clinical work should be able to demonstrate competence in that use. That is a stronger and more proportionate aim than requiring every doctor to purchase the same certificate. RCSI's growing course audience strengthens the case for accessible education, but registration numbers cannot establish that participants have become competent.
On 3 September 2026, RCSI announced its AI in Healthcare course's 20,000th registration and described it as the university's most popular single programme. The audience includes healthcare professionals, students and leaders. The figure should therefore be reported as registrations, not as 20,000 doctors, course completions or verified improvements in practice. RCSI's announcement
The policy position in this article is an editorial argument. It is not a report that the GMC has introduced a universal AI-certificate requirement.
What the RCSI programme offers
As published on 9 September 2026, the RCSI course is self-paced and free to study. It covers core concepts, ethics and governance, clinical applications and administrative uses across four modules. Its assessment includes quizzes, reflection and a final knowledge check. Current course information
After completing the course and passing the knowledge check, learners can choose an official certificate with 10 CPD points for EUR 50, according to the same dated course page. Free access and the optional certificate fee should be explained separately. The currency should not be converted to imply a direct comparison with another provider's sterling subscription.
For a clinician seeking a structured introduction, this is a relevant option. It also demonstrates why introductory AI education should not be reduced to a list of prompting tricks: the advertised curriculum includes governance and the contexts in which the technology is used.
The university's registration milestone is evidence of interest. Whether the learning changes decisions, checking behaviour or incident reporting is a different evaluation question.
What the GMC currently expects
The GMC's guidance on AI and innovative technologies, reviewed on 9 September 2026, applies existing professional responsibilities to new tools. Clinicians remain responsible for their decisions, should work within their competence and should keep up with developments affecting their work. The guidance anticipates education and training as AI enters practice, and addresses reporting concerns and adverse incidents. GMC guidance
That is not the same as requiring everyone to complete a named course. It creates a more practical question: what does this clinician need to understand to use this technology responsibly in this role?
A doctor checking a drafted letter needs different practical knowledge from a clinical safety officer evaluating autonomous triage. A researcher using an agent for statistical analysis faces another set of questions. A shared foundation is useful, but identical training would miss important differences.
Define AI literacy through decisions a clinician can make
A workable curriculum should assess whether someone can identify the task, recognise a meaningful limitation and take the appropriate next step.
For a clinical-reference tool, that might mean distinguishing a supported recommendation from an inference and checking whether a source applies to the patient group and setting. For documentation, it might mean recognising that a generated note has changed an uncertain statement into a confident one. For triage, it might mean understanding what happens when the automated route is outside its scope or cannot complete a handover.
These are proposed competencies, not claims that a short course can certify every use of AI. Their advantage is that they can be demonstrated. A learner can show the source check, explain the corrected note or describe the escalation route.
Knowing the definition of a language model is helpful background. It is not a substitute for knowing what to do when a generated answer is persuasive but inadequately supported.
An original teaching exercise: audit the claim before using it
A department could start with a short, fictional evidence briefing prepared for teaching. Give it an identifiable source and deliberately include one change in meaning. For example, a study of simulated tasks might be described in the briefing as evidence of improved patient outcomes.
Ask the learner to locate the relevant outcome in the source, explain the difference and rewrite the claim. Then ask what additional study would be needed to justify the stronger wording. This is a proposed teaching exercise, not a result from testing any particular AI system.
The September 2026 GPT-4o physician trial provides a real example of why the skill matters. Its published abstract distinguishes controlled performance from real practice and states that harms were not assessed. A reader who notices those boundaries is doing more useful work than simply labelling the paper positive or negative about AI. The physician trial published on 9 September 2026
The exercise can be repeated with a different kind of evidence: registrations versus active users, estimated hours versus observed time, or a device's registration versus regulatory approval. The learning goal is preserving the meaning of evidence as it passes into a decision.
Role-specific practice should follow the introduction
For a doctor using a scribe, an appropriate exercise could involve comparing a synthetic transcript with a draft note. The learner checks who said what, which symptoms were denied, which decisions were made and which actions remain pending. The expected output is a corrected note and an explanation of the important changes.
For a practice evaluating triage software, use a fictional request that falls outside the normal automated route. The learner identifies who receives it and how completion is checked. The expected output is an understandable handover, not a memorised list of supplier credentials.
For a clinician using research agents, ask them to explain how a small eligibility rule becomes executable code. The expected output is a testable rule and a clear point at which statistical expertise is required.
These exercises need not require access to identifiable patient data. The examples can be synthetic, with the purpose and expected checks agreed beforehand. They should be presented as local teaching, not as validation of a commercial product.
Make the requirement proportionate
A blanket annual certificate risks rewarding attendance rather than capability. A more proportionate policy would connect training to the actual tools and responsibilities involved.
Before using a new workflow, the clinician should understand its intended purpose, important limitations, local data rules, checking requirements and escalation process. After a material change, learning should address what changed. Following a relevant incident, the team should examine the failure and whether practice needs to change.
This is an editorial proposal for service design, not a newly enacted regulatory requirement. It also leaves room for clinicians who do not use a particular technology to avoid irrelevant product training.
Employers have a role too. Requiring knowledge without providing time, access to suitable training or an understandable local policy would be an incomplete implementation plan. Competence is easier to assess when the organisation has first explained what the job requires.
A CPD record should show learning, not just generated prose
A useful record identifies the original learning need, the activity completed, what changed in the learner's understanding and how that will affect their work. It should also preserve unresolved questions rather than presenting every session as a completed improvement.
The clinician should be able to explain the reflection without relying on an automatically generated paragraph. A polished record that describes learning that did not happen is not a satisfactory educational outcome.
For the purpose of this comparison, it matters that this article is published by iatroX and includes its own learning tools. The RCSI course and iatroX's ongoing learning workflow serve different needs; neither should be presented as a substitute for local product training or specialist governance expertise.
Where iatroX fits, and where it does not
In iatroX's September 2026 product information, the Socratic Tutor starts from an attempted question, asks targeted follow-ups and explores the learner's misconception. The study planner uses the examination date, available study time and quiz performance to shape revision. These are learning designs, not evidence that the platform provides a complete AI-literacy qualification. iatroX Tutor and learning methodology
iatroX's September 2026 product information also describes clinician-reviewed simulations and CPD tools alongside question banks. The CPD workflow allows learners to select their profession and context, review and personalise a draft record, and export evidence, including direct FourteenFish export for linked accounts. Completed records remain available after the subscription ends. A learning record is not a claim of universally accredited CME. iatroX CPD
The published UK subscription in September 2026 is £99 paid upfront for a year, equivalent to £8.25 a month billed annually, or £29 a month. It combines question banks, Socratic Tutor, study planner, iatroX Simulations and CPD tools; simulations and CPD are not separate add-ons. Ask-iatroX and the designated free question access remain genuinely free, without trial expiry or a verification gate. These are separate access arrangements from RCSI's course and optional certificate.
For a structured introduction to AI in healthcare, RCSI is a relevant starting point. For continued work on an appropriate clinical learning goal, iatroX's questions, tutoring, simulations and evidence records offer a different workflow. The value is several useful ways to learn one relevant subject, not access to unrelated examinations or a certificate for its own sake.
AI literacy should become part of clinical competence wherever AI is used. The educational test is whether a clinician can recognise a consequential problem and respond appropriately, not merely whether their name appears in a course-registration total.
Frequently asked questions
Has the GMC made an AI-literacy certificate mandatory for every doctor?
The GMC guidance reviewed on 9 September 2026 sets expectations around competence, professional judgement and keeping up to date, rather than a universal requirement to purchase a named certificate. The proposal here is for role-specific competence, not a report of a new blanket rule.
Is RCSI's AI in Healthcare course free?
As published on 9 September 2026, the course is free to study. After completion and a successful knowledge check, an optional certificate with 10 CPD points costs EUR 50.
Does an iatroX CPD record automatically provide accredited CME credit?
No: a personal learning record and formally accredited credit are different. Clinicians should check the requirements of their profession, jurisdiction and relevant scheme before claiming credit.
<!-- Editorial production note, 9 September 2026: Supplier services were not tested hands-on, and proposed scenarios are not observed product results. No independent cost audit or replication of the research studies was performed. Private regulatory dossiers were not reviewed, so the host product's classification and manufacturer certification claims were not independently determined. No existing MHRA/Airlock page was identified for an in-place update; Manchester sandbox news is incorporated into the governance article without replacing a live page or inventing its URL. The Nature PDF's figure rendering was unavailable; trial figures used here also appear in the article's text. The specific partnership announcement dated 9 September was located and used, resolving the brief's earlier uncertainty. -->