The honest boundary this article draws is precise rather than vague: AI can genuinely help structure and interrogate reflection, and it cannot supply the underlying clinical experience or substitute for a doctor's own genuine consideration of what that experience meant, a distinction the RCGP's own guidance states directly rather than leaves implicit.
Six things AI can be asked to do, only some of them appropriate
Proofreading: correcting grammar and clarity in a reflection you have already written, uncontroversial and genuinely useful. Restructuring: reorganising your own content into a clearer format, similarly appropriate since the substance remains entirely yours. Reflection prompts: questions designed to draw out deeper consideration than a first draft captured, a genuinely valuable use this cluster's coverage treats as the correct role for AI in this specific task. Summarising the user's own notes: condensing genuine personal notes into a more usable form, appropriate provided the summary does not introduce content the original notes did not contain. Generating the entire reflection from a bare prompt: the point at which appropriate assistance tips into something the RCGP's guidance specifically warns against. And inventing learning or patient encounters that did not happen: never appropriate under any framing, a fabrication risk this article treats as a bright line rather than a matter of degree.
The confidentiality and patient-data risk
Any AI tool used in drafting a reflection carries the same patient-confidentiality risk this cluster's broader coverage of AI and clinical documentation treats throughout: genuine clinical detail entered into a third-party system, even for entirely legitimate reflective purposes, requires the same anonymisation discipline as any other use, names, dates of birth and identifiable combinations removed or generalised before anything touches an AI tool, regardless of how private or well-governed that specific tool claims to be.
The formulaic-reflection problem
AI-assisted reflections risk a specific, recognisable failure mode: fluent, well-structured prose that says remarkably little, hitting the expected shape of a reflective entry, what happened, what was learned, what will change, without genuine specificity underneath it. Experienced appraisers increasingly recognise this pattern, and a formulaic AI-generated reflection is, if anything, more likely to prompt a probing follow-up question than a rougher, more specific one clearly written by the doctor themselves. The goal of any AI assistance in this task should be the opposite of formulaic, using structure to draw out more genuine specificity, not to paper over its absence.
Comparing the reflection tools across this category
Umbil drafts structured reflections directly from captured clinical activity. BMJ Portfolio offers AI-supported prompts for drafting and refining entries. Clarity's Leo AI layer adds assistance within the platform. General-purpose tools like ChatGPT can be used for reflection support outside any dedicated platform entirely, carrying the same confidentiality discipline and RCGP guidance regardless of which specific tool is used. And iatroX takes a deliberately narrower position: assembling the evidence, the source consulted, the question attempted, the misconception the Tutor identified, and prompting reflection around that evidence, without generating the reflection's substantive content itself.
Acceptable and unacceptable prompts, worked
An acceptable prompt: "Here are my rough notes on a prescribing question I had this week. Help me structure this into a clearer reflection without changing what actually happened." An unacceptable prompt: "Write me a reflection about a time I learned something about anticoagulation management," with no actual notes or experience supplied, inviting fabrication by design.
What every reflection should require the clinician to add personally
What surprised them about the situation. What they previously believed, stated honestly even where it was wrong. What actually changed as a result. What action followed in practice, not merely intention. And what uncertainty remains, since acknowledging the limits of what was resolved is itself evidence of genuine reflective capacity, precisely the quality a fluent but generic AI-generated paragraph most reliably fails to demonstrate.
The RCGP's actual position
Current RCGP appraisal guidance recommends that doctors do not use AI simply to write their reflections and notes, and specifically notes that responsible officers may impose local rules governing this further. Its trainee-facing guidance states the same principle in different language: AI can prompt reflection, and it cannot reflect for the doctor or manufacture the underlying experience.
The iatroX implication
This platform's own CPD feature should never be marketed as one-click reflection completed, a framing that would misrepresent both what the tool does and what the RCGP's guidance permits. The honest description is evidence assembled, reflection prompted, clinician remains the author, precisely the boundary this whole article argues every platform in this category should hold to.
Frequently asked questions
Will an appraiser know if a reflection was AI-assisted?
Not always with certainty, though specificity, genuine uncertainty and personal voice are considerably harder for AI to fabricate convincingly than generic structure, meaning a formulaic entry is a more reliable tell than the mere fact of having used an AI tool for support.
Is declaring AI use required?
Follow your responsible officer's specific local rules on this directly, since the RCGP's own guidance notes that responsible officers may set requirements beyond the College's general position.
Does using AI for reflection prompts count as the doctor's own authentic reflection?
Yes, provided the underlying content, what happened, what was believed, what changed, comes from the doctor's genuine experience and consideration; AI-assisted structure around authentic content is different in kind from AI-generated content standing in for experience that was never actually reflected upon.
Evidence assembled, reflection prompted, you remain the author →