AI literacy has quietly become a clinical competency, and it is not prompt engineering. The doctors qualifying this decade will practise their entire careers alongside AI systems, in education now, in decision support already, in workflows increasingly, and the literacy that protects them and their patients is conceptual: knowing how these systems produce outputs, where they fail, how their failures feel from inside, and when not to use them at all. This page, replacing our earlier general overview of AI in medical education, is the practical framework, organised around what the AAMC's principles for AI in medical education also emphasise: human-centred use, transparency, equity, training, privacy and continuous monitoring.
How generative systems actually work, at the level that matters
The load-bearing concept fits in three sentences. Generative models produce statistically plausible continuations of text, trained on vast corpora; they are engines of fluent likelihood, not databases of verified fact. Grounding, connecting outputs to retrieved, citable sources, is an architectural addition, which is why source-grounded clinical tools and open-ended chatbots behave differently on the same question. And confidence in the prose is a stylistic property, not an epistemic one, models write fluently about what they know and what they hallucinate alike, which single fact explains most AI failure modes a student will meet.
The three failure modes to know by feel
Hallucination: fluent fabrication, invented references, plausible-but-wrong mechanisms, most dangerous exactly where the reader lacks the knowledge to notice; the defence is structural, prefer grounded tools with inspectable citations, and check the load-bearing claim at source. Jurisdiction drift: correct medicine for the wrong country, US thresholds and drug names delivered confidently to UK students, the quiet failure that guideline-grounded UK tools exist to prevent. Automation bias, the human half: the documented tendency to over-weight machine output, which in novices combines with fluency to produce the measured paradox, misleading explanations that cut accuracy while raising confidence: /blog/novice-paradox-ai-confidence-medical-students. Literacy here is metacognitive, treating your own post-explanation confidence as a non-signal, and letting delayed unassisted performance be the verdict: /blog/illusion-of-learning-ai-fluency-vs-recall.
Confidentiality: the rule with no exceptions
The clearest line in the whole framework: patient-identifiable information does not go into consumer AI tools, ever, and near-identifiable information, rare-condition-plus-context combinations, deserves the same caution, because your confidentiality obligations exist independently of any vendor's terms. Educational use has a safe transformation habit: synthetic or fully altered cases, checked for indirect identifiers, with local information-governance guidance consulted whenever placement material, letters, results, images, is involved. Clinical versus educational use is a boundary worth keeping bright in your own head: tools appropriate for studying a topic are not thereby appropriate for a real patient's decisions, where intended-use statements, supervision and your own accountability govern.
Using AI well as a learner, and when not to use it
The evidence-backed method compresses to one rule, attempt first: commit to your own answer before any AI speaks, then use explanation, tutoring and spaced retesting around it, the full protocol at /blog/answer-first-ai-second-clinical-learning and the design argument at /blog/chatgpt-is-not-an-ai-tutor-educational-guardrails. The when-not list matters as much: not for scored assessment, integrity aside, it defeats the measurement; not as sole source for anything consequential, grounded or otherwise; not for patient-specific decisions during placements, which belong to supervision; and not as substitute for building the knowledge that makes verification possible, because the clinician who cannot reason without the tool cannot reason with it safely. Disclosure closes the framework: where AI contributed to work you submit or document, say so per your institution's rules, a habit whose professional version, transparency about tools in clinical reasoning, you will carry for a career.
Frequently asked questions
Should medical schools teach this formally?
Increasingly they are, and the AAMC-style framework above, mechanisms, failure modes, bias, privacy, monitoring, is a reasonable competency skeleton; until your school does, this page plus deliberate habit is a serviceable curriculum.
Which single habit gives the most protection per effort?
Source-checking the one load-bearing claim in any AI answer you intend to rely on; it takes seconds with grounded tools and converts fluency from a risk into a convenience.
Is it acceptable to use AI heavily and still become a good doctor?
Yes, conditionally: heavily and correctly, attempt-first, source-checked, retested unaided, builds capability; heavily and passively rents it, and the examinations, then the wards, audit the difference.
Do I need to understand the mathematics of these models?
No: the working concepts above, plausible continuation, grounding, stylistic confidence, carry the clinically relevant weight; deeper technical literacy is valuable for those building or evaluating systems, and optional for using them safely.
How should I handle AI use in group work and teaching others?
Model the disclosure and attempt-first habits explicitly: state what the tool contributed, show the source-check, and require the group's own answers before the reveal; literacy spreads faster by demonstration than by policy.