Not yet, and possibly not ever in the way the question implies. AI can generate questions, explain answers and tutor conversationally, but the things that make question banks effective, forced retrieval, calibrated difficulty, blueprint coverage, spaced scheduling and performance data, are structural properties that a chat window does not provide by default. The strongest evidence-based answer in 2026 is a hybrid: AI-enhanced question banks, not AI instead of question banks.
Why clinicians are asking
The question is reasonable. If a modern AI can answer any clinical question instantly, explain it patiently and even quiz you on request, paying for a bank of multiple-choice questions can feel like paying for a landline. Platforms are encouraging the thought: OpenEvidence has scored 100 percent on the USMLE and now offers CME, and every major clinical AI is adding learning features. So it is worth being precise about what each tool actually does.
What AI genuinely does well
Conversational AI is superb at explanation on demand: adjusting depth to the learner, rephrasing until a concept lands, connecting a question to pathophysiology, and generating unlimited practice items on any topic. It removes the friction that used to sit between confusion and clarification. For a learner who knows what they do not know, that is transformative.
What question banks do that chat does not
A good bank is not a pile of questions; it is an assessment instrument. Items are written to a blueprint so coverage matches the exam rather than the learner's curiosity. Difficulty and discrimination are known because thousands of candidates have answered each item, so your percentage means something. Distractors encode real misconceptions. And the format forces retrieval: you must commit to an answer before seeing the explanation, which is precisely the effortful act that builds retention. Ad hoc AI-generated questions have none of these properties reliably; they are unvalidated, uncalibrated and often subtly miskeyed.
The learning science that settles it
Decades of cognitive research converge on a few robust effects. Retrieval practice beats re-reading: in classic experiments by Roediger and Karpicke, students who tested themselves retained far more a week later than students who repeatedly studied the material, and Karpicke and Blunt later showed retrieval practice outperforming even elaborative concept mapping. Spacing beats massing: distributing practice across days multiplies retention for the same time invested. And difficulty is desirable: Robert Bjork's work shows that the conditions which make practice feel harder, generation, delay, interleaving, are the conditions that make learning last. Question banks operationalise all three effects; passive answer-reading operationalises none of them.
The failure modes of AI-generated questions
It is worth being specific about why on-demand generated questions cannot yet carry an exam campaign. First, keying: a small but meaningful fraction of generated items are subtly wrong, a defensible distractor keyed as correct, an outdated threshold, and the learner has no way to know which fraction they are holding. Second, distractor quality: strong exam items use wrong answers that encode real misconceptions, which takes editorial craft; generated distractors are often either transparently silly or accidentally defensible. Third, blueprint drift: ask an AI for questions on a topic and you get its sense of the topic, not the exam board's weighting of it, so effort quietly migrates away from what is actually tested. Fourth, no calibration: without response data from other candidates, a score on generated items tells you nothing about where you stand. Used as warm-up or explanation fodder these flaws are tolerable. Used as your primary assessment instrument, they corrupt the very feedback you are practising to obtain.
What the hybrid looks like
The synthesis is already visible in well-designed products. Use a validated, blueprint-mapped bank as the skeleton, so coverage and calibration are guaranteed. Let an adaptive engine schedule repetition, so spacing happens without willpower. Use AI where it is genuinely superior: as the explanation layer, and as a tutor on the questions you get wrong, asking you to reason before revealing, identifying the specific misconception, and rebuilding the concept from the source. That is retrieval practice with a personal teacher attached, and it is better than either component alone.
A pragmatic buying rule
For the clinician deciding where money and hours go, a simple allocation rule follows from all of this. Spend the money, where you spend it, on the validated instrument: the blueprint-mapped bank whose difficulty data and coverage you cannot generate yourself. Spend the hours on retrieval inside it. Use conversational AI, which is increasingly free or bundled, in the roles where validation does not matter: explaining what you got wrong, going deeper on mechanisms, generating warm-up questions on a topic before you face the calibrated ones. And exhaust free tiers before paying twice: several UK banks, iatroX included, keep core exams free, which lets you test whether a platform's retrieval-and-tutoring loop suits you before a subscription enters the question. The rule keeps each tool in the seat it earns.
Where iatroX sits
This hybrid is the design thesis of the iatroX Q-bank: thousands of MCQs grounded in UK guidelines and mapped to exam blueprints, an adaptive engine that combines spaced repetition with active recall to resurface your weakest material, and a Socratic Tutor that opens on missed questions to ask before it answers. The core UK exams are free, so the claim is testable at no cost. AI has not replaced the question bank; it has made the question bank considerably better.