Clinical reasoning is not fact recall, and it is not generating a differential list on request. It is the process between: gathering selectively, weighting findings, committing to a working diagnosis, and knowing what would change your mind. An AI can hand you a flawless differential in two seconds, which is precisely why it can accidentally deskill you. The tutors worth paying for in 2026 are the ones that make you do the reasoning and then examine it.
The four tests of a reasoning tutor
Before comparing products, fix the criteria. Does it ask for your reasoning before revealing its own? Does it challenge your assumptions rather than politely agreeing? Does it reveal information progressively, the way patients do, rather than presenting the whole vignette at once? And does it revisit your errors later, since reasoning habits change through repetition, not epiphany?
How the platforms score
iatroX: the Socratic Tutor engages on missed exam questions, asking what you chose, why, and what finding argued against it, then teaching from the misconception it finds and feeding the topic back into adaptive review. Best-in-class for reasoning from a written vignette you got wrong, which is the moment reasoning is most visible and most correctable.
Geeky Medics: virtual patients reveal information the way real consultations do, only if you ask, and the AI examiner scores the encounter against a mark scheme with written feedback. Best-in-class for consultation reasoning: the selective gathering and communication half of the skill.
Lecturio: a genuinely question-led tutor inside its Q-bank that probes thinking and routes to instruction. Strong on the probe, lighter on longitudinal follow-through.
Neural Consult: its clinical case simulator plus GLIA make a capable reasoning gym across your own materials, in text or voice; the caveat is that cases generated from your notes are unvalidated against any external standard.
UWorld UAsk: superb at explaining reasoning after the fact, particularly why a distractor tempted you, but it is an explainer by design and rarely demands your reasoning first.
ChatGPT Study Mode: will run a decent Socratic case if you prompt it well, and forgets you the moment the tab closes; no error memory means no habit correction.
Written vignettes versus spoken consultations
The two halves of reasoning are trained by different machines. Written-vignette reasoning, the kind the AKT, MRCP and USMLE test, is best trained where the questions are calibrated and the errors are tracked: that is the iatroX and Lecturio territory. Consultation reasoning, the kind OSCEs and real clinics test, needs progressive disclosure and a patient who answers only what you ask: that is Geeky Medics territory, with Neural Consult's simulator a flexible second. A candidate facing both should train both, and should not assume skill transfers automatically between them.
Verdicts
Best for reasoning from a missed exam question: iatroX, because the Socratic engagement lands exactly where your reasoning just failed, and the platform remembers. Best for consultation performance: Geeky Medics. Best explainer of reasoning you have already attempted: UWorld UAsk. Best free improvisation partner: ChatGPT, fenced appropriately.
One habit multiplies whichever you pick: say your answer and your reason out loud, or type it, before any AI speaks. The tutor can only examine reasoning that exists.
A five-minute daily reasoning drill
Reasoning improves through reps with feedback, and five structured minutes daily outperforms an occasional ambitious session. The drill: take one vignette, from your bank's adaptive queue or today's clinic, suitably anonymised. Before any option or AI is consulted, write three lines: your leading diagnosis, the single finding that most supports it, and the single finding that most argues against it. Commit. Then let the machine respond, and grade yourself on the middle line and the last line, not the first, because naming the diagnosis is recall while weighting the evidence is reasoning.
If you were right, one further rep: what one change to the stem would have flipped your answer? That question, which the iatroX Tutor and Geeky Medics' examiner feedback both push toward in their own registers, is the fastest known probe of whether you understood the case or recognised it.
If you were wrong, resist the urge to read the explanation immediately. State your reasoning aloud first, then read, then classify the failure: knowledge gap, cue missed, or weighting error. The classification matters because the fixes differ, revision for the first, slower reading for the second, more contrastive practice for the third, and because a tutor that tracks you, as iatroX's does through its adaptive queue, will only correct the pattern it gets to see.
Five minutes, one case, three written lines, one classification. Run it daily for a month alongside whatever platform you chose above, and the platform will have measurably more reasoning to examine.