Pastest, one of the most established names in UK postgraduate revision, now advertises Tutor Mode: instant, on-demand AI support while you answer questions (last checked August 2026). iatroX's Socratic Tutor approaches the same moment, the learner facing a question, differently: it interrogates rather than assists. This comparison is deliberately about educational methodology, not model quality; our separate grounding and feedback audit of Pastest's AI is at /blog/pastest-ai-for-mrcp-part-1-a-grounding-feedback-and-hallucination-audit and this article does not repeat it. The question here is simpler and more consequential: when AI joins a question bank, what should it actually do?
Two philosophies
The traditional question bank workflow has been linear for decades: question, answer, explanation, next question, and both platforms are breaking that line, in different directions. Pastest's Tutor Mode attaches AI to the existing workflow: while answering, the learner can summon support, clarification, elaboration, help with the stem, keeping the bank's rhythm intact and lowering the cost of confusion in the moment. iatroX's Socratic Tutor restructures the workflow around the error: when the learner gets a question wrong, the AI does not explain; it asks, requiring the learner to expose their reasoning so the actual misconception can be found, corrected against cited guidance, and scheduled for retest. Assistance versus interrogation: both are defensible; they optimise for different things.
The hyponatraemia test
A worked contrast makes the difference concrete. A 68-year-old with hyponatraemia; the learner answers wrongly. The traditional bank gives the correct answer and a paragraph of explanation. An AI assistant, asked, will explain on demand: why this is SIADH, fluently and accurately. The Socratic system instead opens with: which feature makes hypovolaemic hyponatraemia less likely here? Then: what would you expect the urine osmolality to show, and why? Then: what alternative diagnosis, if true, would change your management? The learner who completes that dialogue has not received an explanation; they have rebuilt the reasoning, discovered which step of theirs failed, and produced the correction in their own words, which is the version that survives to exam day. The assistant model's answer is faster; the Socratic model's answer is theirs.
Where Pastest remains particularly strong
Objectivity requires saying this plainly: Pastest's core assets are formidable and AI does not change them. A very large, long-refined question bank; genuine past-paper orientation, which for MRCP especially remains a distinctive advantage; mature analytics; substantial learning content; a polished app; and decades of examination heritage that shows in question quality. For an MRCP candidate whose plan is volume through past-paper-flavoured questions with help on tap, Pastest with Tutor Mode is a strong, coherent product, and nothing in this comparison argues otherwise.
Where integrated AI learning goes further
The Socratic Tutor is one component of a loop, and the loop is the actual argument. Performance across the adaptive bank identifies weaknesses the learner would not have chosen to revise; the Tutor diagnoses and corrects the specific misconception; spaced repetition retests the correction at intervals; the Study Planner reprioritises tomorrow accordingly; and mocks calibrate the whole against the exam. AI attached to a bank improves the experience of each question; AI orchestrating a system changes what happens between questions, which is where most learning is won or lost. The category-level version of this argument, across Pastest, AMBOSS, Geeky Medics and others, is at /blog/how-ai-is-changing-the-medical-question-bank.
Choosing between them
The honest decision rule: if your constraint is question volume and past-paper fidelity with frictionless help, Pastest's model serves it; if your constraint is that errors keep recurring, that revision feels busy but uncalibrated, or that you need the system to decide what you should do next, the orchestrated model is built for exactly that. Plenty of candidates run both, Pastest for past-paper mileage, iatroX for the diagnostic loop, and lose nothing by it. The question worth asking of any 2026 revision product is no longer how many questions does it have. It is: what happens after you get one wrong?
Frequently asked questions
Is on-demand AI help bad for learning?
Not bad, just limited: assistance in the moment reduces confusion, and the evidence on feedback consistently favours approaches that make the learner produce and then correct. The risk with pure assistance is comfort masquerading as progress; the fix is pairing it with forced retrieval, whoever provides it.
Does the Socratic approach slow revision down?
Per question, yes, deliberately; per mark gained, the arithmetic usually inverts, because interrogated errors stop recurring. Speed through questions was never the objective; speed to retained competence is.
Can Tutor Mode be used Socratically by a disciplined learner?
Partly: you can withhold your own peeking, attempt an explanation before summoning help, and interrogate the AI's response with follow-ups, which recovers some of the production effect. What self-discipline cannot recreate is the system-side half, the misconception logged, the retest scheduled, the planner updated, because those live in the architecture, not the conversation.