Using AI for MRCPCH revision: a safe study workflow

Use generative AI for MRCPCH revision as a planning assistant, critical challenger and writing aid - not as your source of truth or a substitute for answering questions yourself. Set a precise study goal, attempt the task unaided, ask the tool to challenge rather than complete your reasoning, verify every factual claim against a current primary source, and finish with a closed-book test. AI output is not evidence of learning; what you can retrieve and apply without it is.

This guide covers non-clinical examination preparation. It contains no clinical teaching, diagnostic, treatment, prescribing, dosing, safeguarding or patient-management guidance.

The direct answer: use AI around your learning, not instead of it

Generative AI can reduce friction around planning, reflection and organisation. It can help turn an unclear goal into a study session, pose questions about your revision process, challenge an explanation you have already written, or convert your own verified notes into a retrieval schedule. It can also produce confident errors, invented references, bland encouragement and plausible explanations that you recognise without truly learning.

The safest rule is simple:

Use AI to improve the process around an MRCPCH task, but require current authoritative sources for factual content and unaided performance for evidence of learning.

No study shows that using a particular AI tool improves MRCPCH results. Research across health-professions education suggests that generative AI can support practice, feedback, resource creation and self-regulated learning, but the evidence is developing and concerns include over-reliance, academic integrity and reduced critical engagement. Your workflow therefore matters more than the brand of tool.

Key points

• Attempt the learning task before consulting AI, so that the tool responds to visible gaps rather than replacing the effort that reveals them.
• Treat every AI-generated factual claim, citation and examination detail as unverified until you have checked it against a current primary or authoritative source.
• End every AI-assisted session with a closed-book explanation, retrieval prompt or fresh single best answer question completed without AI.

What generative AI can and cannot establish

It can generate useful language quickly

A modern language model predicts plausible text from patterns in its training and current context. That makes it useful for drafting, reorganising, comparing and questioning. It can propose several ways to structure a week, turn a vague reflection into specific prompts, or make your own explanation shorter and clearer.

Plausible language is not the same as verified knowledge. A polished answer may contain a wrong date, an outdated examination rule, a reference that does not support the claim, or a citation that does not exist. Fluency can make those faults harder to notice.

It cannot prove that you know the material

Reading an answer and recognising its wording creates familiarity. The examination requires you to select a single best answer independently. If the tool supplies the explanation, identifies the distinction and decides what matters, you may complete the interaction without practising retrieval or decision-making.

Your evidence should come from a later task performed without assistance: reconstruct the explanation, answer a fresh question, state why alternatives are weaker, or retrieve the plan and execute it.

It cannot replace current official MRCPCH information

The RCPCH maintains current pages for examination structure, syllabi, regulations, dates and application guidance. These are time-sensitive. If an AI answer describes eligibility, timing, booking, permitted materials or examination conduct, open the relevant current RCPCH page and check the wording yourself.

The current theory structure page describes three computer-based single best answer examinations. The regulations treat malpractice seriously and prohibit unauthorised communication, concealed devices and sharing or obtaining recalled examination content. Do not use AI or any other external assistance during an examination, and follow the current candidate instructions for your sitting.

What the research says

Generative AI research is expanding faster than long-term educational evidence. Short studies, varied tools and different outcome measures limit strong conclusions.

Health-professions education reports several possible uses

A 2025 systematic review of 33 studies in health-professions education grouped reported uses into practising, inquiry, production and knowledge acquisition. It also identified concerns including effects on critical thinking, academic integrity and peer interaction. The review maps how tools are being used; it does not establish a universal best workflow or an MRCPCH pass-rate benefit.

A 2026 systematic review and meta-analysis of randomised trials in medical education reported favourable learning outcomes in some comparisons. The studies largely measured short-term educational outcomes across heterogeneous settings, so the result should not be read as proof that routine AI use improves postgraduate examination performance.

Structured use appears more defensible than passive reliance

A systematic review of generative AI and self-regulated learning identified affordances across goal-setting, resource analysis, monitoring and feedback. Another review of AI-supported self-regulated learning found a small evidence base with varied study designs. Together, they suggest that AI may help learners manage a cycle, but they do not show that handing the cycle to the tool is beneficial.

Recent reviews of AI and critical thinking describe a conditional pattern: guided activities that require critique, verification and reflection can support thinking, while passive answer acceptance can encourage cognitive offloading. This distinction fits an exam-preparation workflow. The useful question is not simply "Did I use AI?" but "Which cognitive work did I retain?"

Cognitive load depends on the learner and task

A systematic review of generative AI and cognitive load found that effects depend on task design, scaffolding and prior knowledge. A tool may remove unhelpful administrative load, but it can also add prompts, choices and verification work. If checking the output takes longer than doing a familiar task directly, AI has not made that session more efficient.

Use GUARD for an AI-assisted revision session

GUARD is a PaedVault editorial framework, not an RCPCH method or a validated educational score.

G - Give the session one precise goal

Define the outcome before opening the tool. Examples include:

• build a realistic seven-day timetable from fixed rota constraints;
• create five reflection questions about recurring question-reading errors;
• challenge the logic of an explanation you have already written; or
• turn your own verified error log into retrieval prompts without adding new facts.

Avoid the unbounded prompt "Teach me everything for MRCPCH". It transfers control of scope, accuracy and emphasis to a system that does not know your syllabus gaps.

State constraints explicitly. Tell the tool not to add clinical facts, examination recalls or unsupported policy claims. Ask it to mark uncertainty and separate suggestions from claims requiring a source.

U - Undertake the task unaided first

Write your answer, plan or explanation before asking for help. Even a two-minute attempt reveals what you can retrieve, where your reasoning stops and which assumptions deserve checking.

For question review, record:

1. what you chose;
2. why you chose it;
3. which clue or distinction you used;
4. why the strongest alternative seemed weaker; and
5. what type of error you think occurred.

Then use AI to interrogate that account. This protects the core principle of self-explanation: your reasoning becomes observable before feedback changes it.

A - Ask for challenge, not completion

Prompts that preserve your cognitive work are more useful than prompts that request a finished answer.

Try:

Here is my non-clinical revision plan. Identify assumptions, scheduling conflicts and missing feedback loops. Do not add medical content or claim that the plan is evidence-based.

Or:

Here is my explanation based on a source I have checked. Ask five questions that test whether I can distinguish the ideas. Do not supply the answers until I respond.

Or:

Categorise these anonymised error-log entries by process: knowledge gap, retrieval failure, misreading, option discrimination, timing or confidence. Preserve my wording and flag ambiguous cases rather than guessing.

The tool should create productive resistance. If it immediately supplies the final version, ask it to withhold conclusions, request your reasoning and present one challenge at a time.

R - Reference-check every claim

Verification means opening the source, finding the relevant passage, checking its date and confirming that it supports the exact claim. A list of links generated by AI is not verification.

Use a source hierarchy:

1. current RCPCH pages for MRCPCH format, policy, booking, eligibility and logistics;
2. original research, systematic reviews or consensus guidance for learning claims;
3. established explanatory resources for context; and
4. AI output only as a lead to investigate.

Check references individually. Confirm the title, authors, journal, year and identifier. Then read enough of the paper to distinguish what was studied from what the AI inferred. If you cannot locate or inspect a source, remove the claim or label it as an unverified suggestion.

For a sustainable source set, follow the guide to choosing MRCPCH revision resources. More generated material is not automatically better coverage.

D - Demonstrate learning without the tool

Close the AI window and perform a transfer test. Choose one:

• retrieve the explanation on a blank page;
• answer fresh questions under timed conditions;
• explain the distinction aloud without notes;
• reconstruct the study plan and its next actions; or
• return after a delay and complete the same task without prompts.

Score the unaided result, not the quality of the AI conversation. If performance is weak, shorten the target, revisit the verified source and schedule another retrieval attempt. The closed-book blurting guide offers a simple way to expose omissions.

Five useful non-clinical AI tasks

1. Stress-test a timetable

Give the tool your available sessions, fixed commitments and desired review intervals. Ask it to identify overload, missing buffer time and goals that lack an observable completion criterion. Check the final schedule yourself and keep ownership of what is realistic.

Do not let the tool infer readiness from hours alone. Use question performance, recall after delay and completed feedback actions. A weekly MRCPCH revision review can turn those signals into the next plan.

2. Generate metacognitive questions

Ask for questions about how you studied rather than new medical facts. Useful prompts examine why you changed an answer, whether confidence matched accuracy, what evidence altered a decision and which part of the process failed.

This pairs well with confidence calibration. AI may group your anonymised observations, but your recorded confidence and subsequent performance provide the evidence.

3. Improve the clarity of your own verified explanation

Paste a short explanation you wrote from a checked source and ask for unclear wording, unsupported leaps or undefined terms. Require the tool to preserve meaning and flag any proposed factual addition separately. Compare every revision with the source before accepting it.

4. Convert verified notes into retrieval prompts

Provide only material you have permission to use and have already checked. Ask the tool to create short prompts without answers, then inspect each prompt for ambiguity. Remove anything that introduces a new claim. Schedule the retained prompts using spaced repetition without overload.

5. Analyse revision-process patterns

AI can help cluster de-identified error-log entries or compare two weeks of study behaviour. It might notice that several "knowledge gaps" are actually rushed reading or that late sessions repeatedly go unfinished. Treat the grouping as a hypothesis and confirm it against the original entries.

Tasks to avoid

Do not use AI as the sole author of practice questions

AI can generate plausible stems, options and explanations, but plausibility does not establish factual accuracy, syllabus fit or a defensible single best answer. Poorly controlled generated questions may rehearse ambiguity or error.

If you experiment with generated prompts, restrict them to verified user-provided material, inspect every item and never treat them as equivalent to official sample questions or an editorially governed question bank. Use the official sample-paper guide to practise the intended format.

Do not ask it for personalised clinical guidance

This workflow is for non-clinical examination preparation. Do not enter patient details, request diagnostic or treatment decisions, or use a public AI system to process confidential information. Use approved clinical systems and local governance for clinical work.

Do not upload material you should not share

Before pasting anything, consider confidentiality, copyright, contractual restrictions and the tool's current data controls. Remove personal and organisational identifiers. Do not upload live examination content, recalled questions, paid question-bank material or documents you are not authorised to share.

Do not use it to predict your pass result

Encouraging language, a generated score or an AI judgement of readiness has no validated relationship with your MRCPCH outcome. Build a readiness decision from timed first-seen performance, syllabus coverage, error patterns, delayed recall and practical examination arrangements.

Do not use external assistance during the examination

MRCPCH regulations prohibit unauthorised communication and devices, and protect examination content. Follow the current candidate instructions. AI use during an examination would not be a revision strategy; it would be external assistance.

A prompt template that preserves independent thinking

Use this template for a non-clinical task:

Goal: [one revision-process outcome]. Context: [constraints and my unaided attempt]. Role: challenge my process, not supply medical content. Ask one question at a time. Do not invent MRCPCH rules, references or results. Mark every factual claim that requires verification. Do not use recalled examination content. Finish by asking me to state the result unaided and choose a delayed retest.

The template does not make output reliable. It makes the boundaries visible and reminds you where verification and unaided testing belong.

A ten-minute verification routine

When AI produces a useful-looking answer:

1. highlight every factual claim;
2. separate time-sensitive MRCPCH claims from general learning suggestions;
3. open the current official page or research source;
4. confirm that the source supports the exact wording;
5. delete invented, overstated or unnecessary claims;
6. record the source and access date where it matters; and
7. test the final learning outcome without AI.

If verification repeatedly takes longer than the original task, stop using AI for that task. Efficiency includes checking costs.

Protect your long-term independence

Set deliberate no-AI zones. Timed question blocks, closed-book recall, error diagnosis and readiness checks should regularly happen without assistance. Compare performance with and without the tool. If AI-supported work looks strong but independent performance does not improve, the workflow is producing polished activity rather than learning.

Keep a brief usage log for two weeks:

• task delegated to AI;
• reason for using it;
• verification time;
• unaided outcome;
• delayed outcome; and
• decision to keep, change or stop the use case.

This turns convenience into an experiment. Retain only uses that save time or improve a measurable learning process without transferring essential reasoning.

Conclusion

Generative AI can support MRCPCH preparation when it improves planning, questioning, organisation or reflection while leaving factual authority and cognitive work with the candidate. Use GUARD: give the session a precise goal, undertake an unaided attempt, ask for challenge, reference-check claims and demonstrate learning without the tool.

The boundary is the point. Current RCPCH sources govern examination information, verified educational evidence informs learning choices, and unaided retrieval shows what you know. If AI makes the output look better but removes the thinking you need to practise, it is working against the examination goal.

Frequently asked questions

Can I use ChatGPT or another AI tool for MRCPCH revision?

Yes, for non-clinical support such as planning, reflection, challenging your own explanation or creating prompts from material you have already verified and are permitted to use. Do not treat the tool as an authoritative source. Attempt the task first, check factual claims against current primary sources and finish with an unaided test.

Can AI generate reliable MRCPCH practice questions?

It can generate plausible questions, but that does not establish accuracy, syllabus fit or a defensible single best answer. Do not treat unreviewed AI questions as equivalent to official sample papers or an editorially governed question bank. If you use generated prompts at all, restrict them to verified source material and inspect every item before practice.

How do I verify an AI-generated explanation?

Identify each factual claim, open the current primary or authoritative source, find the relevant passage and confirm that it supports the exact wording. Check every citation exists and matches the claim. Remove anything you cannot verify, then explain the corrected point without AI and test it later.

Will using AI weaken my independent reasoning?

It can if you routinely accept completed answers before making your own attempt. Reviews suggest that effects depend on task design: structured critique and verification may support thinking, while passive reliance can encourage cognitive offloading. Protect independence with unaided attempts, no-AI question blocks and delayed retrieval tests.

Can I use AI during an MRCPCH examination?

No. Follow the current RCPCH regulations and candidate instructions. They prohibit unauthorised communication and devices and protect examination content. AI or any other external assistance must not be used during the examination, and recalled examination material must not be obtained, shared or entered into a tool.

Sources

1. Royal College of Paediatrics and Child Health. Theory exams: structure and syllabi. Updated 27 August 2026.
2. Royal College of Paediatrics and Child Health. Regulations and rules for MRCPCH and DCH examinations. Current regulations and appendices.
3. Moorhouse N, et al. Generative artificial intelligence in health professions education: a systematic review. 2025.
4. Effects of generative artificial intelligence on learning outcomes in medical education: a systematic review and meta-analysis of randomised controlled trials. 2026.
5. Generative artificial intelligence and self-regulated learning: a systematic review. Computers & Education. 2025.
6. AI-empowered self-regulated learning: a systematic review. npj Science of Learning. 2025.
7. Generative AI and critical thinking in education: a systematic review. 2026.
8. Generative artificial intelligence and cognitive load in education: a systematic review. 2026.