AI for residents and registrars: the workflows that buy back time in training

How to think about AI in postgraduate training

Residents, registrars and fellows have the opposite problem to medical students. The reading list is shorter, the unprotected hours are fewer, and the documentation load is heavier. AI's job in training is to compress the repetitive prose around clinical work so the limited supervised hours go to deliberate practice. Treat the model as a fast junior who drafts well, hallucinates occasionally, and must never sign anything.

The GMC's 2024 update to Good Medical Practice and the equivalent ACGME 2024 common program requirements both land in the same place: AI tools may support training, but the trainee remains the accountable clinician for every entry, every decision and every patient record. That accountability does not transfer to the model, the vendor, or your training programme director.

The five jobs worth automating in training

1. Logbook entries and case reflections

Draft eLogbook narratives, eportfolio reflections and ARCP CBDs from bullet-point notes. The Joint Committee on Surgical Training accepts AI-drafted logbook prose at ARCP from 2025 onward, provided the trainee personally verifies content and identifiers. ChatGPT or Claude will turn five bullet points into a 250-word reflection in under a minute. Read every line, fix any clinical drift, sign it yourself.

2. Teaching slide decks for juniors and medical students

Registrars teach. Gamma, Beautiful.ai and Canva's AI features turn a topic outline into a 12-slide teaching deck in about 90 seconds. They are excellent at structure and weak at medical imagery, so replace stock diagrams with figures from Radiopaedia, BMJ Best Practice or a sketch-first illustration tool. The deck stays yours; the AI absorbs the typesetting and layout.

3. Journal-club preparation

Upload the paper to NotebookLM. The free tier handles up to one million tokens of source material per notebook (Google, 2025), so RCTs, guidelines and Cochrane reviews all fit without trimming. Ask for population, intervention, primary outcome, secondary outcomes, key limitation and the strongest critique. Cross-check against the paper. You arrive at journal club with a pre-built skeleton and time for the actual argument.

4. Operative and procedural teaching diagrams

Drawing each step of a procedure is the fastest way to teach a junior how it goes. Generic text-to-image models still hallucinate surgical anatomy, so for teaching diagrams use a sketch-first tool such as Angiosome. Sketch the step on an iPad, render a clean labelled image in under a minute, build a teaching set for an entire procedure in roughly 20 minutes. The topology comes from you; the rendering comes from the model.

5. Exam revision (MRCP, MRCS, FRCA, FRCR, USMLE Step 3, ABIM)

Use ChatGPT or Claude for explanation and viva rehearsal, NotebookLM for grounded guideline review, and Anki with FSRS for retention. The stack is the same as medical school; the volume per topic is lower and the bar for clinical reasoning is higher. Pair AI explanations with a question bank (Passmedicine, eMRCS, UWorld) for actual exam calibration.

The honest time-cost vs time-saving table

Numbers below are median estimates from informal trainee surveys (UK Foundation, IMT and Core Surgical, 2024-2025). They assume you already have a clear prompt and verify the output before submission. Without verification the time saving is illusory and the clinical risk is real.

Hard limits that protect your registration

Postgraduate training operates under a tighter regulatory frame than medical school. The four limits below are not optional and are not soft. Crossing them risks an ARCP outcome 3 or 4, a GMC referral, or a coroner's letter.

A workable weekly stack for a trainee

1. Monday on-call write-up: bullet the cases on the train home, paste into ChatGPT with an eportfolio prompt, edit on arrival, sign.
2. Tuesday teaching session: outline 10 bullets, drop into Gamma, replace stock images with Radiopaedia or sketch-first diagrams, present.
3. Wednesday journal club: load paper into NotebookLM, generate PICO summary, draft three critique questions, cross-check against the paper.
4. Thursday operative teaching: sketch each step on an iPad, render with Angiosome, assemble into a one-page teaching aid for SHOs.
5. Friday revision: Anki review, then 20 minutes of viva rehearsal with Claude playing the examiner.

AI scribes, ambient documentation and what is actually live in the NHS

Ambient AI scribes (Abridge, Heidi Health, Nuance DAX Copilot, Tortus, Nabla) are appearing in NHS trusts and US health systems through 2025 and 2026. The 2024 NEJM Catalyst evaluation by Tierney and colleagues of an ambient scribe across 3,442 Permanente physicians reported a fall in documentation burden and a modest rise in patient-facing time, with high physician satisfaction. The evidence for nursing and junior doctor cohorts is thinner.

Practical rule: only use a scribe inside an approved trust or department pilot, with patient consent recorded in the notes, and only sign documentation you have read in full. NHS England's 2024 ambient AI guidance is explicit that the clinician remains the author of record regardless of who or what generated the draft.

What AI does not help with in training

• Procedural skill. No model replaces deliberate practice under supervision; the only way to learn a chest drain is to insert one with a senior watching.
• Clinical judgement under uncertainty. AI hedges in ways patients cannot. The model never has to call the family at midnight.
• Breaking bad news and ethical conversations. The model has read every textbook on SPIKES; it cannot have the conversation.
• Audit data collection. AI summarises columns you give it; it will not chart-review the notes for you.
• Calibration against real exam questions. AI explanations are good; AI question generation is currently below the standard of dedicated banks.