Daily Clinical Briefing · Educational reading for healthcare professionals

What gets lost when AI writes the notes?

November 13, 2026

Accurate documentation can still bypass the thinking that helps clinicians notice inconsistencies.

PEARL OF THE DAY

Before entering patient details into an AI tool, check how it handles personal information; clear, useful output provides no assurance about that.

Summary

Writing a clinical note helps organise the patient's story. Choosing what to include and trying to make the account coherent can reveal inconsistencies that need another question. Automating the record may remove some of that work as well as the typing.

Artificial intelligence (AI) can produce consultation notes, referral letters and discharge summaries, creating opportunities for more direct engagement with patients. An accurate record is useful. But checking that record for errors is a different task from working through the clinical story while writing it, and accuracy alone does not show that the clinician has retained that thinking or can recall the patient.

The Acute Medicine Podcast's discussion of AI makes this a practical concern for teams considering automated documentation. Clinicians still need to organise the account and decide whether the findings fit together. Time saved in producing the record tells us little about whether those activities have been preserved.

Similar problems arise when AI supplies an interpretation. Pattern matching and probabilistic associations can produce a plausible answer, sometimes one that is largely correct. Near-correct answers may be harder to challenge than obvious mistakes because identifying the important error requires clinical expertise.

A final human check offers less reassurance than it first appears. Automation bias can lead a clinician to favour the generated answer over conflicting clinical information. Fatigue and workload can further limit the scrutiny that a busy registrar gives the result, even while remaining responsible for reviewing it.

Clinical expertise therefore needs to remain involved throughout the work, rather than being confined to approving a finished output. In teaching, ask learners to explain how they weigh the findings before comparing their reasoning with an AI answer. Agreement with the generated conclusion is insufficient evidence that they have worked through the problem.

The hyponatraemia example makes that approach concrete. Clinical dehydration and high urea need consideration alongside an apparently normal serum osmolality. Asking which findings deserve greater weight creates a useful discussion; moving straight to the generated conclusion can bypass it.

There are useful communication tasks too. AI can help make patient letters easier to understand, provided clinicians consider how the system handles personal information. Useful explanations can be retained for reuse, and patients who bring AI-generated information deserve a discussion of its content rather than dismissal.

For a team introducing automated notes, the practical decision is how clinicians will continue to assemble the story and investigate what does not fit. Reviewing the wording of the completed note cannot, by itself, establish that this work has happened.

More from today’s podcasts

AI

Automated documentation and patient explanations offer useful roles for AI, but plausible errors, biased training data and automation bias require ongoing clinical judgement; preserving the thinking involved in note-taking and learning remains part of safe use.

Behind the Knife General Surgery Oral Board Review – Sample Episode 11: Inguinal-Femoral Hernia

A painful incarcerated groin hernia can threaten bowel viability despite normal vital signs; decompression before sedation and observation after reduction address immediate risks, while contamination, prior surgery and defect anatomy guide repair.

Ep 232 – Umbilical hernia

Umbilical hernia assessment depends on reducibility, fascial defect size and skin condition; small asymptomatic childhood defects often close spontaneously, whereas adult and para-umbilical hernias do not, and suspected strangulation requires urgent surgical care.

Diagnosis and Treatment of Acute Ischemic Stroke

Acute ischaemic stroke assessment combines clinical disability with imaging of vessel occlusion, infarct core and salvageable penumbra; reperfusion decisions require consideration of collateral flow and tissue viability rather than elapsed time alone.

Establishing Potosi Virus as a Human Pathogen - Part 2

Fever or new neurological symptoms during B-cell therapy warrant a detailed exposure history, including ill children, gatherings and arthropod contact; practical precautions should acknowledge treatment benefits and support a fulfilling life.

Establishing Potosi Virus as a Human Pathogen - Part 1

Anti-CD20 therapy can produce slowly progressive infection with muted cerebrospinal fluid and antibody responses; nearly normal results do not reliably exclude infection, and specialist input with consideration of next-generation testing may help identify the pathogen.

Facial Pain Series - Part 2

Brief electric pain in the throat or deep ear requires attention to distribution and triggers rather than automatic attribution to trigeminal neuralgia; associated syncope, overlapping ear innervation and secondary causes guide further assessment.

E230: Melanie Clarkson, Jancis Kinsman and Sam Worster - Accreditation and Routes into Advanced Practice

Advanced practice accreditation in therapeutic radiography depends on evidence of autonomy and breadth across four pillars; supported learning routes, usable study time and observed clinical decisions help demonstrate competence beyond a job title.

What to change on your next shift

When a system writes the notes, keep actively organising the clinical story and checking for inconsistencies. Ask learners to explain how they weigh competing findings before comparing their reasoning with an AI answer.

Questions from today’s episodes

A registrar working a busy overnight shift receives an AI interpretation that conflicts with the clinical assessment. The registrar accepts the generated answer despite those conflicting findings. Which risk does this illustrate, and does requiring a final clinician check reliably remove it?

This is automation bias: favouring automated information despite conflicting clinical evidence. A final human check does not reliably remove the risk because fatigue and workload can limit critical scrutiny; clinical judgement needs to remain involved throughout the task.

An adult presents with a painful groin mass, vomiting and abdominal distension. The hernia is successfully reduced, and the mass disappears. Does successful reduction establish bowel viability, and what follow-up assessment is needed?

No. Bowel viability remains uncertain after reduction. Arrange close observation for 12 to 24 hours with serial abdominal examinations and a surgical repair plan; worsening findings require exploration.

An adult receiving anti-CD20 therapy develops slowly progressive neurological disability. Cerebrospinal fluid is nearly normal, and serum and cerebrospinal fluid antibody tests are negative. Can these results reliably exclude infection?

No. B-cell depletion can mute both the cerebrospinal fluid response and antibody production, while infection may progress slowly. Seek infectious diseases and allergy/immunology input and consider next-generation testing to help identify the pathogen.

How these briefings are prepared

Daily Clinical Briefings are prepared using ChatGPT Pro from the show notes and educational output for that day’s episodes. Iain Beardsell then checks the briefing for accuracy.

This describes the current briefing process. Item-specific review dates are shown only where documented. The original podcasts remain the sources; MedPod’s briefing is additional educational material.

Read about MedPod’s editorial approach

Report an error or source concern — please include the article title and link, without patient-identifying information.

Continue learning with MedPod Learn

Browse and listen to medical podcasts free in the app.
Premium adds notes, MCQs, reflections and activity tracking, with exportable learning records.
Try the St Emlyn’s sample