A clinical summary becomes misleading when it makes the encounter sound more certain than it was. "Possible diagnosis", "patient reports a diagnosis" and "diagnosis confirmed" describe different states of knowledge. The same applies to a test considered, requested or completed. Improving the prose must not quietly change the status of the information.
Uncertainty is part of the clinical content
Consider a fictional consultation about intermittent palpitations. The clinician discusses several explanations and arranges further assessment. The patient remembers being told years earlier that the symptoms might be related to stress, but no supporting record is available.
A summary stating "anxiety-related palpitations, investigations normal" would introduce two unjustified conclusions. The explanation remains uncertain, and the planned investigation has not happened. Neither problem is solved by making the sentence grammatically elegant.
A faithful account might instead state that the patient reports a previous possible explanation, that the current cause remains unresolved, and that further assessment has been arranged. The precise wording depends on what actually occurred, but the distinction between history, interpretation and action should survive compression.
The clinical risk in this constructed example is not a particular diagnosis. It is that a later reader could stop looking for information because the record appears to have settled the question already.
Certainty has more than one dimension
A useful review separates diagnostic certainty from evidential origin. A patient can confidently report that a specialist made a diagnosis while the reviewing clinician has not yet seen the specialist's letter. That is different from saying the diagnosis is unlikely, and different again from independently confirming it.
Time is another dimension. A previously confirmed condition may no longer be active. An earlier negative investigation may not answer a new question arising from changed symptoms. An accurate historical statement can become misleading when its date disappears.
Technical standards already recognise some of these distinctions. The published FHIR R4 condition-verification terminology, checked on 10 October 2026, separates unconfirmed, provisional, differential, confirmed, refuted and entered-in-error states. These are information-model categories, not evidence that a particular AI product implements them correctly.
The practical lesson is to avoid forcing every sentence into a binary choice between true and false. Sometimes the correct record is that a relevant fact has not yet been established.
Conditional plans need their conditions
Now change the fictional encounter. The clinician says that a referral may be appropriate if the next assessment identifies a particular concern. A summary reading "referred to specialist" converts a contingency into a completed action.
Even "specialist referral planned" may be too strong if the decision remains conditional. The missing condition tells the next professional why the referral has not happened and what information is needed before deciding.
The same problem can affect documentation of consent. Discussion of an option is not consent to proceed, and a patient wanting time to consider an option is not necessarily refusing it. The GMC's guidance on recording decisions, checked on 10 October 2026, emphasises an accurate account of the information exchange that led to a decision.
A proposed review should therefore ask what was discussed, what was agreed, what remains conditional and what has actually occurred. These questions are more useful than searching only for obvious factual inventions.
Preserve corrections and changes of mind
A consultation can move from one working interpretation to another. The clinician may initially consider a benign explanation, then change the plan after hearing an additional symptom. Summarisation should not preserve the opening hypothesis while dropping the information that displaced it.
Equally, a note should not present every discarded possibility as an active concern. The purpose is not to retain a transcript-length differential. It is to represent the final state of the encounter accurately while retaining earlier reasoning where it explains a consequential decision.
One proposed interface could distinguish patient-reported history, clinician findings, working interpretation and agreed next steps. Another could highlight newly generated certainty words for review. Neither design should be advertised as protective until users have tested whether it improves meaningful error detection.
The preferable design is the one that helps the clinician understand the change, not simply the one with more coloured labels or more text.
Audit promotions and demotions of certainty
A useful local audit would identify whether a statement becomes more definite or less definite during generation. Both directions can matter. Turning a suspected condition into a confirmed diagnosis may encourage inappropriate action, while weakening a documented urgent concern into a vague possibility may reduce the response it receives.
A research anchor is the medical-summarisation framework published in npj Digital Medicine on 13 May 2025. It separates types of output error and considers their clinical significance. The authors were affiliated with TORTUS; the study is not an independent validation of every scribe or a measured rate of uncertainty loss in routine NHS care.
An original evaluation could use fictional encounters with explicit changes in diagnosis status, completed actions and pending results. Adjudicators would agree what the source supports before examining generated outputs. The assessment should include whether clinicians identify errors during realistic review, not only whether experts find them afterwards.
No such local audit results are reported here. The proposal is a way to ask a better question than whether the summary generally sounds plausible.
Teach the distinction outside the live consultation
A practical teaching exercise is to give learners a polished summary and ask them to mark which statements are observed, reported, inferred or unknown. They then inspect the underlying fictional dialogue and explain what they would change.
The exercise should reward justified restraint as well as correct conclusions. Writing "not established" can demonstrate better understanding than supplying a confident answer to a question the case cannot resolve.
As described by iatroX in October 2026, Socratic Tutor works from an attempted question and explores the learner's reasoning. It offers a relevant educational format for examining mistaken assumptions, not a validated system for detecting uncertainty errors in clinical records.
A better summary need not be longer. It should preserve the qualifications that change the meaning, remove unnecessary repetition and leave the reader clear about what still needs to be established. Clinical uncertainty is not untidy wording to be edited away; it is information that care may depend on.
Frequently asked questions
Is a more decisive clinical note necessarily a better note?
No. A note is better when its certainty matches the evidence and the decision actually reached during the encounter.
What is the difference between a patient-reported diagnosis and a confirmed diagnosis?
The first identifies the source of the information. Whether the diagnosis has been independently established is a separate question that should remain clear where relevant.
Should AI summaries keep every possibility mentioned during a consultation?
Not necessarily. They should accurately represent the final assessment and preserve earlier uncertainty or changes of reasoning when these are important to the plan.
Practise identifying the assumptions behind a clinical answer →