An AI clinical note can contain the right words but attribute them to the wrong person. Distinguishing speakers is only part of the task: the system must also preserve whom a statement concerns, whether it is a report or a hypothesis, and whether someone later corrected it. A fluent summary can conceal all three mistakes.
The speaker and the subject are different questions
Consider this fictional paediatric consultation. A parent says, "His sister was vomiting yesterday." The clinician asks whether the child being assessed has also vomited. The parent answers no. A generated note that records vomiting in the patient has not merely misspelt a word. It has moved information between people.
A different error would record the parent's account as a finding observed by the clinician. Another would preserve the initial statement but omit the subsequent correction. These are distinct failures, and each calls for a different check.
Speaker diarisation, the technical task of separating speech by speaker, should therefore not be treated as proof of correct clinical attribution. Even a correctly identified parent can be talking about a sibling, quoting a teacher or describing something they did not witness. The review question is not simply "Who said this?" but "What relationship does this statement have to the patient's care?"
Why the distinction matters in the record
The GMC's guidance on keeping records when protecting children and young people, checked on 10 October 2026, emphasises recording concerns, findings, decisions and relevant conversations. Its purpose is a usable account for care, not an undifferentiated transcript of everything audible in the room.
Applied to the fictional encounter, a useful note would distinguish the child's symptoms from household context. It would not convert a question into a positive finding or a parent's uncertainty into an established history.
The same reasoning extends beyond paediatrics. In a memory assessment, a patient's account and a relative's observations may differ without either being deliberately misleading. Combining them into one confident narrative could remove the very disagreement the next clinician needs to understand. In a consultation involving an interpreter, words spoken in the first person may convey the patient's account rather than the interpreter's own experience.
These are proposed failure scenarios, not reported incidents involving a named scribe.
Give the system a role map, not an assumption
A practical deployment should establish how participant roles are supplied and corrected. An introduction identifying the patient, clinician and accompanying person may help, but the organisation should test whether the product actually uses that information reliably. A spoken introduction is not itself a validated control.
The role map must also accommodate change. Someone may join by telephone halfway through, an interpreter may replace another interpreter, or the clinician may leave briefly while a nurse enters. A system designed around two stable speakers should not silently treat a more complicated encounter as equivalent.
Where the product cannot establish attribution, an explicit uncertainty is preferable to a confident guess. The clinician can then resolve the relevant point before signing the record. Requiring a name for every utterance is unnecessary; preserving the origin of consequential information is the goal.
This does not justify creating biometric voice profiles or collecting extra identifying information by default. The local assessment should consider the minimum information needed for the task, alongside the approved data-handling arrangements.
Review disagreements rather than smoothing them away
Imagine a second fictional encounter. An older patient reports taking their medicines regularly. Their accompanying relative says several packets appear untouched. The clinician has not yet established what explains the discrepancy.
A poor summary might choose one account and write either "adherent" or "non-adherent". A more faithful note would preserve both reports, the unresolved question and the plan to clarify it. The appropriate clinical conclusion cannot be obtained merely by deciding which speaker sounded more confident.
A reviewer should look particularly closely at statements concerning symptoms, medicine use, safeguarding, consent and agreed actions. That prioritisation does not excuse ignoring the rest of the note. It identifies places where altered attribution could materially change what happens next.
As checked on 10 October 2026, NHS England's guidance for ambient-scribe users requires meaningful checking each time the tool is used and correction before information enters the record. Multi-speaker complexity should influence how the service makes that review achievable.
Test attribution separately from transcription
A proposed evaluation should include paired encounters containing similar words but different relationships. In one, the patient describes their own previous operation; in another, they describe a parent's operation. In one, the clinician states a finding; in another, the clinician asks whether the finding occurred.
The reference assessment should identify the speaker, subject, information source and any correction. Reviewers can then distinguish a speech-recognition error from an attribution error or an unsupported summary. A single overall accuracy figure would obscure these differences.
Include interruptions, overlapping speech, remote participants and interpreters where those occur in the intended service. The test should also examine the final approved note, because a product error caught during review has a different consequence from an error that remains after sign-off.
This is an evaluation design, not a claim that iatroX has performed a comparative scribe test. Results would need to come from an actual, appropriately governed assessment of the specified product and version.
Make correction reach the relevant destinations
When incorrect attribution has already entered a record, editing the latest visible sentence may be insufficient. The organisation should identify whether the information also appeared in a referral, problem list or patient message, and use its authorised correction process.
The correction should explain the distinction that matters: for example, that a symptom concerned a relative rather than the patient. It should not add an unsupported assertion about who caused the error. The investigation may need to distinguish the original speech, the generated draft and later human edits.
Teams can practise this skill without exposing real consultations. A fictional transcript containing competing accounts can become a discussion about which facts are established, which remain uncertain and what should be documented. As described by iatroX in October 2026, its CPD tools support reviewed learning records; they do not certify the accuracy of an AI-generated medical note.
The useful standard is faithful attribution, not simply tidy prose. A record should help the next reader understand who knew what, not require them to reconstruct it from a confident but ambiguous sentence.
Frequently asked questions
Is speaker identification enough to prevent attribution errors?
No. A correctly identified speaker may be discussing another person, quoting someone else or offering an uncertain interpretation.
Should an AI note resolve disagreement between a patient and a relative?
Not by silently choosing an account. Relevant disagreement should remain visible until the clinician has enough information to interpret it.
How can a team test this without using identifiable recordings?
Use fictional multi-speaker encounters with predefined attribution and correction points. Assess both the generated draft and the record produced after professional review.
Turn documentation review into a focused learning activity →