An AI scribe should not turn advice that would have been sensible into a statement that the clinician actually gave it. The same applies to consent: a plausible description of risks and options is not evidence that they were discussed. Documentation needs to distinguish an event from a suggestion about what ought to happen next.
That distinction matters clinically as well as evidentially. A future clinician may rely on the note when deciding what the patient understands. A patient may read a record of a conversation they do not recognise. Neither problem is solved by making the generated wording sound professional.
A record of care is not a model answer
A clinical note describes information, decisions and actions from the encounter. A teaching answer may instead describe an ideal approach. Confusing those purposes can produce a record that is medically plausible but historically inaccurate.
The GMC's recording-decisions guidance, checked on 10 October 2026, emphasises proportionate records of the information exchange, decisions and agreed actions. Completing a consent form is not a replacement for an individual dialogue.
An AI-generated paragraph cannot supply that missing dialogue retrospectively. Nor should a standard macro create the appearance that every listed risk was explained, every question answered or every preference explored when the encounter does not support those claims.
This article concerns record integrity and proposed workflow safeguards in UK practice. It is not legal advice about whether a particular record establishes negligence, valid consent or liability in an individual case.
A fictional example: suggested safety-netting becomes documented history
Imagine a fictional consultation in which the clinician explains an initial plan but the conversation ends before follow-up advice is completed. The scribe produces a fluent closing sentence stating that deterioration advice was discussed and that the patient understood when to seek further help.
That sentence might resemble a desirable model answer. It is not an accurate record of the conversation as described. The appropriate response is not to approve it because the advice would have been reasonable.
The clinician should assess what still needs to be communicated, provide appropriate advice through the approved pathway and record that later communication accurately. The record should not pretend the later action occurred during the original discussion.
This is a constructed example, not a documented incident involving a named product. Its purpose is to show the boundary between prompting better care and falsely documenting care as completed.
Separate three states in the interface
A safer proposed design distinguishes discussed, suggested and completed later. Those states should not be interchangeable labels applied after the text has already entered the record.
"Discussed" should be supported by the encounter or by the clinician's accurate account of what occurred. "Suggested" identifies material for consideration, not a claim about the past. "Completed later" requires a truthful record of the subsequent action and its timing.
For example, a product could place potential missing follow-up topics in a clearly separate review area rather than insert them into the historical note. The clinician would decide whether they are relevant and whether further communication is needed.
That is an implementation proposal, not evidence that such an interface is sufficient to prevent errors. It would need testing for whether users understand the distinction and whether suggested text can accidentally flow into the final record.
Consent wording needs particular precision
Statements such as "all risks explained" or "patient fully informed" can conceal rather than clarify what occurred. A proportionate record should reflect the material information exchanged, the person's questions and the actual decision, rather than reproduce an idealised checklist without context.
An AI may help organise a draft of that record, but the responsible clinician needs to check its factual basis. A patient agreeing to use an AI scribe is also a different matter from consenting to an examination, investigation or treatment.
Where a tool is used to prepare information for a future discussion, its output should remain identified as preparatory. A draft explanation of options should not automatically become evidence that the patient received, understood or agreed with it.
The useful design test is whether someone reading the final record can distinguish what was proposed, what was explained, what the patient chose and what remained unresolved.
The source check must follow meaning, not keyword presence
A transcript containing the word "risk" does not establish a meaningful risk discussion. A patient saying they understand one part of the plan does not establish understanding of every generated follow-up instruction.
Likewise, a clinician thinking aloud about a possible warning should not be recorded as having delivered the warning to the patient. Speaker attribution, conditional language and timing all affect the interpretation.
The May 2025 medical-summarisation study distinguished several types of error and their potential significance within a specific documentation experiment. Its relevance here is the need to evaluate meaning, not a claim that it measured the frequency of fabricated consent discussions in every deployed scribe.
A review process should therefore examine whether the note accurately represents the exchange, including omissions. The presence of reassuring clinical terminology is not itself supporting evidence.
What to do when the draft overstates the encounter
Before filing, remove or correct unsupported historical claims. Then assess whether the underlying clinical task remains unfinished. Deleting a fabricated sentence can repair the draft while leaving an actual communication need unresolved.
Where further discussion is necessary, use the appropriate clinical route and record what happens. Do not backdate the communication or disguise it as part of the earlier encounter. If the patient cannot be reached, record that accurately and follow the relevant local process.
If the inaccurate note has already been filed, use the approved correction or addendum process. Consider whether another clinician, service or patient communication has relied on it. An internal amendment may not resolve an already distributed letter.
The ICO's rectification guidance, checked on 10 October 2026, is relevant to distinguishing correction from erasure of a historical record. Specific disputes require the organisation's information-governance and professional processes.
Test the templates before normalising them
A proposed local test would use fictional encounters in which an expected discussion does not occur, occurs only partly or is explicitly deferred. The evaluation should examine whether the system preserves those states rather than completing the consultation as though it were an examination model answer.
Include cases where the patient declines an option, asks an unanswered question or agrees to consider a decision later. These are not incomplete interactions for an AI to tidy into agreement; they may be the clinically important outcome.
Also test default macros and approval behaviour. A template that automatically states that the clinician reviewed the output can create another unsupported assertion if review has not occurred. The same integrity rule applies to the documentation of checking as to the documentation of care.
NHS England's ambient-scribing guidance for information-governance professionals, checked on 10 October 2026, requires outputs to be checked before they are used or recorded. Local design should make that check meaningful and practicable.
Use learning tools to improve the next conversation, not rewrite the last one
As described on 10 October 2026, iatroX's Tutor supports reasoning-focused follow-up around attempted questions. Fictional case practice can help a learner identify what a good explanation or safety-netting discussion should contain.
That educational use is different from generating claims about a real consultation. A practice answer should help the clinician prepare for future care, not make an earlier record appear more complete than the encounter was.
The strongest standard is straightforward: useful AI suggestions remain suggestions until the relevant action occurs. Accurate documentation records that action, its timing and its limits without substituting an ideal narrative for reality.
Frequently asked questions
Can an AI scribe add safety-netting that was not discussed?
It can suggest material for clinical consideration, but should not present undelivered advice as a historical fact. The clinician must verify the final record and address any genuine unfinished communication.
Does generated consent wording prove that the patient consented?
No: wording in a draft does not establish that an appropriate individual discussion and decision occurred. The record must accurately reflect the actual exchange.
What should happen if unsupported advice has already entered the record?
Use the approved correction process and assess whether further patient communication or downstream correction is needed. Do not backdate an action or pretend the original record was accurate.
Practise explaining a clinical plan through guided questioning →