Document the clinical decision and its material basis: what information mattered, what was checked, what remained uncertain and why the plan was chosen. Where AI made a consequential contribution, record that proportionately and in line with local policy. A transcript of the chatbot exchange is not a substitute for an intelligible clinical account.
This article is dated 10 October 2026. It discusses UK professional standards and proposed documentation practice, not a universal legal requirement to record every AI interaction. GMC guidance on clinical records requires clear, accurate, contemporaneous records with proportionate detail; the task is to preserve useful reasoning, not maximise note length.
Start with what the next clinician needs to understand
A fictional GP considers a patient's persistent symptom, checks a source-linked reference answer and decides that the patient's circumstances require assessment outside the routine pathway. The useful record explains the relevant facts, the reason for that decision and the follow-up arrangements.
"Discussed with AI" says little about the decision. A pasted answer may say too much about general medicine while omitting what made this patient different. The record should make the link between evidence and action visible without requiring the next clinician to reconstruct a conversation with software.
Ask a simple question: if the AI exchange were unavailable tomorrow, could another professional understand the important clinical judgement from the record? If not, the note may be relying on an external explanation rather than documenting the care delivered.
Distinguish background reference use from material influence
Not every search has the same significance. Looking up the wording of a general recommendation differs from changing a management plan because a system highlighted a contraindication or proposed an alternative interpretation.
A proposed proportionality test is whether the information meaningfully changed the assessment, action, communication or follow-up. If it did, record the relevant reasoning and source verification. If it did not, avoid turning routine reference activity into a lengthy technical diary unless the local process requires a particular entry.
This is an editorial framework, not a new GMC rule. Organisations may specify additional documentation for particular systems, regulated functions or research settings. Follow those requirements while ensuring that the patient record remains understandable and clinically useful.
Write what was verified, not merely that verification happened
"AI checked" is ambiguous. It could mean that a second model agreed, that the clinician glanced at a citation or that the original source was read and applied to the patient's circumstances.
A more useful account identifies the proposition and the relevant source when that detail matters. It might say that current local guidance was checked, that the patient fell outside a standard pathway and that an alternative assessment was arranged. The entry should not imply that a source supports a claim it does not make.
The same scrutiny applies to iatroX. As described in its methodology checked on 10 October 2026, source retrieval and grounding are design features intended to support checking. A citation is a route to evidence, not a legal shield or proof that the recommendation fits the patient.
Three fictional examples of proportionate documentation
The following examples are educational and contain no real patient information. They illustrate structure rather than wording that should be copied regardless of the encounter.
A reference-supported decision might read: "Relevant UK guidance reviewed after a clinical-reference search. The usual pathway does not address the combination of findings discussed today. Explained the uncertainty and agreed further clinical assessment, with follow-up responsibility recorded below." The actual record would state the clinically necessary findings and plan; the general example deliberately avoids presenting a treatment instruction.
A disagreement might read: "The AI-generated summary described the previous problem as confirmed. The source correspondence records it as a possibility only. Corrected the summary and retained diagnostic uncertainty in the assessment." This records the meaningful discrepancy rather than simply saying that the clinician disagreed with AI.
An information gap might read: "The suggested plan depended on a result not available during this consultation. Did not treat the missing result as normal. Arranged retrieval and review, with the responsible team and interim plan documented." Here, the important contribution is recognition of missing evidence, not acceptance of a generated recommendation.
Separate what happened from what should have happened
A generated note must not document advice, consent or examination findings merely because the model thinks they would be appropriate. If a draft adds a useful suggestion that was not discussed, it remains a proposed action, not a record of a completed conversation.
The NHS ambient-scribing guidance for professionals, checked on 10 October 2026, is a relevant starting point for reviewing generated documentation before it is relied upon. The clinician's task is to check meaning as well as wording.
In a fictional consultation, the model adds a clear follow-up instruction that the clinician had not actually given. The appropriate response is not simply to sign the improved note. The clinician must decide whether the advice is appropriate and, if needed, communicate it through the correct process before documenting it as delivered. A record should not retroactively invent the consultation.
Record disagreement and uncertainty without defensive inflation
A good note can be concise while making an important decision traceable. It should identify uncertainty that affects the plan, not append every remote possibility the model generated. Excess detail can make the clinically important point harder to find.
When rejecting an AI suggestion, explain the clinical reason where material: missing information, an inapplicable population, a conflicting source or a different patient priority. There is usually little value in a generic declaration that "clinical judgement prevailed" without saying what judgement was exercised.
Similarly, avoid documenting an absolute assurance that the output was correct. Record the checks actually performed and the decision reached. A later reader should be able to distinguish knowledge available at the time from information that emerged afterwards.
Keep technical audit information in the appropriate place
Product versions, event logs and detailed retrieval records may matter during an investigation. That does not mean every technical identifier belongs in the visible consultation narrative.
A proposed organisational design separates the clinical record from the technical audit trail while maintaining an authorised link where necessary. The patient record explains care; the technical record helps reconstruct system behaviour. Retention, access and incident preservation should be arranged through governance processes rather than personal exports.
ICO data-minimisation guidance is relevant to that design: collect the information needed for the specified purpose, rather than treating unlimited logging as inherently safer. Clinical usefulness and investigatory value should be identified explicitly, not assumed to justify every copy.
Use reflection to improve the process, not rewrite history
A clinician may identify a learning need after noticing an unsupported recommendation or an important exception. That can become a separate reflective activity using fictional or appropriately generalised material.
The reflection should describe what the clinician actually learned, not invent a more sophisticated decision process than occurred at the time. It should not be used to backfill the original record. Any clinical correction should follow the appropriate amendment process and preserve the distinction between the original entry and later information.
The best documentation is neither silent about material AI influence nor dominated by it. It gives a clear account of the patient, evidence, judgement and plan, while preserving an appropriate route to examine the technology when that becomes necessary.
Frequently asked questions
Must every AI search be mentioned in the patient record?
This article does not establish such a universal requirement. Follow local rules and record material clinical influence proportionately, with the reasoning and final decision made clear.
Should the complete chatbot transcript be copied into the notes?
Not as a substitute for a clinical account. Relevant technical evidence may need a separate approved audit process, while the patient record should explain the care and reasoning.
Does recording that a citation was checked protect against liability?
It is evidence of a step taken, not immunity. The source must support the relevant claim and be applied appropriately to the patient and setting.
Explore source-linked evidence for a general clinical question →