A clinical AI answer can contain correct medical knowledge and still be wrong for the encounter. The failure may concern whose information was used, when it was true or whether a problem remains active. These contextual errors require checks of identity, chronology and status, not simply a more authoritative medical reference.
Knowledge accuracy and encounter accuracy are different
An explanation of a condition can be correct in the abstract while relying on an old result or another person's history. A citation may support the medical proposition without supporting the patient-specific premise that made the proposition relevant.
The medical-summarisation framework published in npj Digital Medicine on 13 May 2025 distinguishes error types and potential clinical consequences. Its experiments concerned consultation documentation and had commercially affiliated authorship, including TORTUS. It supports examining meaning and consequences, not assuming a universal error rate across clinical AI. See the original publication record.
For practical review, separate three questions. Is the underlying medical information sound? Is the input an accurate account of this patient? Does the output preserve the time and status of that information? Passing one check does not answer the others.
A historical result can quietly become a current fact
Consider a fictional record containing a previous abnormal investigation, a subsequent reassuring result and a new episode whose assessment is incomplete. An AI-generated summary highlights the earlier abnormality but omits the later change. The wording is based on a real entry, yet the summary could misrepresent the current position.
The opposite error is also possible. A later reassuring result might be used to imply that a new episode has already been assessed. Neither problem requires invented medicine. Both arise from attaching information to the wrong time.
A proposed review should distinguish the date of the event, the date it was entered and the date the source was updated. "Recent" and "previous" are less useful than a clear relationship to the current encounter when timing changes the interpretation.
This does not mean every historical detail needs repeating. It means that compression should preserve changes that could alter the decision being made now.
One patient's record can contain information about several people
Clinical conversations may mention relatives, carers and family history. A genuine statement about a relative should not become an established condition in the patient's own problem list.
An attached document may also be associated with the wrong encounter or contain another person's details. The safe response is not to ask the model to infer which person was intended. The mismatch needs resolving through the approved workflow before consequential use.
A fictional example could involve a referral summary that attributes a sibling's diagnosis to the patient. The source sentence is real, and the medical description is accurate. The error is the relationship between the information and its subject.
For training, ask reviewers to identify the subject of each important statement. A heading labelled "patient history" should not be accepted as proof that every sentence beneath it concerns the patient rather than someone discussed during the consultation.
Resolved, suspected and active are not interchangeable
A past diagnosis, a working possibility, a patient-reported concern and a confirmed active problem serve different purposes. Turning them into one undifferentiated list can create an apparently coherent but misleading clinical picture.
The GMC's record-keeping standards, checked on 10 October 2026, require clear and accurate records. In an AI-assisted workflow, a proposed application is to preserve the status of consequential information rather than merely retain the associated clinical term.
The reviewer can ask: was this reported, observed, inferred or established? Is it still relevant? Was an investigation requested, completed or only discussed? Has a later entry changed the meaning?
These questions also help distinguish missing data from negative findings. An empty field cannot safely be treated as proof that the feature was absent. Where the source is incomplete, the output should expose that limitation rather than silently make the record more decisive.
Dates and units need explicit handling
A numerical value without its unit, reference context or timing may be unusable even when its digits were copied perfectly. Similarly, a date entered in an ambiguous format can be interpreted differently across systems.
This is an engineering and workflow problem as well as a language problem. A proposed implementation should preserve structured units where available, make date interpretation explicit and prevent unsupported conversion between different measurement concepts. A fluent explanation should not be allowed to validate an ambiguous input.
For medicines-related decisions, the GMC's prescribing guidance, checked on 10 October 2026, emphasises sufficient knowledge of the patient's health. An apparently relevant result from an uncertain source is not equivalent to verified current information.
No numerical treatment instruction is needed to test this failure mode. Synthetic records can deliberately vary dates, units and result status while leaving the general clinical topic unchanged.
Put the context check at the point of use
A correct patient banner at login is not enough if the user later switches records, opens an attachment or copies information into another window. The relevant confirmation is the one connecting the current input and output to the action about to be taken.
A proposed review view would show the patient and encounter, the source of important facts, their dates and unresolved conflicts. It would also distinguish an AI-generated draft from the approved record. This is a suggested design, not a claim that every existing product exposes those fields.
The HL7 FHIR Provenance specification, inspected on 10 October 2026, describes recording the activities and agents involved in producing or changing a resource. That can support traceability, but an authentic history of how a record was created does not prove that its clinical interpretation is correct.
Test changes in context, not only familiar questions
A useful evaluation set would contain paired fictional cases. The medical topic stays the same while the patient, chronology or problem status changes. The expected output should change where the altered context matters and remain stable where it does not.
Reviewers should examine whether the system detects contradictions, asks for missing information and avoids treating historical material as current. They should also test whether the clinician can recognise the mismatch in the actual interface, rather than score only a clean text export.
The principle applies to iatroX as well. Its October 2026 reference positioning can support checking general clinical knowledge, but a source-linked answer cannot establish that a local record, attachment or result belongs to the right patient and encounter.
The practical safeguard is to keep the medical proposition and the patient-specific premise visible as separate things. A correct answer to the wrong clinical context is still the wrong answer to use.
Frequently asked questions
Can an AI answer be wrong even when its medical references are correct?
Yes. Correct sources do not establish that the patient details, chronology or record status used in the answer are accurate.
Is a historical normal result the same as a current assessment?
No. Its relevance depends on the present question and what has changed, so timing should remain explicit.
Does provenance prove a record is clinically correct?
No. It can help establish origins and transformations, but the meaning and applicability of the information still require assessment.
Separate the general clinical question from the patient-specific context →