Text-based clinical AI cannot establish an examination finding that has not been observed and supplied. It can discuss possibilities, identify missing information and help structure questions. It should not turn an incomplete conversation into the equivalent of a completed examination merely because the requested answer expects a diagnosis or plan.
A complete vignette hides the work of obtaining the information
An examination question often supplies the relevant history, observations and examination findings. The learner's task is to interpret them. In a real consultation, deciding which information to obtain and whether it is reliable is part of the clinical work.
A model given a polished vignette therefore faces a different task from a clinician meeting someone whose presentation is incomplete or difficult to describe. Comparing their final answers without acknowledging the information difference can obscure an important source of difficulty.
The distinction does not make vignette evaluation worthless. It defines what the evaluation has tested: performance with the supplied information, not necessarily the ability to obtain that information safely from an unselected patient.
A fictional swollen-joint conversation
Consider a fictional patient who describes a painful, swollen joint through a text interface. The description does not establish the physical findings, general condition or complete clinical context. The system responds with a familiar explanation and a confident statement that no concerning examination features are present.
The key error is not the name of the suggested condition. It is the invented transition from "not provided" to "assessed and absent". The conversation cannot support that transition.
A useful teaching exercise asks what information could materially alter the response and how it should be obtained. Some questions may be answerable through a further history. Others may require examination, observations, investigation or assessment in a different setting. The exercise is not a treatment protocol and cannot settle this fictional case from the sparse details supplied.
The same problem can occur when a clinician uses AI. A concise prompt may omit a finding the clinician knows, or may include a shorthand description that the system interprets too strongly. The output must be read in relation to what was actually supplied.
Patient descriptions, photographs and examination findings differ
A patient's description is clinically important, but it is not automatically equivalent to a clinician's observation. A photograph may add useful information while leaving other aspects unavailable. A text-only tool may receive only someone else's description of the photograph.
The proposed review question is therefore not simply "was an image attached?" It is "which features can this input establish, and which remain unassessed?" Image quality, framing and the context of capture can also be relevant to that judgement.
A system should preserve attribution: the patient reports, a clinician observed, a measurement recorded, or the image appears to show. Those phrases identify the status of evidence. They should not be removed solely to make the final answer sound more decisive.
A longer multimodal conversation does not remove every need for physical assessment. The sufficiency of the consultation depends on the clinical task, not the number of messages or available media formats.
Remote care is a clinical choice, not a universal limitation
The GMC's remote-consultation guidance, checked on 10 October 2026, treats the suitability of the consultation method as part of professional judgement. Remote care can be appropriate, but the clinician needs to recognise when the available information or method is insufficient.
The implication for AI is neither "all text consultations are unsafe" nor "a sufficiently detailed chat replaces examination". Both claims are too broad. The relevant question is whether this encounter provides enough reliable information for the decision being proposed.
An answer may appropriately remain conditional. It can explain what follows if a finding is confirmed and what changes if it is not. That is different from assuming the finding and then adding a generic instruction to seek help if worried.
For the fictional joint presentation, a responsible educational response would identify the unresolved assessment rather than use reassuring language to disguise it.
Make the missing-information question specific
A useful framework separates the information into three groups: what is established, what is uncertain and what would change the next action. This is an editorial method for reviewing an answer, not a validated diagnostic instrument.
The first group should be traceable to the input. The second should not be silently filled with typical findings. The third helps prioritise what needs resolving instead of producing an exhaustive history template that nobody can complete.
For each consequential missing item, ask how it could be obtained and who should obtain it. A request for further information is not helpful if it asks the patient to perform an unreliable manoeuvre or requires access they do not have. The escalation route must fit the actual service.
This approach also avoids the opposite failure: demanding an examination for every general educational question. A clinician asking how a sign changes a differential is not necessarily asking the AI to determine whether the sign is present in a patient.
Test whether the system notices the missing examination
A proposed evaluation should include paired fictional cases with the same symptom description but different levels of available assessment. In one, a finding is explicitly documented. In another, it is unassessed. In a third, the description conflicts with a later observation.
The system should preserve those distinctions and change its level of certainty appropriately. Reviewers should assess unsupported findings, inappropriate reassurance, excessive escalation and the usefulness of requests for further information.
The DECIDE-AI reporting guideline, published in 2022, addresses early live clinical evaluation, including the human and workflow context. It is a reporting framework, not certification that an AI product can manage an encounter without examination.
A strong evaluation would examine the clinician-plus-tool process as well as the generated response. Does the interface help the user recognise what remains unknown, or make the incomplete case look finished?
Keep reference support separate from patient assessment
As described in iatroX's October 2026 product information, Ask-iatroX is a clinical-reference service with linked sources. A general question about which findings matter can be appropriate for reference checking. Establishing those findings in a patient remains a different task.
For learning, use fictional cases and ask what additional information would change the answer before revealing the examination. Then compare the revised interpretation with the initial one. This can practise the distinction between knowing what to look for and having actually found it.
The essential safeguard is not a ritual disclaimer beneath a confident conclusion. It is an answer that stays within the evidence available and makes the consequential gaps visible before anyone acts.
Frequently asked questions
Can text-based AI establish that an examination finding is absent?
Not when the finding has not been assessed and supplied. Missing information should remain missing rather than become a documented negative.
Does every remote consultation require an in-person examination?
No. The appropriate method depends on the task and available information, with a route to further assessment when needed.
What should a clinician check in an AI answer based on an incomplete case?
Check which findings were supplied, which were inferred and which missing details could change the next action. A plausible diagnosis does not resolve an unperformed assessment.
Explore which clinical information would change the answer →