A clinical AI answer can contain no invented facts and still be inadequate for the task. It may describe a reasonable option while omitting the condition that makes that option inappropriate, or explain an investigation without saying who must act on its result. Checking only whether the visible sentences are true will not detect everything that is missing.
The useful question is not whether an answer contains every fact about a condition. It is whether it contains the information needed for the decision the reader is trying to make. Completeness is task-specific, and an exhaustive textbook entry can be less useful than a short answer that preserves the important exception.
Omission is different from fabrication
An invented examination finding is an addition. A real examination finding excluded from a summary is an omission. Both can matter, but they need different detection methods. To find an addition, compare the output with the available evidence. To find an omission, establish what relevant evidence or decision requirement should have appeared in the first place.
A paper published in npj Digital Medicine on 13 May 2025 offered a framework for evaluating hallucinations, omissions and their potential clinical significance in generated medical documentation. Its experiments used primary-care consultation transcripts from the PriMock dataset and clinician annotation. The work was commercially affiliated with TORTUS; it should not be treated as an independent assessment of every scribe or reference product. Its useful contribution here is the separation of error types and consequences, not a universal error rate. See the original study.
A reference answer is not identical to a consultation summary. Nevertheless, the methodological lesson is relevant: counting unsupported statements alone leaves an important part of answer quality unmeasured.
A fictional plan with a missing owner
Consider a fictional educational case. A clinician asks an assistant to help organise follow-up after an investigation has been requested. The output accurately summarises the reason for the test, explains that the result will influence the plan and suggests discussing the result with the patient.
It never identifies who will check whether the result arrives, who will review it, or what happens if the patient is not reached. None of its sentences is necessarily false. Yet the plan is incomplete as an operational handover.
Adding another paragraph about the underlying condition would not solve this omission. The missing information concerns ownership and completion. The clinician needs to establish the responsible person or team through the actual service, not ask the language model to invent a plausible name.
Now change the task. If the original question asked only what the investigation measures, the same absence of an operational owner might be entirely appropriate. This is why an evaluation must specify the intended use before deciding what counts as missing.
Four places to look for consequential gaps
The first is eligibility. Does the answer identify the population or circumstances to which its recommendation applies? A recommendation limited to a particular group should not silently become general advice.
The second is exceptions and competing factors. Does the answer preserve the conditions under which the usual approach changes? In a real clinical decision, relevant medicines information, patient factors and local guidance must be checked rather than assumed from a short response.
The third is uncertainty. Does the output distinguish missing information from an explicit negative finding? "Not described" is not the same as "excluded", and a confident summary should not erase that difference.
The fourth is what happens next. Does the task require monitoring, reassessment, escalation, communication or confirmation that another team has accepted responsibility? Where it does, the answer should not stop at a plausible immediate action.
These are proposed review categories, not a clinical protocol for a particular condition. They help the reader locate the question that needs authoritative checking or further assessment.
Short answers should expose their boundaries
The solution is not to make every response longer. More text can conceal the important qualification among routine material. A better short answer states its scope, gives the main point and makes the consequential limitations visible.
For example, an educational response might say that it is describing a general approach rather than resolving eligibility for an individual patient. It can identify the information that would change the answer and link to the relevant section of the source. This is more useful than appending a broad disclaimer after a detailed but overconfident plan.
The same principle applies to expandable interfaces. Important exclusions should not become invisible merely because they sit behind a collapsed section. Optional background can be layered; a condition essential to interpreting the main recommendation belongs close to that recommendation.
A reader should therefore assess both the words and their presentation. Was the caveat available only after several additional actions? Was the uncertainty framed as routine background even though it could change management? Those are questions for usability testing, not just factual scoring.
A completeness check that starts with the task
Before reading the AI answer, write the decision in one sentence. For a fictional training exercise, that might be: "Determine what additional information is needed before this plan can be accepted." This reduces the temptation to judge success by how polished the answer looks.
Then identify the few missing facts that could alter that decision. Check whether the answer addresses them, explicitly asks for them or acknowledges that they remain unknown. Do not let it fill them through inference disguised as documentation.
Next, inspect the relevant original source. Look beyond the cited heading to the surrounding qualifications. Finally, identify the next human or organisational action. Some missing information requires a conversation, an examination or access to an approved record, not another search.
This proposed process should remain proportionate. It is not an instruction to re-create a complete medical textbook during every consultation. It is a way to focus scrutiny on information that could change what happens to the patient.
How to test omission detection fairly
A useful evaluation would give reviewers a defined task, the full authorised input and agreed criteria for important content. Reviewers would distinguish an irrelevant omission from one capable of changing interpretation, action or follow-up.
Test cases should include apparently routine requests, conflicting information and negative findings that are easy to compress incorrectly. Where the input itself is incomplete, score whether the system exposes that limitation rather than pretending it knows the answer.
The comparator should be the actual alternative, including its omissions and review burden. A manually written note or conventional search can also be incomplete. The question is whether the assisted workflow improves the balance of useful information, error detection and effort under realistic conditions.
The DECIDE-AI reporting framework, published in 2022, is relevant when moving from technical testing into early live clinical evaluation. It is a reporting framework, not certification that a product is safe.
Apply the same scrutiny to iatroX
As described in its published methodology, checked on 10 October 2026, iatroX uses retrieval, source grounding and output checks in its clinical-reference workflow. These are design features, not proof that an answer includes every relevant exception.
A useful learning sequence is to form an initial view, inspect the answer's supporting passage and ask what missing fact would change the conclusion. Where uncertainty remains, follow the source or seek appropriate clinical support. The objective is better verification, not replacing one unexamined answer with another.
Frequently asked questions
Can an AI answer be unsafe without hallucinating?
Yes: an answer can omit information necessary for its intended use even when its included statements are accurate. The significance depends on the task and the consequence of the missing information.
Should every clinical AI answer include every possible exception?
No, but it should preserve the exceptions and uncertainties needed to interpret its main claim. Excessive detail is not a substitute for identifying what could materially change the decision.
What is the quickest useful completeness question?
Ask what missing information would change the interpretation, action or follow-up. Then verify that issue through the appropriate source, patient assessment or approved clinical system.