An original audit method for separating a real source, a supported claim and a convincing explanation when evaluating evidence-linked clinical AI.
EvidenceMD makes citations and visible clinical reasoning central to its public positioning. Those features can help a clinician inspect an answer, provided the inspection reaches the original evidence. A visible explanation and an authentic reference are useful objects to examine; neither independently establishes that a clinical conclusion is correct.
The practical question is therefore how to check an EvidenceMD answer, and how to apply the same standard to iatroX. This article is published by iatroX and reviews public documentation checked on 26 October 2026. We did not run a hands-on comparison. The sample claims and audit below are fictional teaching material, not quotations or observed failures from either service.
Separate three questions before assessing the answer
First, does the source exist? A title, journal, date and identifier should lead to the intended publication. Second, does the source support the associated claim? A real paper can be attached to a statement about an outcome or population it did not study. Third, does that supported claim justify the conclusion in the stated setting? Even a faithful study summary may not resolve the clinical question.
These checks require different evidence. Finding a paper in an index answers an identity question. Reading its methods and results answers a support question. Comparing its population, setting and outcomes with the current question addresses applicability. Treating the first check as if it completed all three creates false reassurance.
The PubMed user guide describes finding a specific citation by its title or bibliographic details and following full-text links. That is a useful recovery route when an AI citation is incomplete or opens the wrong page. A database record can verify identity, but you may still need the full article to evaluate a detailed claim.
What visible reasoning adds, and what it cannot prove
EvidenceMD describes a visible reasoning process linked to medical sources. This may give the reader more intermediate assumptions to challenge. For example, an explanation that explicitly assumes a particular population makes that assumption easier to inspect than a conclusion offered without context.
The explanation still needs checking. Turpin and colleagues' 2023 research demonstrated that explanations from the models they tested could omit influences on their answers and rationalise biased conclusions. That study did not evaluate EvidenceMD or iatroX. Its relevance is conceptual: readable reasoning should not automatically be treated as a faithful account of an AI system's internal process.
iatroX's methodology describes retrieval, grounding and output checks. Those descriptions likewise do not remove the need to inspect a particular answer. A fair evaluation asks what the reader can verify and whether the conclusion follows, rather than equating a named architecture or a long explanation with correctness.
A worked fictional audit: when the claim grows beyond the evidence
Consider an invented answer containing this sentence: "The intervention reduces hospital admissions across all age groups." Suppose the cited fictional study actually compares a symptom score in adults attending a specialist clinic. It reports an improvement in that score, but does not measure hospital admissions and excludes children.
The citation may be real within the exercise, yet two unsupported extensions have occurred. The outcome has changed from a symptom score to admissions, and the population has expanded from selected adults to everyone. A fluent explanation of the intervention's mechanism would not repair either gap. The issue is the evidence-to-claim relationship.
An audit entry should identify each gap separately. Record the claim, the precise source passage and the reason it fails to support the statement. Then ask the assistant to narrow or withdraw the claim. A useful correction would describe the measured outcome and population, and acknowledge that the broader question remains unanswered by that source.
| Audit item | Fictional source finding | Required change |
|---|---|---|
| Population | Adults in one specialist setting | Remove the all-age generalisation |
| Outcome | Symptom score | Do not substitute hospital admissions |
| Comparison | A specified comparator | Retain that comparison in the summary |
| Conclusion | Limited to the study question | State what remains unresolved |
This exercise illustrates a check, not an assertion that either product produces these mistakes. It is useful because it tests what source verification must accomplish even when the citation itself is genuine.
Read the assumptions as carefully as the conclusion
Before following every reference, inspect the inputs the answer relies on. Has it assumed a country, age range, disease stage, treatment exposure or care setting? Were those details supplied, inferred or left unspecified? An answer can accurately quote evidence and still answer the wrong question because its starting assumptions are wrong.
Use a short annotation system. Mark supplied facts as "given", missing details as "unknown" and explicit modelling choices as "assumed". If the explanation treats an unknown as given, ask for a revision. If several reasonable assumptions lead to different source requirements, request those branches separately rather than accepting one seamless answer.
This is particularly helpful when an AI combines several publications. Ask which claim comes from which source and where synthesis begins. A conclusion supported by the assistant's interpretation should be labelled as such. It should not appear to be a direct finding merely because citations are nearby.
Build a source-checking record you can reuse
For a teaching group or local product evaluation, keep the original synthetic question, date, product configuration if visible and the generated answer. Select a few decision-relevant claims before opening the sources. This makes the exercise more informative than collecting only the examples that happen to look good or bad.
Record source identity, claim support, applicability and unresolved uncertainty separately. Avoid a single pass/fail column that hides different failure types. A broken link may be an access problem; a fabricated citation is an identity problem; a real study attached to an unrelated conclusion is a support problem. They call for different corrections and affect different parts of the workflow.
Also record the effort required. If you must reconstruct every citation and rewrite the main conclusion, the answer has not completed much verified work for you. If it directs you efficiently to the right passages and preserves their limitations, the benefit is clearer. These are practical observations for the task tested, not evidence of general clinical safety or a population-wide accuracy ranking.
Make your shortlist follow the verification task
EvidenceMD may interest readers who value a visible explanation and are evaluating its wider suite. iatroX may interest readers who want referenced clinical questions connected to UK-focused learning. Their public feature descriptions should guide what you test, while the resulting evidence should guide what you trust.
For a broader selection framework, read EvidenceMD review: a practical evaluation. For guideline questions, use the NICE guidance comparison workflow. In each setting, the useful endpoint is a defensible relationship between the question, the source and the conclusion.
FAQs
Are citations proof that an AI answer is correct?
No. They make verification possible. You still need to confirm source identity, whether the source supports the associated claim and whether the finding applies to the question being answered.
Does visible reasoning make EvidenceMD more accurate than iatroX?
This article does not establish comparative accuracy. Visible assumptions may help readers inspect an answer, but explanation length and transparency claims are not substitutes for appropriate comparative evidence.
What should I do if a cited article is inaccessible?
Check its bibliographic identity through a reliable index and seek legitimate full-text access. If the necessary passage cannot be checked, record that limitation rather than treating the claim as verified.
Put the audit into practice: ask a synthetic evidence question in askiatroX, select one important claim and follow it back to the source before accepting the explanation.