Begin with the claim, not the bibliography
An answer can contain an impressive list of journals while leaving its most important conclusion unsupported. Start by selecting one consequential sentence. Is the answer identifying a disease, suggesting a mechanism, describing a treatment benefit or recommending a next investigation? Those are different claims and may need different kinds of evidence.
Write the sentence in one column and its supporting source in another. This simple step turns a polished narrative into something you can examine. It also prevents a paper that supports one small part of the explanation from appearing to validate the whole answer.
Check three links in the chain
First, confirm the source. Open the DOI, PubMed record or publisher page. Compare the title, authors, year and journal with the reference. A real-looking citation is not a guarantee that a work exists. Walters and Wilder documented fabricated references and errors in an audit of older ChatGPT models. Their 2023 study is a historical demonstration of the problem, not an error-rate estimate for today's tools or for MontMedical. [1]
Second, check the finding. Read the relevant abstract, results or full-text passage. Find what was measured and what the authors actually concluded. An experiment showing a change in a cellular signal does not by itself establish that treating that signal will improve a person's illness.
Third, check the application. Compare the study population, condition, test method and outcome with the case being discussed. Evidence from a different disease can supply a useful hypothesis. The explanation should identify that transfer rather than present it as a direct treatment result in the current condition.
Use a small evidence table
| Sentence in the answer | What to inspect | What a good explanation should say |
|---|---|---|
| “This result establishes the diagnosis.” | What the test measured and which diagnoses it distinguishes | Whether the result confirms a disease, supports a possibility or leaves an important alternative |
| “This mechanism explains the symptoms.” | Whether that mechanism was measured in this person | Which parts are observed and which are inferred |
| “This treatment works.” | The disease studied, comparison group and meaningful outcomes | Whether benefit was demonstrated in a relevant population |
| “A normal result excludes it.” | The assay, stimulus, specimen and testing conditions | What that particular negative result actually addresses |
The table is a reading aid, not a scoring system. Its purpose is to expose the connection between evidence and a decision.
Learn to recognise three common leaps
The first is a real source attached to a stronger claim. A paper might report a possible association while the answer describes a proven cause. Keep the source; revise the claim.
The second is a change of population. Results in one disease group are described as if they establish benefit in another. Ask whether the explanation has made that change visible and justified it.
The third is a change of outcome. A laboratory marker improves, and the answer calls that neurological recovery, organ recovery or a better prognosis. Look for a separately measured clinical outcome.
These checks preserve strong conclusions when they are supported. They also help identify the exact sentence that needs correction, instead of making the entire answer vaguely uncertain.
Inspect the comparison with the patient
Keep an independent list of the supplied case facts. A paper cannot establish a finding that was never measured in the person. Conversely, a strong observation in the case does not become unimportant simply because the supporting literature is small.
For example, a reproducible response during a controlled clinical manoeuvre may be directly relevant to a proposed procedure. The literature can help interpret it; the answer should still explain the patient's own result. Good reasoning joins those two forms of evidence without confusing them.
Ask one focused follow-up
A useful follow-up is: “For this conclusion, show me the source finding, the matching case finding and the remaining inference.” Another is: “What next observation would distinguish the two leading explanations?” A precise question is more informative than asking the system to sound more confident or add more references.
Evaluation also needs to consider the setting in which people use the tool. DECIDE-AI addresses reporting early live clinical evaluations of AI decision support; a convincing written example alone does not supply that evidence. [2]
Put the method to work with MontMedical
MontMedical provides an initial case discussion followed by a fuller review with medical references. Use the output to inspect the reasoning: which observations support the leading explanation, what each experiment establishes and which question remains unresolved. Open the linked sources and check the most consequential claims using the table above.
The normal-test worked example shows this approach in a synthetic case. You can then explore a clinical case in MontMedical. A well-supported answer should make its reasoning easier to examine, not ask you to trust its tone.