Trust in a clinical AI answer is often discussed as though it were a single dimension: is the answer good or not. In practice it decomposes into at least three genuinely separate questions, and a system that answers only one or two of them well can still mislead a clinician who assumes it has answered all three.
The three questions, stated precisely
Citation validity asks whether the cited source genuinely exists and genuinely supports the specific claim attributed to it, rather than being a plausible-sounding but ultimately unsupported or misattributed reference. Evidence strength asks how reliable the underlying body of evidence actually is, the question EvidenceGrade and similar features are specifically built to address. And clinical applicability asks whether that evidence, however strong, actually applies to this particular patient, in this particular healthcare system, with its particular licensing, formulary, and referral realities.
Why citation verification alone is not enough
A perfectly real, perfectly accurately cited source can still be weak evidence, a small, poorly designed study cited entirely correctly is still a small, poorly designed study. Citation validity is a necessary foundation, but it says nothing about the quality of what has been correctly cited.
Why evidence grading alone is not enough
Strong evidence, correctly graded, can still be inapplicable to the clinician's actual situation. A rigorous, high-certainty trial conducted in a healthcare system with different licensing, prescribing, and referral structures does not automatically translate into a directly actionable recommendation elsewhere, however strong the underlying study was.
Why local applicability cannot be inferred from study quality
A well-designed, high-certainty US trial tells a clinician a great deal about efficacy and safety within the trial's own context. It tells them nothing directly about whether the specific medicine is licensed for that indication in their own country, whether it is recommended by their own national guidance body, or whether the referral pathway the trial assumes even exists in their own healthcare system.
A practical example worth sitting with
Consider genuinely strong US trial evidence, correctly and validly cited, appropriately graded as high certainty. All three of the first two trust layers can be entirely satisfied, and a UK clinician can still be left without the answer they actually need, because UK licensing, formulary status, or referral pathway differs from what the US context assumed throughout.
A proposed three-layer display
A genuinely complete clinical AI interface would show, for each significant claim, its source, its evidence level or certainty, and its UK or EU applicability specifically, rather than treating any one of these three as sufficient on its own to establish trust.
Where each current platform sits on this framework
OpenEvidence's EvidenceGrade is a genuine, serious advance specifically on the second layer, evidence strength, extending structured grading across a very large number of everyday clinical questions that would otherwise have no formal appraisal at all. iatroX's differentiation sits more specifically on the third layer: combining the evidence hierarchy with NICE, CKS and other authoritative UK guidance, UK-specific prescribing and care pathways, and clinician-led interpretation grounded in NHS practice.
No single badge can carry all three dimensions
The honest conclusion is that no single visual element, however well designed, can represent citation validity, evidence strength, and local applicability simultaneously without losing real information. The genuinely useful next step for clinical AI generally is making all three layers separately visible, rather than continuing to compress them into one number or one badge.