OpenAI's headline figures for its 1 September Epic evaluation, physicians rating 99.1% of responses safe, create a genuine opportunity to explain a distinction that applies to every clinical AI product in this category, not only this one: safe, accurate, useful and clinically effective are four separate evaluation constructs, and a response can pass one test while failing another entirely.
The four constructs, defined
Safe: the response does not create a clear risk of harm, a threshold test focused on avoiding the worst outcomes rather than confirming the best ones. Accurate: the factual content is correct, a genuinely different and often stricter bar than safety alone, since an inaccurate response can still avoid creating a clear risk of harm while being wrong. Useful: the response meaningfully improves the clinician's actual task, a practical test independent of both safety and accuracy, since a perfectly safe and accurate response that is too long, too generic or too poorly timed to actually help does not clear this bar. And clinically effective: the response changes workflow or patient outcomes beneficially, the outcome-level test that sits above and beyond the other three, requiring evidence that using the tool made a measurable difference to care, not merely that its outputs were individually sound.
Worked examples showing the gaps between them
A summary can be safe but omit the most important result: nothing in what it includes creates risk, and the single most consequential piece of information the chart contained is simply absent, a genuinely dangerous gap a narrow safety definition can miss entirely. A response can be accurate but too long to use: every fact correct, and delivered in a form a clinician does not have time to read during a real consultation, failing the useful test despite passing the accurate one cleanly. A recommendation can be useful but unsupported by the cited source: genuinely helpful direction that a clinician could act on productively, resting on a citation that, checked directly, does not actually say what the recommendation claims it says, the citation-fidelity failure this cluster's evidence-literacy coverage treats throughout. An AI tool can save time while increasing downstream investigation: genuinely useful at the point of use, and generating a pattern of additional testing that offsets or exceeds the time saved once the full care episode is considered, failing the clinically effective test even while passing the useful one. And a technically accurate summary can amplify an incorrect chart entry: faithfully representing exactly what the record says, while the record itself contains a stale or duplicated error, producing an accurate summary of inaccurate underlying information, the specific risk this cluster's dedicated copy-forward-errors analysis examines in depth.
The reusable editorial rubric
Eight dimensions worth applying to any clinical AI product review, this platform's own products included, rather than treating any single headline figure as a complete verdict. Evidence accuracy: is the underlying clinical claim correct. Source fidelity: does the cited source actually support the specific claim attached to it. Clinical relevance: does the response address what the clinician actually needed to know, not merely what was technically asked. UK applicability: does the guidance, medicine information or pathway reflect UK-specific practice where that matters. Time saved: does using the tool genuinely reduce the time a task would otherwise take, measured realistically rather than assumed. Harmful omission: has anything critical been left out, the failure mode a narrow safety definition can miss. Appropriate uncertainty: does the response communicate genuine uncertainty honestly rather than presenting a confident answer the underlying evidence does not support. And effect on learning and judgement: does repeated use of the tool build or erode the clinician's own independent capability over time, the longer-horizon question this cluster's coverage of AI-assisted clinical education treats throughout.
Applying this to any headline evaluation figure
Whenever a vendor reports a single evaluation number, safety, accuracy or otherwise, the productive response is to ask which of the four constructs it actually measured, and which of the remaining constructs, and the eight rubric dimensions beneath them, it left untested. OpenAI's 99.1% figure, examined in full in this cluster's dedicated analysis, measured safety specifically, on a defined case set, under conditions that remain only partially disclosed, leaving accuracy, usefulness and clinical effectiveness each as separate questions the headline figure does not itself answer.
Frequently asked questions
Can a tool be clinically effective without being individually accurate on every response?
In principle a tool could produce net benefit despite occasional inaccuracy if the overall workflow catches and corrects errors before they matter, though this would require outcome-level evidence specifically, not inferred from accuracy figures alone, which is precisely why the four constructs need measuring separately rather than assumed to track each other.
Should this rubric be applied to iatroX's own products too?
Yes, explicitly: the rubric is built to be applied evenhandedly, and any platform in this category, this one included, should expect its own claims measured against all four constructs and all eight dimensions rather than judged more leniently than a competitor's.
Is safety the most important of the four constructs?
It is the necessary floor rather than the ceiling: a response failing the safety test is disqualified regardless of how accurate, useful or effective it might otherwise be, and a response clearing the safety floor still needs to clear the other three tests before it represents genuine clinical value.