How this editorial can be challenged
What would it take for promising imaging AI to become dependable help rather than another result waiting for clinical meaning?
A prediction changes care only through an action. The same improvement in ranking can create benefit, unnecessary follow-up or no practical change depending on thresholds, service capacity and the response to errors.
Waiting for complete outcome evidence before any use can delay worthwhile assistance, especially where staff are stretched and missed findings already harm patients. A demand for certainty can itself become a barrier.
The choice is not unrestricted rollout or indefinite delay. Bounded adjunctive use with explicit evidence collection can preserve learning while making it possible to stop, adapt or expand according to patient outcomes.
The new imaging papers report detection or risk discrimination, not demonstrated mortality reductions. We have not inspected their inaccessible full manuscripts, and the adolescent survey cannot establish clinical benefit or causation.
Prospective comparative evidence of better patient outcomes, acceptable false-positive burdens and performance across relevant groups would justify stronger confidence; higher scores without useful changes in care would not.
Start with the person, not the chart
I want an AI system to help a clinician catch something that would otherwise be missed. I want a worrying examination to lead to the right next step, not an avoidable delay. Those are ordinary hopes, and they are more compelling to me than a contest over whether technology deserves an optimistic or pessimistic headline.
That is why the medical papers in today's briefing interest me. They contain promising signals and important limits. We can acknowledge both without turning every caveat into a verdict against the work. A result does not have to be the finish line to be worth reaching.
A score can open a door
The fracture review and BREST-AI study address different questions: detecting an existing injury and estimating future cancer risk. Each asks whether a system can extract something useful from an image. Neither, by itself, tells us that a patient will recover sooner, avoid a late diagnosis or experience less harm.
I see that distinction as the beginning of a research agenda, not an excuse to stop. If a model separates risk groups more effectively, ask which action that separation should change. If a tool flags a suspected fracture, ask whether the clinician reaches a better decision. The next study should follow the proposed benefit into the place where somebody might actually receive it.
Help is not one thing
The adolescent survey offers a different reminder. Some respondents found chatbot support helpful, and many wanted additional help. A person can feel heard without having the underlying problem resolved. The first experience matters. It just does not supply evidence for the second.
Medical products can blur a similar boundary. An interface may reassure, a score may discriminate, and a service may improve outcomes. Those are three different achievements. A health system should be able to welcome one without advertising all three. Patients should not have to discover the distinction after making a decision.
Do not demand a perfect world before learning
The strongest objection to a strict evidence argument is that existing care is not perfect either. Waiting has costs. Clinicians work with uncertainty every day, and useful assistance may be especially valuable where time and expertise are scarce. An insistence on complete certainty would be neither realistic nor humane.
That is a reason to design careful learning, not to stop asking what happens. NICE's approach to selected fracture tools allows adjunctive use while evidence is generated, with conditions. It offers a useful principle: make the scope explicit, preserve clinical responsibility and decide in advance what would justify continuing, changing or stopping. Conditional use should be a real learning arrangement, not a permanent exemption from answering the question.
Build the receiving end
A hypothetical risk model that identifies more people for follow-up creates a practical question: who can offer that follow-up? If appointments are unavailable, a better ranking may not deliver the expected benefit. If the threshold produces too many unnecessary investigations, extra activity can consume the very capacity intended to help.
These are possibilities to test, not findings established by today's studies. They suggest what a sensible evaluation should count alongside accuracy: the decisions changed, time to appropriate care, unnecessary procedures, missed cases and experiences across patient groups. The team purchasing the tool should know who owns each consequence. A model developer cannot answer every operational question alone, and a hospital cannot outsource responsibility for the entire pathway to a score.
Give optimism a harder job
My challenge to medical AI is not to look less impressive. It is to make the next claim more useful. Show the patient pathway, state the comparison, explain where the evidence is weak and let outcomes decide which applications deserve to grow. That would make positive coverage stronger, because the reason for confidence would be visible.
And my challenge to those of us writing about AI is to keep room for earned enthusiasm. We should be able to celebrate a well-designed result without converting it into a miracle, and describe a limitation without making progress sound futile. A patient needs more than a promising score. The purpose of the work is to make sure they eventually get it.
Read the reporting
Opinion is ours. The factual record is linked below.
Kuo and colleagues — fracture-detection umbrella review, npj Digital Medicine Wei and colleagues — BREST-AI external validation, npj Breast Cancer NICE HTG739 — conditional fracture-AI use and evidence requirements Okechukwu, Bear, Fazel and Soneson — adolescent support survey, npj Mental Health Research