Summary
A Nature Medicine editorial argues that clinical AI needs to be evaluated in real care settings, not just on performance metrics. It highlights workflow fit, local adaptation, safety and ongoing monitoring as conditions for useful deployment.
A Nature Medicine editorial published on 14 October 2026 argues that clinical AI should be judged by its effects in routine healthcare, not only by performance on retrospective datasets or controlled reader studies. Drawing on work in the journal’s issue on real-world deployment, it identifies workflow integration, local adaptation, safety and continued oversight as central challenges. The editorial is a synthesis, not a new clinical trial.
The gap matters as health systems adopt AI tools for tasks including medical-image interpretation, clinical documentation and decision support. The editorial cites a survey in which more than 90% of US health systems had begun deploying AI tools, while 53% or fewer found them highly useful. It says evidence of effects on patients, healthcare workers and health systems remains limited.
Clinical performance depends on how a tool is used
Strong results in retrospective evaluations do not automatically translate into value in clinical practice. A system has to fit the care pathway and the work of the people using it. The editorial describes a report on an AI eye-care system in China that found the technology feasible only when embedded at every stage of the care pathway.
Implementation also depends on local conditions. Healthcare settings differ in their data infrastructure, specialist workforce, computing resources and models of care, so a system that works in one hospital may need substantial adaptation elsewhere. The editorial points to work on a diabetic-retinopathy AI system deployed across India, Thailand and Australia: implementation strategies differed according to local priorities and clinical workflows. It also notes that resource-constrained settings may need more “frugal” systems designed around local capacity.
The editorial highlights safety and reliability over time as further concerns. It says research has found large language models can generate inaccurate or misleading medical advice, and cites safety concerns in mental-health applications. It also describes a report in which changes to model versions and electronic health-record documentation could contribute to performance drift after deployment of the ChatEHR system. These examples underpin its call for ongoing monitoring, with clear responsibility for responding to emerging risks.
Evaluate effects in care, not just on benchmarks
The editorial calls for evaluation to begin early in development and continue before, during and after deployment. It argues that studies should assess clinical outcomes, care delivery, cost-effectiveness and health-service performance alongside technical measures. Prospective approaches—including hybrid effectiveness–implementation trials and quasi-experimental studies—can examine both whether an intervention works and how it is adopted in real settings.
It also stresses that clinicians and patients need to find a tool useful and acceptable in practice. In a pilot evaluation cited by the editorial, clinician adoption of an AI decision-support system declined over four weeks as engagement fell. The editorial argues that prospective real-world studies should examine effects on workflows and staff experience, while incorporating patient perspectives that can reveal barriers developers may overlook.
The practical test, in the editorial’s account, is whether AI helps deliver better care. That requires health systems to treat deployment as an ongoing clinical and organisational process, rather than assuming that a strong model score alone predicts benefit.