Generative and agentic artificial intelligence systems in medicine may now change faster than clinical evidence can be generated, reviewed and implemented, according to a Perspective published in Discover Artificial Intelligence. The article, authored by Mohammad Hossein Imani and Amir Hossein Imani, argues that a clinically evaluated AI system can silently stop being the system that was evaluated, because updates to the foundation model, prompt, retrieval layer, generation settings, tools, memory, interface or workflow can alter behaviour without changing the product name visible to investigators or users.
The authors synthesize existing evidence on repeated-run instability, cross-version performance, human–AI interaction, agentic systems and post-deployment monitoring. Their central observation is that the traditional assumption behind clinical evaluation — that an intervention remains stable once it has been tested — does not hold for software that can be updated continuously and, in some cases, autonomously.
In response, the paper proposes VERSION-AI, described as a testable evidence-refresh framework. The framework separates two questions that the authors say are often collapsed into one: what changed in the system, and what are the clinical consequences if the changed system fails. Change classes labelled C0 through C8 describe the system or context modification, while a distinct clinical-consequence tier reflects intended use, decision dependence, time sensitivity, and the likelihood and severity of harm from incorrect, omitted or delayed actions.
The intersection of these two axes determines the minimum evidence response. Depending on the combination, evidence may remain current after an audit, or the change may require usability testing or repeated-run bridging, cross-version or local validation, shadow evaluation, continuous monitoring, rollback capability, or renewed prospective clinical assessment. The framework is illustrated with a pathway in which a clinical AI claim applies to an evaluated system snapshot; a detected modification is first classified by what changed and then assessed independently for clinical consequence, with lifecycle controls maintaining evidence only within the validated configuration, intended use and clinical context.
Importantly, the authors are explicit about the status of their proposal. Because VERSION-AI has not yet been prospectively validated, it is presented as a research and governance proposal rather than a regulatory instrument. The framework is offered as a testable pathway, not a settled standard, and the paper calls for empirical work to determine whether it performs as intended.
The Perspective’s conclusion states that generative and agentic AI challenge the assumption that a clinically evaluated intervention remains stable, and that evidence should therefore attach to a time-stamped system snapshot operating in a defined population, workflow and oversight arrangement. Transparent reporting, the authors argue, is necessary but insufficient: after a material change, transferability of earlier evidence should be demonstrated rather than assumed.
A key conceptual claim of the paper is that change class and clinical risk should not be collapsed into a single hierarchy. Two systems could undergo the same type of modification — for example, an updated retrieval layer — yet warrant very different evidence responses depending on how much a clinical decision depends on the output and how severe the harm from an error would be. Keeping the two axes separate, the authors contend, allows the required evidence to reflect both what changed and what failure would mean for patients.
The authors identify the empirical questions now facing the field. Validation is needed, they write, to determine whether this two-axis approach improves consistency, detects clinically meaningful regression, and remains feasible across health systems with different resources. Until such validation exists, the framework’s practical value remains an open research question rather than a demonstrated result.
The article also notes a data limitation: no datasets were generated or analysed during the study, which is consistent with its nature as a synthesis and proposal rather than an empirical study. The paper’s overarching message to clinicians, developers and governance bodies is captured in a single principle from the abstract: evidence transfer after a material system change should be demonstrated rather than assumed.
Subject of Research: A governance framework for maintaining current clinical evidence as medical AI systems are updated.
Article Title: Keeping clinical evidence current as medical artificial intelligence systems change
Article References: Imani, M. H., & Imani, A. H. (2026). Keeping clinical evidence current as medical artificial intelligence systems change. Discover Artificial Intelligence, 6(1), Article 1654. https://doi.org/10.1007/s44163-026-02636-8
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02636-8
Keywords: medical artificial intelligence, generative AI, agentic AI, clinical evidence, VERSION-AI, post-deployment monitoring, model updates, validation, health governance, shadow evaluation, rollback, patient safety
News Source: Denise Maddox. (October 28, 2026). New framework asks when medical AI updates require fresh clinical evidence. Scienmag.
