From Regulatory Authorization to Clinical Accountability: A Control Preventability Framework for Healthcare AI

Abstract

The integration of artificial intelligence into healthcare has accelerated dramatically, with the FDA authorizing 1,430 AI/ML-enabled medical devices by the end of 2025 including 331 authorizations in that year alone yet the evidentiary foundation for many deployed tools remains limited, and accountability frameworks have not kept pace with technological adoption. This integrative evidence–regulatory analysis synthesizes FDA authorization data, selected high-relevance clinical evidence from randomized trials and systematic reviews, and primary regulatory and governance materials from the FDA, European Union, and World Health Organization. The analysis reveals that among 1,357 FDA-cleared AI devices identified through December 2025, only 2.5% were linked to registered prospective trials and only 0.2% had identified evidence evaluating patient-centered outcomes. A randomized clinical trial involving 44 physicians found that exposure to erroneous large language model recommendations reduced diagnostic reasoning accuracy by 14.0 percentage points (95% CI, −18.9 to −9.1; P<0.0001) despite prior AI literacy training, while a prospective observational study of a sepsis early warning system found provider confirmation of alerts was associated with an adjusted absolute mortality reduction of 3.3% (95% CI, 1.7%–5.1%). A systematic review and meta-analysis of 83 studies found no significant overall difference in diagnostic performance between generative AI and physicians, though generative AI performed significantly worse than expert physicians. Under the July 2026 Digital Omnibus, AI systems incorporated into medical devices remain in Section A of Annex I of the EU AI Act, subject to the high-risk framework integrated with existing medical-device regulations, with Chapter III obligations deferred to August 2028. The available evidence suggests that publicly identifiable rigorous prospective clinical validation is uncommon among FDA-authorized AI devices, and accountability frameworks remain fragmented, creating a risk that clinicians may bear disproportionate responsibility for harms arising from systems they do not fully control. This paper proposes a Control–Preventability Principle that allocates accountability according to control, foreseeability, preventability, and meaningful oversight capacity. Closing the evidence–accountability gap requires strengthened evidentiary standards, lifecycle governance, and clear allocation of responsibility across developers, deploying institutions, and clinicians, with accountability following control over preventable risk rather than merely the physical location of the final clinical decision.