Thresholds of agency in medical artificial intelligence: A liability framework for healthcare professionals under European Union law

The rapid integration of sophisticated artificial intelligence (AI) into clinical practice represents a paradigm shift for healthcare, fundamentally challenging traditional conceptions of medical li-ability and demanding a new legal framework. This article confronts this challenge by proposing a novel “Threshold Typology” to systematically analyze the liability of healthcare professionals within the European Union’s complex and multi-layered regulatory environment. We argue that the question of liability is not monolithic but turns on which of three distinct thresholds of agency a professional cross. The analysis first delineates the Regulatory Threshold, established by the pre-ventive and ex-ante obligations of the Artificial Intelligence Act (AI Act), the General Data Pro-tection Regulation (GDPR), and the Medical Device Regulation (MDR). This threshold focuses on compliance with duties of risk management, human oversight, data governance, and vigilance, where breaches can inform subsequent liability determinations. The Professional Threshold is then examined, defined by the fault-based standards of national malpractice law, which are pro-cedurally adapted by the proposed AI Liability Directive (AILD) through mechanisms like dis-closure of evidence and rebuttable presumptions of causality. Finally, the Product Threshold is explored, grounded in the strict liability regime of the revised Product Liability Directive (rPLD), which becomes directly relevant when healthcare professionals substantially modify AI systems, effectively transitioning into the role of a producer. By meticulously dissecting the interplay be-tween these five core EU instruments: the AI Act, the AILD, the rPLD, the GDPR, and the MDR, this article provides an indispensable doctrinal map. It demonstrates that liability is contingent, distributed, and highly context-specific, depending on whether the professional acts as a user, an overseer, or a modifier of the AI system. The Threshold Typology thus serves as a vital analytical tool, translating a fragmented legal architecture into a coherent and operational framework for legal scholars, practitioners, and healthcare professionals, while signaling a broader shift from reactive compensation towards a proactive governance of medical AI.