Rethinking Health Care Regulatory Sandboxes Under the European Union AI Act

Abstract AI models have enabled substantial advances in health care. Since the entry into force of the European Union (EU) AI Act in 2024, these systems have become subject to a common regulatory framework in the EU intended to harmonize AI governance across all member states. One of the act’s distinctive features is that it accommodates existing economic, technological, and regulatory differences among European countries by leaving certain implementation responsibilities to national authorities. This is evident in regulatory sandboxes, supervised environments for developing, testing, and validating AI systems before deployment. In this viewpoint, we provide a policy-oriented analysis of health care regulatory sandboxes under the EU AI Act and propose recommendations for their design. These recommendations address the need to document model compression techniques that may affect the classification of large high-risk systems above the 10 25 floating point operations per second threshold, incorporate explainability into AI systems from the design stage, identify and document biases in health care datasets, clarify the treatment of federated learning parameter updates under the General Data Protection Regulation, and strengthen the oversight of AI agents in clinical environments. The implementation timeline makes this discussion particularly timely given the approaching deadline of August 2, 2027, by which all member states must have established at least one national regulatory sandbox.