Research ethics governance with responsible AI sandboxes

Abstract

University research ethics committees (REC) face challenges in overseeing artificial intelligence (AI) research. Historically rooted in biomedical and social science paradigms, REC were not designed to evaluate the epistemic, temporal, and normative complexities of AI and machine learning research. The EU’s AI Act exacerbates this tension by exempting academic research from its scope while at the same time promoting the application of ethics guidelines, thereby creating a zone of normative ambiguity. This paper critically examines the resulting governance vacuum. We argue that conventional ethical review processes are inadequate in many cases for reasons inherent in AI research, which is often iterative and interdisciplinary, characterized by shifting goals and emerging risks, as well as because of the normative and socio-technical co-construction of AI technology development. We propose the Responsible Artificial Intelligence Sandbox as a model for research ethics governance. It reframes the role of REC from static evaluators to co-constructors of ethical oversight within experimental research environments. Drawing on insights from regulatory sandboxes in EU law and national contexts, this conceptual model enables dynamic, participatory, and reflexive engagement with ethics throughout the research lifecycle. Two main contributions are made: we diagnose structural misalignments of existing research ethics infrastructure and conceptualize responsible AI sandboxes as an institutional and methodological innovation that aligns ethical governance with the nature of research on and with AI.