‘We’re Not Doing Ethical AI, We’re Doing Legal AI’: How Compliance Shapes Responsible AI Practice

Abstract

As AI systems increasingly shape high-stakes decisions, ethics frameworks and toolkits have been introduced to promote fairness, accountability, and transparency. Yet in practice, developers work in environments where regulatory compliance fundamentally reshapes and often displaces deliberation about ethics. Drawing on 20 interviews with AI/ML practitioners across North America, Europe, and the UK, we provide an empirical account of how compliance operates not merely as a floor for ethical practice but frequently as its ceiling, and in some cases, as a substitute that forecloses deliberation about ethics. Across our interviews, ethics work persisted in development workflows only where it could be rendered as auditable evidence embedded in existing process gates; the underlying values that could not be converted into artifacts were routinely dropped. We further show that whether practitioners resist or absorb compliance logic depends not on individual values but on structural position within organizations. Our findings have implications for how AI ethics resources are designed, how accountability is assigned across development organizations, and how AI governance scholarship should account for the gap between jurisdiction-bound regulation and cross-jurisdictional AI deployment.