The Responsibility–Control Gap in AI: Organizational Governance and the Reconfiguration of Control through Sovereign AI

Abstract

Organizations are increasingly held formally responsible for AI-driven decisions while control over the systems producing those decisions is distributed across external data, models, and infrastructure they do not own or fully influence. This study examines how this structural tension, referred to as the responsibility-control gap, is constructed in practice, and how Sovereign AI initiatives reconfigure the conditions under which responsibility can be exercised. Drawing on ten semi-structured interviews with senior professionals across defence, telecommunications, transportation, manufacturing, and AI consultancy sectors in Sweden, the study finds that responsibility for AI-driven outputs is attributed to individual users across all cases examined, yet the practical capacity to act on that responsibility remains limited. Across three dimensions of organizational control, decision-making authority, resource access, and capacity to intervene, formal responsibility consistently exceeds the practical capacity to act on it. Internal governance limitations, including shadow IT, weak data governance, and the leadership-capability gap, constitute a second structural layer of the gap that operates independently of external infrastructure dependency. Sovereign AI initiatives meaningfully reduce the external layer by expanding organizational control over data, infrastructure, and model configuration, but do not address the internal layer. The study further shows that Sovereign AI is most productively understood not as a binary condition but as a deliberate risk positioning decision. Governance maturity is identified as a structural requirement that must be in place before sovereignty initiatives can reduce rather than merely relocate the gap. The study contributes an analytical framework for diagnosing where control is practically absent, and introduces a distinction between operational and epistemic resource access that has not been theorized in existing frameworks.