Responsibility Without Punishment? Prescriptive Control and the Crisis of the Punitive Model in Distributed AI

Distributed AI systems increasingly challenge one of the foundational assumptions of modern legal responsibility: the possibility of coherently attributing conduct, consequences, and sanctions to identifiable subjects. Contemporary approaches to AI governance generally address this difficulty by extending traditional frameworks of accountability and liability across multiple actors involved in the lifecycle of AI systems, while leaving their underlying attributional structure largely intact. This paper argues that legal punishability should not be understood as an intrinsic property of socio-technical systems, but as a function of the degree of prescriptive control and causal traceability that can effectively be exercised over system behavior. As these conditions weaken in distributed and adaptive AI systems, traditional models of punitive attribution progressively lose both operational effectiveness and normative coherence. On this basis, the paper develops a typology of governance configurations based on different combinations of prescriptive control and causal traceability. It argues that, beyond certain thresholds of distributed complexity, punitive governance risks becoming increasingly arbitrary, detached from sufficiently stable conditions of imputability. The paper concludes that governance in distributed AI systems must progressively shift from models centered primarily on attribution and sanction toward forms of continuous systemic intervention oriented toward modifying the operational conditions from which behavior emerges.