AI agency and criminal responsibility: a category mistake

The deployment of AI systems has intensified debate about criminal responsibility, giving rise to discussions of the responsibility gap and, at the level of criminal law, the retribution gap. Existing scholarship is divided between those who deny that AI systems could ever be criminal law agents and those who argue for their inclusion within the criminal law framework. This paper takes a different approach. It challenges the agentic framing underlying the debate without engaging directly with specific positions. The central argument is that applying the framework of criminal agency to current and foreseeable AI systems constitutes a category mistake. The paper proceeds in three steps. First, it examines why agentic framing is so tempting, identifying cognitive, linguistic, legal, and philosophical reasons for its persistence. Second, it elaborates the conditions an entity must satisfy to qualify as a criminal law agent, drawing on discussions of both responsibility and punishment, and shows that no existing AI system meets them. Third, it argues that abandoning agentic framing redirects attention toward the human duties, decisions, and failures that precede AI-related harms, where criminal responsibility properly belongs. Accepting this conclusion carries urgent practical implications for the regulation of AI-related harms.