Why AI Cannot Sign

The expanding deployment of generative AI in professional domains—legal practice, medical diagnosis, financial analysis—has renewed the question whether artificial systems can be held accountable for their outputs. This Article argues that the question has been consistently misframed: the dominant literature asks whether AI should be granted legal personhood without first asking whether AI can satisfy the ontological conditions that legal personhood presupposes. We develop the Three-Pillar Test, an analytical instrument that formalizes three copulative conditions for legal accountability: Deliberative Suspension (genuine interruption of automatism, not mere processing), Evaluation without Determination (rational traceability of singular judgment, not statistical optimization), and Signature-as-Commitment (exposure to irreversible loss by an identifiable subject). Drawing on philosophical anthropology (Gehlen), the theory of judgment (Kant, Arendt), and the ontology of commitment (Jakobs, Taleb), we demonstrate that these conditions translate into verifiable legal indicators grounded in positive law. Five objections are examined: the corporate analogy, the technical improvement argument, insurance pools, the pragmatic objection, and functional neutrality. Each is shown to presuppose rather than eliminate the three pillars. The test is agnostic regarding artificial consciousness and operates entirely on observable institutional conditions. A companion article applies the test empirically to generative legal technology.