AI Hallucination as a Burden of Proof Problem

Artificial intelligence and legal systems both seek to determine truth under uncertainty in complex environments, but they diverge fundamentally in how admissibility is established. Contemporary AI systems rely on probabilistic generation, producing candidate outputs across a broad search space before applying validation, thereby permitting the inclusion of plausible but unverified states. Legal systems, by contrast, do not admit claims through exploratory or probabilistic generation; they require that defined conditions—grounded in evidence and burden of proof—be satisfied before a claim is considered. This paper advances the thesis that AI hallucination is, in part, a consequence of this reversed sequencing. It proposes that repositioning validation to the beginning of the reasoning process—through a validation-first, constraint-based framework—can reduce the admission of unsupported outputs, improve computational efficiency, and align AI systems more closely with established structures of verified truth.