Inference Is Not Decision: The Axiomatic Non-Trustworthiness of Always-Answer AI Systems

Abstract

Description The preceding papers in this series argued that explainability cannot ground accountability in probabilistic AI systems, that hallucination is the expected epistemic consequence of forced answer generation, and that accountable AI therefore requires a formal system language of non-decision, decision boundaries, and commitment. Building on these results, the present paper formulates their operational consequence: probabilistic inference is not yet a decision. It begins from a simple axiom: a technical system that is structurally required to generate an answer to every input cannot be operated in a trustworthy manner. Trustworthiness presupposes not only the capacity to act, but the formally representable capacity, under defined conditions, not to act. Where this capacity is absent, uncertainty is not governed but suppressed, and probabilistic plausibility is released as if it were an authorized decision. The paper argues that this limitation cannot be resolved by explainability, evaluation, or post-hoc oversight alone. Explanations may reconstruct how an output was generated, but they cannot authorize whether that output should have been released in the first place. Likewise, guardrails, human review, and governance procedures remain secondary as long as inference and release remain structurally coupled. The decisive question is therefore not whether a system can generate plausible outputs, but whether it can formally distinguish between generated possibility and admissible decision. The contribution of the paper is to make this threshold explicit. It does not argue against generative AI as such, nor against probabilistic systems as tools of exploration, support, or reasoning. It identifies a prior operational condition for trustworthy deployment in responsibility-bearing domains: a system must be capable, in a justified, testable, and formalized way, of withholding output. Only under this condition can responsibility attach to release rather than be reconstructed after the fact. By establishing this distinction, the paper turns the analysis from semantic accountability to structural legitimacy. If trustworthy deployment depends on a capacity for formalized non-release, then the next question becomes unavoidable: where does this capacity come from? The following paper addresses this by arguing that decision legitimacy cannot emerge from inference itself, but requires prior structure that determines in advance what may be said, when, and under which constraints.