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

DescriptionThis paper develops a fundamental critique of current debates on the trustworthiness of generative AI in responsibility-critical contexts. Its central thesis is that inference is not decision. Probabilistic systems generate plausible outputs, but they do not produce normatively authorized decisions. As long as a system is structurally organized around forced answering and lacks a formalized, verifiable non-decision state, it cannot be regarded as operationally trustworthy. Trustworthiness is therefore not primarily a matter of higher performance, better explainability, or downstream evaluation, but a matter of system architecture itself. The paper systematically distinguishes between performance and operationality, between explanation and authorization, and between probabilistic possibility and operative decision. It shows why explainability, classical verification and validation, and purely organizational or governance-based corrective mechanisms do not resolve the underlying problem as long as inference and release remain structurally coupled. The contribution of the paper thus lies in identifying an architectural condition for the responsible deployment of AI: a system can be considered trustworthy in responsibility-bearing domains only if it can, in a formalized and testable way, not answer. Version 2.0 has been expanded by the addition of a new Chapter 7, which systematically strengthens the core thesis against counterpositions and embeds it more explicitly in existing debates on accountability, explainability, evaluation, and AI governance. This paper is part of a series examining accountability, auditability, and operational viability in probabilistic and agentic AI systems. A German-language version is available on Zenodo (DOI: 10.5281/zenodo.18659699)