Understanding Is Not Authority: What Turing and Searle Could Not Ask About Acting Machines

Alan Turing and John Searle established two of the most durable tests in the philosophy of artificial intelligence. Turing displaced the unstable question of whether machines think toward publicly observable performance. Searle argued that successful symbol manipulation does not, by itself, establish understanding or intentionality. Contemporary language models and agentic systems intensify rather than settle their disagreement. Models now generate fluent language, use tools, maintain memory, form internal representations, interact with environments, and initiate actions through software systems. Yet the field still lacks a disciplined architecture for separating two questions that are frequently collapsed: what a machine may be said to understand, and what that understanding may authorize it to do. This paper proposes Understanding Is Not Authority as a bounded field-theoretic and constitutional framework for that missing region. Its central thesis is that understanding status and authority status are independent objects. A system may earn a limited, mission-relative, receiver-indexed, transportable, reconstructible, contestable, consequence-sensitive, and revocable understanding status without thereby obtaining permission to recommend, decide, execute, persist, or delegate. Conversely, a system may possess technical permissions while lacking adequate understanding of the objects and consequences over which it acts. The paper introduces the Understanding Constitution, the Authority Constitution, the Understanding-to-Action Passage, the Understanding-Authority Matrix, the Understanding-Authority Independence Law, the No-Self-Authorization Law, the Authority Conservation Principle, and the Understanding-Authority Cross-Test. It further defines failure classes including Epistemic-to-Deontic Collapse, Receipt-to-Authority Failure, Comprehension Laundering, Delegation Amplification, and Successful-Action Validation. The proposal does not claim that machine consciousness has been established, that understanding can be reduced to one score, that human understanding is fully reconstructible, or that authorization can be derived from model confidence. No empirical validation is reported. The contribution is conceptual, formal, and architectural: a falsifiable research program for studying what begins after a machine appears to understand but before its output becomes consequential authority. The governing conclusion is that a machine may understand and still have no right to act.