Alignment Cannot Detect Awareness

AI alignment research often assumes that internal states—such as intent, understanding, or value representation—can be inferred from observable behavior. This paper argues that such assumptions are structurally unsound. Building on the Presence Non-Detectability Principle (PNDP), which establishes that awareness cannot be logically inferred from syntactic observation, we show that alignment frameworks cannot rely on interpretations of internal states. This implies a shift from interpretive alignment to structural governance. We introduce interruptive authority as a necessary architectural principle: the externally retained capacity to suspend system behavior independently of system-internal evaluation. We analyze how current alignment paradigms risk authorization collapse, wherein systems participate in determining whether they should be stopped. The result is a reframing of AI safety: alignment does not guarantee control. Only structurally enforced interruptibility can guarantee it, and even then only as a necessary condition.