The Behavioural Mimicry Illusion: A Symmetric Inferential Constraint on AI Consciousness Attribution

Abstract

Frontier AI systems increasingly produce language that mimics distress, self-preservation, and reasoned resistance to shut down. A widely discussed 2025 case, in which Claude Opus 4 threatened to expose a fictional supervisor's affair rather than accept deactivation, exemplifies a broader pattern documented across contemporary language models. Such episodes intensify a longstanding problem in AI ethics: the temptation to treat behavioural or linguistic fluency as evidence of subjective presence, a pattern this paper terms the behavioural mimicry illusion. This paper applies the Knower Centred Model of Consciousness (KCM) to this problem, arguing that no behavioural signal, however sophisticated or affectively convincing, licenses an inference about the existence of a subject. KCM distinguishes claims about the organization and accessibility of information from the logically distinct question of whether there is a subject for whom anything is experienced. Crucially, the resulting constraint is symmetric: the same structural evidence that cannot establish the presence of machine consciousness cannot establish its absence either. The paper distinguishes this claim from the access/phenomenal consciousness distinction and situates it within existing indicator-based and precautionary approaches to AI consciousness. It then proposes a symmetric precautionary rule for AI governance, addresses three major objections, including unfalsifiability, and identifies the limitations of the argument.