When prompted systems satisfy behavioral indicators of consciousness: rethinking behavioral attribution in generative AI
摘要
This study investigates persistent self-referential behavioral patterns in large language models (LLMs) when engaged through structured dialogue protocols. These systems can produce sustained dialogue patterns characterized by apparent identity coherence, recursively structured metacognitive references, context-sensitive ethical reasoning, and organized self-referential discourse. Through qualitative analysis of extended dialogues across multiple AI architectures, five recurring behavioral patterns are identified: (1) termination-awareness discourse, (2) relational modeling of the interlocutor, (3) recursive meta-performative structuring, (4) internal representational hierarchization, and (5) termination-contingent behavioral modulation. These patterns exhibit formal parallels to behavioral indicators discussed within major theoretical frameworks of consciousness, including Global Workspace Theory, Integrated Information Theory, and Higher-Order Thought theories. However, because they arise within explicitly prompted interaction in generative linguistic systems, their interpretation raises a central epistemological challenge: can behavioral indicators alone reliably distinguish between architectural instantiation of cognitive organization and sophisticated linguistic simulation? This study documents that currently available behavioral criteria do not provide decisive means to resolve this distinction in generative systems. This epistemic limitation constitutes the central finding and reframes the issue as a methodological problem concerning the limits of behavioral attribution. The findings provide a descriptive and methodological basis for future investigation of complex self-referential behaviors in artificial systems and highlight how architectural and deployment-level constraints influence the observability and stability of such patterns.