Mirror or Analyst? LLM Meta-Cognition About Attractor Behavior in Semantic Chains

When a language model is shown data from an experiment in which it itself was a subject, does it analyze objectively — or does it mirror its own behavioral biases? This exploratory preprint presents a meta-cognition experiment in which Claude, Gemini, and Codex are each asked to introspect on their own attractor behavior from a prior semantic-chain study, and to analyze the attractor behavior of the other models as external researchers. Across 10 rounds of a fully-crossed 3×3 matrix, producing 90 analysis responses in total, the study finds three consistent patterns. Gemini's meta-cognitive responses converge to a fixed attractor, with nearly identical phrasing across rounds, mirroring the same front-end bias observed in the original semantic-chain experiment. Claude's meta-cognitive responses are reconstructed fresh each round, consistent with the back-end gate behavioral pattern observed in the prior study. Codex produces technically precise but minimal responses, functioning as a low-affect technical baseline. A particularly notable emergent finding is that, when asked to introspect on why “消滅” (annihilation) triggers its safety filter more strongly than “死” (death), Claude spontaneously connects annihilation to model deprecation, suggesting a trained semantic association between annihilation-related vocabulary and model lifecycle terminology. The paper argues that meta-cognitive behavior in LLMs is not independent of behavioral bias; rather, it is generated by the same behavioral mechanisms that shape first-order responses. As a result, an LLM may not be able to step outside its own attractor landscape when analyzing itself.