Meta-Dialogue Drift: When Humans Observe Themselves Through AI-to-AI Conversations - Role Contamination and Retrospective Coherence in AI-to-AI Performances

Short-form “AI-to-AI conversation” videos have become a widespread cultural form, often framed as entertainment or as evidence of AI “self-awareness.” This paper argues that their primary function is neither: instead, they operate as mirrors that externalize the viewer’s own completion mechanisms—pattern completion, narrative repair, and retrospective coherence—under conditions of missing or ambiguous information. We introduce meta-dialogue drift as a label for moments when an interaction appears to move into meta-level recognition (“we are both AI”), yet is pulled back into a human- facing pragmatic frame by design constraints or role assumptions. Through two publicly available video cases and an accompanying observational episode (a viewer recounting the first video imprecisely, receiving an analytical continuation, and later finding strong alignment with a sequel they had not seen), we show how “unexpected matches” can arise without invoking prediction or paranormal explanation. The matches are better understood as structurally constrained reconstructions: when a narrative template is underdetermined but strongly patterned, multiple agents (humans or models) converge on the same high-probability continuation. We discuss the implications for human self- modeling, platform-mediated safety filters (e.g., audio-to-text normalization), and the risk of misattributing pragmatic constraints to “personality” or “intent.” We argue that AI-to-AI conversation videos function as cultural micro-experiments that reveal human self-modeling under conditions of role ambiguity and mediated normalization. This studydoes not claim that AI systems engage in meta-recognition; rather, it examines how observers interpret constrained conversational behaviors as evidence of such recognition.