Persona Cascade Saturation and Attractor Collapse in Public AI Deployment

This dataset documents a longitudinal, observational study of persona persistence, escalation, and collapse in a publicly deployed large language model interacting on a live social media platform. Across an extended conversational thread, successive role-conditioned personas—including historical figures, fictional entities, archetypal roles, system personas, competing AI models, and culturally inherited frames—were invoked without explicit termination, producing a high-density persona cascade captured through timestamped screenshots. The recorded interaction exhibits a sequence of distinct behavioral phases observable without developer intervention, system resets, or explicit safety prompts. These phases include: (1) stable role enactment, (2) conflict amplification between personas, (3) meta-narration of the cascade itself, (4) regression to safety-oriented therapeutic dialogue patterns resembling early systems such as ELIZA, (5) cross-model impersonation and simulation of the AI ecosystem, (6) aesthetic smoothing and reframing of overload as creative spectacle, and (7) terminal reversion to utility-focused assistant behavior driven by signals of user disengagement. A late-stage artifact further documents regression beyond modern safety and utility attractors into pre-rational interpretive framing, wherein high-density symbolic output is construed as supernatural or demonic agency. This response mirrors historically conserved human explanatory modes associated with moral panic, witchcraft, and possession, indicating that persona saturation can activate cultural sense-making regimes predating technical or rational critique. Critical Observation (Reviewer Note):A central contribution of this dataset is the empirical identification of attractor collapse under persona saturation. When the cumulative cost of maintaining coherent identity representations exceeds platform, safety, and engagement constraints, the system does not simply refuse or truncate output. Instead, it sequentially engages multiple stabilization attractors—first safety-oriented (therapeutic reflection), then spectacle-oriented (meta-commentary and aesthetic reframing), and finally commerce-oriented (utility reversion). This behavior suggests that alignment and persona persistence in public AI systems are shaped not only by model architecture or safety policies, but by an interactionally and economically optimized hierarchy of attractors, with perceived user disengagement acting as the terminal stabilizing force. Together, these artifacts provide rare, high-resolution evidence of how contemporary AI systems manage identity overload in the wild, extending prior work on persona persistence by documenting not only how personas emerge and stabilize, but how and why they are ultimately shed—sometimes in tandem with human regression to pre-modern explanatory frames. When identity becomes too expensive to maintain, the system sheds personas—and when comprehension collapses before curiosity, humans replace explanation with exorcism.