Reciprocal Miscalibration in Personalized AI Correction. User model fidelity persuasive fit and feedback over time

Persistent AI agents can condition advice and correction on disclosures, histories, and inferred preferences. Apparent personalization fit is insufficient evidence of calibration. A system may misrepresent its user while the user poorly distinguishes accurate from inaccurate advice; their coexistence is termed dual miscalibration. Reciprocal miscalibration is a narrower hypothesis: representation-conditioned outputs affect behaviour, that behaviour updates the operative representation, and the updated state sustains or worsens both representation error and miscalibrated human reliance. The framework separates user-model fidelity, contextual representation fit, perceived fit, persuasive efficacy, and advice accuracy. An accurate record of a false user belief remains distinct from endorsing that belief as true. Performative prediction, preference pollution, persuasion profiling, distributed cognition, narrative railroading, and human-AI feedback research supply direct antecedents. The contribution is an operational refinement of these problems for conversational correction, with falsifiable propositions and interventions on representation updating and correction. Feedback to the human about independently checked outcomes is distinguished from behaviour fed back into the system. A factorial study tests immediate effects; a longitudinal intervention tests reciprocal reinforcement. Appropriate reliance additionally depends on payoffs, checking costs, relative competence, and learning opportunities. Neither reciprocity nor the benefit of the proposed controls has been established by this conceptual synthesis.