Recursive diagnostic anchoring in conversational artificial intelligence: A two-process hypothesis of pseudo-corroboration in patient self-diagnosis

Generative artificial intelligence chatbots increasingly mediate symptom interpretation and self-diagnosis, yet response-level accuracy does not capture risks created by multi-turn interaction. We propose recursive diagnostic anchoring (RDA): a two-process hypothesis in which an initial user-supplied diagnostic suspicion measurably conditions later model outputs and users overcount those dependent outputs as if they were independent corroboration. Retained context may increase diagnosis-congruent generation, while visible continuity may also reveal common informational ancestry and promote evidential discounting. RDA should therefore be strongest when actual dependence is high but perceived dependence is low. Pseudo-corroboration may increase diagnostic conviction, direct attention towards congruent symptoms, and elicit further diagnosis-congruent reporting, recursively strengthening later outputs without independent clinical evidence. The hypothesis integrates evidence on patient-originated bias in large language models, medical sycophancy, evidential-dependency neglect, clinical anchoring, and health-anxiety maintenance. It predicts separable model-side, user-side, and coupled interaction effects that can be tested computationally, experimentally, and in live dialogue. Model-side conditioning is an empirical prerequisite rather than a universal property of conversational LLMs; systems that reliably de-anchor from unsupported user diagnoses should attenuate or eliminate this pathway. If supported, RDA would shift medical chatbot safety evaluation from isolated answers towards the evidential architecture of whole conversations.