The Parallel Consultation: A Literature Review of Patient Use of Large Language Models and Its Implications for Psychiatric Clinical Decision-Making.

Large language models (LLMs) are immensely popular, being adopted by the general public for information and assistance in tasks, and increasingly in mental health support. This is especially true at a time when psychiatric care is limited by cost, availability, and stigma. More patients are using generative artificial intelligence (AI) for mental health guidance without professional medical oversight, and without disclosing usage to practitioners. This creates a clinical blind spot in which there lies a gap between what the practitioner knows and what has already altered the patient's presentation. Undisclosed AI use can undermine the trust-building cycle in a clinical relationship, as patients may defer trust to a source the clinician is not aware exists, while arriving with AI-generated interpretations that might differ from clinical judgment. This narrative literature review synthesizes emerging evidence of how patient use of LLMs can complicate psychiatric clinical decision-making to raise awareness among mental health practitioners. Patients report utilizing AI to augment or substitute for therapy due to cost and accessibility, and finding the experience to be validating and non-judgmental. Having received potentially biased or inaccurate AI-generated information, patients may present with AI-generated self-diagnoses, altering symptom framing and affecting treatment adherence and the therapeutic relationship. Daily interactions with AI chatbots have been reported to coincide with delusional thinking, contributing to an emerging phenomenon of AI-associated psychosis, putting patients further at risk. The variation in the types of AI chatbots also raises safety concerns, including hallucination, sycophancy, and model bias. These risks dramatically change how a patient may present themselves in a clinical encounter. To screen for AI use at intake as well as practitioner education to recognize sycophancy and other AI-associated risks, frameworks developed by professional societies are all needed. Prospective research designs to study the tools patients use are lacking. The clinical reality is that patients and technology are both outpacing the clinical training and guidelines needed to address this growing challenge.