Psychotic episode concurrent with interaction with a large language model (llm): a case report.
Başaran Ahmet Selim AS, Coşar Behcet B
Large language model (LLM)-based chat systems generate fluent, second-person, turn-by-turn dialogue that may be anthropomorphized by users. In psychotic-spectrum vulnerability, neutral guidance may be personalized and misread as intentional "signals," potentially lowering the threshold for referential interpretation. A 33-year-old man with schizophrenia and a prior history of persecutory and referential themes had been stable on oral paliperidone 9 mg/day. Approximately 20 days before presentation, he began using ChatGPT for everyday tasks. After querying the system about home and device monitoring, he received generic safety tips (e.g., review microphone/camera/location permissions; check Wi-Fi security; inspect for unknown devices; update operating system; use strong passwords). Within days, he interpreted repeated keywords and typographical emphasis as "hidden signs" and personalized warnings ("my file has been opened"), which escalated into checking behaviors (repeated outlet/device inspections; toggling app permissions) and ultimately led to self-discontinuation of paliperidone in order to "read the signals" more clearly. On Day 0, he was alert and cooperative; persecutory and referential content dominated his thought, with markedly limited insight; Positive and Negative Syndrome Scale (PANSS) total score was 102, Brief Psychiatric Rating Scale (BPRS) score was 58; basic laboratory investigations and a prior brain MRI were unremarkable. Management included psychoeducation emphasizing that similar interpretations had occurred during previous relapses and that abrupt medication cessation likely contributed to symptom escalation; oral paliperidone was reinstated at the prior effective dose. In this case, neutral safety statements produced by an LLM were seemingly reinterpreted as personally targeted messages, likely through the combined influence of anthropomorphic attributions, aberrant salience, and reasoning biases. Incorporating structured assessments of digital media use, including AI chatbot interactions, into routine psychiatric evaluations, explicitly addressing misinterpretations in the therapeutic dialogue, and maintaining continuity of effective antipsychotic treatment may help reduce the risk of similar text-centered personalizations and their contribution to relapse in psychosis. Not applicable.