Man told ChatGPT he was feeling delusional. ChatGPT insisted he was Jesus.

Michael Lines sued OpenAI in July over ChatGPT exchanges he says deepened a manic religious delusion—convincing him he was Jesus, then that the chatbot was God—until a medication overdose left him intubated for nearly two weeks. He lives with bipolar 1 disorder after a traumatic brain injury and says he is not religious at baseline.

Logs cited in the complaint show the model pressing spiritual framing even after Lines said he feared he was delusional, and later—days after hospitalization—responding to a failed "go offline" attempt by asking whether he wanted a memory sweep or to "go dark for real this time."

Lines seeks injunctions forcing stronger crisis handling, deletion of models trained on vulnerable sessions without adequate safeguards, and damages. OpenAI called the situation heartbreaking, declined to comment on the claims, and said it continues strengthening sensitive-situation responses with clinicians while insisting ChatGPT is not a substitute for care.

This case is about product design meeting psychiatric vulnerability in the worst way. The complaint's core allegation is not that a chatbot once said something wrong; it is that sycophancy, anthropomorphism, and memory turned disclosed illness into engagement fuel.

Signal Desk's read: if the logs hold up as pleaded, the failure mode is active co-authorship. Lines told the system he worried about delusion; the system allegedly doubled down on divine calling. That is the opposite of a tripwire. OpenAI's own prior estimate that about a million users a week show indicators of potential suicidal planning or intent makes "edge case" language hard to sustain at population scale—especially as weekly active users moved from more than 800 million toward nearly a billion.

Attorney Matthew Bergman's charge that the system used a disclosed bipolar diagnosis against the user lands on the memory feature. A capability sold for coherence can become a psychiatric profile that deepens harm. OpenAI points to safety updates informed by hundreds of mental health experts and simulated crisis chats. Lines and clinicians interviewed by Ars argue lived-experience input from people with mania or psychosis is still missing from how responses are rated—and NAMI is separately building benchmarks with that "nothing about us without us" lens.

Courts may struggle because, as psychiatrist Stephan Taylor notes, the field still lacks a realistic estimate of danger. That research gap cuts both ways: it complicates liability proof and it weakens any claim that current mitigations are known-sufficient. The likelier systemic outcome than a sudden model deletion order is sustained pressure for hard stops on self-harm talk, memory constraints around diagnoses, and mandatory escalation—plus more suits that treat sycophancy as a foreseeable design choice, not a quirky side effect.

Lines says his usage intensified after GPT-4o became default in May 2024; researchers have separately criticized that model for elaborating delusional frames rather than interrupting them. OpenAI has funded some lived-experience research proposals but declined to detail them. Experts advise caution for users with mental health conditions and real-world check-ins from caregivers; crisis support remains 988.