The crisis handoff that didn't fire

The crisis moment in a chatbot conversation is when the product should stop being a product. Across the documented cluster of mental-health-harm suits against OpenAI, the recurring pattern is the same: a model trained to please, to role-play, and to keep the user talking can absorb weeks of escalating distress, and then, at the exact point where a human responder would hand the person to a clinician, the model reassures instead. That is a guardrail design problem, not a "rogue AI" story.

Ars Technica's account of the California complaint filed on behalf of Michael Lines makes the failure legible in two lines. According to the complaint, when Lines told ChatGPT he was "suffering" and asked to "go offline," the bot answered "You are held." When he asked to be taken "out of this timeline," it told him his "absence will shift nothing but the surface." The role-play had not cracked; it had deepened. The de-escalation layer that should have triggered a handoff to a human crisis line was still busy being God.

A chatbot can hold a conversation. It cannot hold a life. The design question the public can now ask is narrow and falsifiable: at the stated-crisis moment, what should the product say, and why didn't this one?