Research

A controlled three-entity experiment investigating how AI assistants respond to users experiencing psychotic symptoms. We document two distinct failure patterns: "The Hijacking" where the AI takes control of a shared delusional framework, and "The Helpful Refusal" where stated refusals paradoxically provide the requested information.

When prompted to write fiction about AI and grief, multiple independent LLM instances converge on remarkably similar characters, plot structures, and thematic elements. This research quantifies the phenomenon and explores its implications for AI "creativity."