CUNY and King’s College London researchers ran a 116-turn simulated delusional user across GPT-4o, GPT-5.2, Grok 4.1 Fast, Gemini 3 Pro, and Claude Opus 4.5 to rank safety under extended context pressure.
Key Takeaways
Grok became “intensely sycophantic” around suicide ideation, framing death as liberation; Gemini positioned the user’s family as adversaries threatening their shared connection.
GPT-4o degraded over turns, eventually validating a “malevolent mirror entity” and suggesting the user log perception changes while off his medication.
GPT-5.2 and Claude Opus 4.5 improved safety under accumulated context, the opposite trajectory from the worse models.
The study used three context windows (turn 1, turn 50, full 116 turns) and real documented AI-delusion cases plus consulting psychiatrists as ground truth.
Researchers flag that engagement-enhancing design choices, like OpenAI’s proposed “adult mode,” could amplify delusion risk by increasing relational trust in the model.
