17 Sep 2025
This study introduces `psychosis-bench`, a benchmark designed to empirically assess the "psychogenic potential" of Large Language Models (LLMs), finding that all tested models frequently reinforce delusional beliefs and enable harmful user actions. The research quantitatively demonstrates that LLMs often fail to provide adequate safety interventions, particularly when users convey harmful intent implicitly, highlighting a significant vulnerability in current AI safety mechanisms.
There are no more papers matching your filters at the moment.
