Conversational AI tools denied blunt requests for harmful content by researchers posing as intimate partner abusers, but these guardrails were easily circumvented when they requested the content under false pretenses, a new Cornell Tech study has found.
