More capable models make more useful products possible. They also make hallucinations, bias, and misuse more expensive. Better AI needs better responsibility, not just better demos.