“It is easier to direct a smarter AI system with safety guardrails than a less intelligent one.”