Limit the use of fully autonomous AI systems to make high-risk decisions

Meaningful human oversight should also be present, particularly in high-impact public and private uses, such as in healthcare triage and treatment, or criminal justice and law enforcement.

Current AI systems can appear confident while remaining uncertain, difficult to audit, and prone to rare but significant failures. The immediate risk is not whether AI systems are sufficiently intelligent or wise, but that decision-making power is being delegated before reliability, uncertainty handling, and accountability are sufficiently understood. Delaying action increases the likelihood that unsafe practices become embedded and their effects harder to reverse.