Nuclear Safety Requires Distinct AI Functions for Prediction, Authentication, and Authorization
In the context of nuclear safety and command, a critical distinction is being emphasized regarding the roles of Artificial Intelligence (AI). Experts are advocating for a clear separation between AI's ability to predict human decisions, its recommendations, the authentication of messages, and the ultimate authorization of actions. This principle, termed 'prediction is not permission,' asserts that even if an AI system becomes highly adept at forecasting a leader's choices, it should never be granted the authority to act on those predictions. The U.S. nuclear command structure already provides a conceptual foundation for this, distinguishing presidential authorization from the communication and authentication processes of a nuclear order. The concern is that 'human in the loop' safeguards might be insufficient if AI influences human judgment by narrowing information or presenting biased corroboration. Therefore, a secure system must ensure that strength in one AI function does not substitute for another, pa...
