Nuclear command and control refers to the people, procedures, and technical systems that detect possible attacks, inform leaders, and carry out a decision to use nuclear weapons. As AI improves at processing sensor data and supporting decisions under time pressure, a sharp debate has emerged over whether it belongs anywhere near these systems, and where the firm line against automation should be drawn.
The dominant position, shared across much of the policy community, is that a human must always make the decision to use a nuclear weapon. In 2023, US lawmakers introduced the Block Nuclear Launch by Autonomous Artificial Intelligence Act, which would prohibit federal funds for “any system that launches a nuclear weapon” or selects nuclear targets “without meaningful human control.” In November 2024, Presidents Biden and Xi affirmed in a joint statement “the need to maintain human control over the decision to use nuclear weapons,” the first time the two governments had said so together.
Even with broad agreement on keeping humans in charge of the launch decision itself, the harder questions concern everything upstream. AI used for early warning, intelligence fusion, or option generation could speed up decisions in ways that compress the time leaders have to think, or could feed them flawed analysis with unwarranted confidence. Critics warn that automation bias, the tendency to trust a computer’s output, could erode human control in practice even where it is preserved on paper. Others argue that the same systems are vulnerable to hacking, spoofing, and unfamiliar situations that AI handles poorly.
For a general reader, this is the highest-stakes case of a recurring question: how much should automation be trusted in decisions where a mistake is catastrophic and irreversible. The emerging consensus, keep a human firmly in the loop on the gravest choices while remaining wary of subtler automation creep, is a useful template for thinking about AI in any high-consequence system.