Can artificial intelligence be kept out of the decision to use nuclear weapons?
The human-control norm for nuclear command, control and communications: what the UN resolution of December 2025 demands, why no nuclear-armed state voted for it, and what 'human in the loop' does and does not settle.
The question
Nuclear-armed states are integrating machine learning into early warning, intelligence analysis, targeting, and the communications systems that would carry a launch order. Should there be a binding rule that a human being, not an algorithm, makes the decision to use a nuclear weapon — and if so, what would such a rule actually constrain?
Why it matters
The stability of nuclear deterrence has always rested on decisions made by people under time pressure with imperfect information. Automating any part of that chain changes the failure modes: faster, less transparent, and prone to errors that no one in the chain can see. The episodes in which human judgment stopped a false alarm from becoming a launch — 1983 is the one everyone cites — are the argument for keeping people in the loop. The argument on the other side is that people are slow and fallible too, and that an adversary who automates first gains an advantage.
Where things stand
On 1 December 2025 the UN General Assembly adopted its first resolution on the subject (A/RES/80/23), which “demands that, pending the total elimination of nuclear weapons, human control and oversight is maintained over command, control and communications systems of nuclear weapons.” It passed 118 to 9 with 44 abstentions. The nine against included the United States, Russia, the United Kingdom, France, Israel and North Korea; China, India and Pakistan abstained. No nuclear-armed state voted for it. The resolution is not binding.
Separately, the United States and China stated in November 2024 that decisions on nuclear use should remain under human control — a bilateral political statement, not an agreement.
The case for a binding human-control rule
- The consequences of a nuclear use decision are unique, and delegating any part of it to a system whose reasoning cannot be inspected is an unacceptable risk.
- A norm stated now, before the systems are fielded, shapes their design; one stated afterward has to be retrofitted.
- The 2025 vote shows the norm has broad support among the states that would suffer the consequences without holding the weapons.
The case against — or for caution
- “Human in the loop” is undefined. Every nuclear state already says a human decides; the question is what the human is looking at, how much time the system leaves them, and whether they can meaningfully disagree with an automated recommendation. A rule that does not address those is a slogan.
- Some uses of AI make nuclear command safer — better early-warning discrimination, faster detection of a cyber intrusion — and a blanket rule could forbid them.
- The nuclear-armed states’ objection is partly procedural: they want such norms negotiated among themselves, not imposed by a General Assembly majority. Whether that is a reason or an excuse is itself contested.
What to read
- UN General Assembly resolution 80/23 (1 December 2025), full text.
- The November 2024 U.S.–China statement on human control of nuclear-use decisions.
- SIPRI Yearbook 2026, chapter 8 (the AI-in-NC3 developments of 2025).
- RSIS commentary IP26002 on resolution 80/23 (2026).