NPT Review Conference Specifies 'Maintaining Human Control' in Draft Final Document to Address AI Risks to Nuclear Weapons
⚡ What Happened
A draft final document was released at the NPT Review Conference, calling on all states to maintain "human control over nuclear weapons" in order to reduce the risk of AI and other emerging technologies intervening in nuclear weapons decision-making. This represents the first serious attempt to establish international norms at the intersection of AI and nuclear weapons, and carries critical significance as the military AI arms race accelerates. Full-scale negotiations among states toward adopting the final document will now begin, with agreement among nuclear-weapon states being the central focal point.
The NPT Review Conference is held once every five years; the previous 2022 conference failed to adopt a final document due to Russia's opposition. The explicit inclusion of "human control over nuclear weapons" in this draft signals that international concerns over military applications of AI have reached a tipping point. The US, China, and Russia are all advancing AI integration into their nuclear command and control systems, making the risk of automating early warning and launch decisions a realistic threat. Since the 2023 AI Safety Summit, progress on regulating lethal autonomous weapons systems (LAWS) has been slow, while the intersection of nuclear weapons and AI has rapidly emerged as an "existential risk to humanity." The inclusion in the draft means that at least the conference chair and some major states share this concern, but the document is likely to remain a political statement without legally binding force. Consensus among all five nuclear-weapon states (P5) is required for adoption, which remains the greatest structural barrier.
🔍 What makes this draft particularly noteworthy is the ambiguity of the phrase "human control." Since no nuclear-weapon state discloses the details of its nuclear command and control systems, there remains room for each to expand AI-assisted functions while claiming that "human control is maintained." In fact, the US NC3 (Nuclear Command, Control, and Communications) modernization program and Russia's "Perimeter" system already incorporate AI-like elements. Behind the draft's inclusion lies not only pressure from non-nuclear-weapon states and civil society, but also US and Russian intentions to check China's rapid expansion of military AI investment. Even if the final document is adopted, there will be no verification mechanisms whatsoever, making its effectiveness as a norm extremely limited.
📰 Source: NHK
🧭 Why This Is Moving Now
domain=geopolitics
🔮 Scenarios Ahead
🎯 Incentive Map
| Player | True Incentive | Underlying Vulnerability | Predicted Action |
|---|---|---|---|
| United States | Balancing the maintenance of military AI superiority with checking China. Seeks to lead norm-setting while pursuing language that does not constrain its own NC3 modernization | Fixation on maintaining the credibility of extended nuclear deterrence for allies. Accepting AI norms too readily risks constraining the flexibility of its deterrent | Endorses the principle of "human control" while opposing the specification of definitions or verification measures. Seeks adoption through ambiguous language |
| Russia | Maximizing the coercive value of nuclear weapons in the context of the Ukraine war. Uses the NPT forum as a diplomatic card to criticize NATO nuclear sharing | Structural dependence on nuclear forces to compensate for conventional military inferiority. AI integration is key to efficiency gains and cannot be constrained | Escalates confrontation over NATO nuclear sharing and Ukraine-related language rather than AI-nuclear language, and is willing to accept non-adoption of the final document |
| China | Continuing rapid modernization and expansion of nuclear forces while securing international standing as a "responsible nuclear-weapon state" | Vulnerability to international pressure regarding nuclear force transparency. Faces a dilemma of participating in norm-setting while concealing the extent of AI integration | Offers superficial endorsement of the "human control" principle while strongly opposing any concrete transparency measures or verification mechanisms |
⚠️ Pre-Mortem — Conditions Under Which This Prediction Fails
- The US, China, and Russia make an unexpected compromise, adopting the final document while retaining the AI-nuclear language (a scenario in which the P5's political interests temporarily align)
- The AI-related language in the draft is sufficiently diluted that no nuclear-weapon state has reason to oppose it — a structural possibility of adoption in an effectively harmless form
- The political pressure arising from the precedent of consecutive NPT failures — "we must produce results this time" — may be underestimated (a rebound bias from the previous failure)
Fear-Setting / When this prediction fails
- This probability fails if the US and China reach a bilateral AI-nuclear safety agreement before the conference ends, creating momentum for consensus adoption.
- This probability fails if the final document language on AI is diluted to merely 'noting concerns' — a formulation too weak for any state to oppose — enabling consensus adoption.
- This probability fails if Russia, seeking diplomatic rehabilitation, unexpectedly supports the final document to demonstrate multilateral engagement.
HIT condition: HIT if the 2026 NPT Review Conference fails to adopt a final document containing language on AI and nuclear weapons by consensus
Resolution date: 2026-05-21
