Skip to content
OGERIA — Observatory of Global Evidence on Risks in AISynthesis report · 2026 ed.
Updated 2 Oct 2026

Chapter 01 · Risk map

Military and autonomous weapons

AI in nuclear command and control

AI does not press the button: it compresses decision time and presents opaque recommendations to whoever does.

Severity
Existential
Horizon
3–10 years
Evidence
Projected
Consensus
Low

The popular image —an AI with its finger on the button— is the least plausible version of this risk, and concentrating on it means losing sight of the one that does have a mechanism. The central contribution of SIPRI’s June 2025 analysis is to relocate the problem: AI integration can influence nuclear escalation even when it happens outside nuclear weapon systems. Non-nuclear applications compress decision timelines; the opaque recommendations of a decision-support system bias the decision-maker towards acting; and autonomy in a system with counterforce potential threatens the integrity of deterrent forces [972]Impact of Military Artificial Intelligence on Nuclear Escalation RiskChernavskikh, Vladislav; Palayer, Jules · 2025 · reportView source ↗Accessed on 9 September 2026. AI operates on inadvertent and accidental escalation, not on deliberate escalation.

On the declared-control side there are two texts and neither binds. The 2022 Nuclear Posture Review states that in all cases the United States will maintain a human in the loop for critical actions [334]2022 Nuclear Posture ReviewU.S. Department of Defense · 2022 · official documentView source ↗Accessed on 9 September 2026. In Lima, on 16 November 2024, Biden and Xi affirmed the need to maintain human control over the decision to use nuclear weapons [1097]Readout of President Joe Biden's Meeting with President Xi Jinping of the People's Republic of ChinaThe White House · 2024 · official documentView source ↗Accessed on 9 September 2026: one paragraph of a presidential statement, with no definition of what counts as human control and no verification mechanism.

What this does not demonstrate. There is no observation of AI-induced nuclear escalation, and the only adjacent empirical evidence is from the lab and without real weapons: Rivera and co-authors find that five commercial models, in simulated war games, show escalation patterns that are hard to predict and in rare cases reach nuclear deployment [916]Escalation Risks from Language Models in Military and Diplomatic Decision-MakingRivera, Juan-Pablo; Mukobi, Gabriel; Reuel, Anka et al. · 2024 · paperView source ↗Accessed on 9 September 2026. A text simulator is not a command system. The existentialExistential riskHarm that would wipe out humanity or permanently close off its chance of recovering. A catastrophic risk is enormous, but it can be recovered from.For exampleA fire that burns down your house is catastrophic. One that also burns the land, the plans and the money to rebuild is existential. severity of this entry comes from the weapon, not from the strength of the evidence about AI: it is the case where the observatory assigns the highest severity with one of the weakest empirical backings.

Chain of materialisation

  1. PreconditionObserved

    Doctrine declares human control and has no verification mechanism

    The 2022 Nuclear Posture Review states that in all cases the United States will maintain a human in the loop for critical actions. In Lima, November 2024, Biden and Xi affirmed the need to maintain human control over the decision to use nuclear weapons. It is a paragraph in a presidential communiqué with no definition of human control, no verification and no legal obligation.

    Precedents: Biden and Xi affirm human control over nuclear weapons employment

  2. TriggerLab

    In simulation, models escalate in hard-to-predict ways

    Rivera and co-authors evaluated five commercial models in simulated wargames: they show hard-to-predict escalation forms and patterns, tend to develop arms-race dynamics and in rare cases reach nuclear deployment. It is language-agent behaviour in a simulator, not a weapon with an agenda.

    Observed and demonstrated evidence ends here. What follows is projection.

  3. CascadeProjected

    The risk enters from outside the nuclear system

    SIPRI relocates the problem: AI integration can influence nuclear escalation even when it occurs outside nuclear weapon systems. Non-nuclear applications compress decision timelines, opaque recommendations bias decision-makers towards acting, and autonomy in counterforce-capable systems threatens deterrence integrity.

  4. ImpactSpeculative

    Inadvertent or accidental escalation to nuclear use

    Nobody has observed AI-induced nuclear escalation. It is the only risk in the observatory classified as existential in severity, and that classification comes from the weapon, not from evidence about AI.

See on the map →Report a mistake in this entry →

Sources

  1. [334] 2022 Nuclear Posture Review · Departamento de Defensa de Estados Unidos 2022
  2. [1097] Readout of President Joe Biden's Meeting with President Xi Jinping of the People's Republic of China · Casa Blanca de Estados Unidos 2024
  3. [972] Impact of Military Artificial Intelligence on Nuclear Escalation Risk · Stockholm International Peace Research Institute 2025
  4. [916] Escalation Risks from Language Models in Military and Diplomatic Decision-Making · Rivera, Juan-Pablo 2024
  5. [290] Pentagon-Anthropic Dispute over Autonomous Weapon Systems: Potential Issues for Congress · Congressional Research Service 2026 archived copy only
  6. [444] Elements of Deterrence: Strategy, Technology, and Complexity in Global Politics · Oxford University Press 2024 verified through Crossref
  7. [507] An Overview of Catastrophic AI Risks · Center for AI Safety 2023

Ask OGERIA

It answers only with what the observatory publishes and can be wrong: check the entries it cites. Your questions are sent to an AI model, so don't write personal data. More in the privacy policy.

Up to 500 characters.

Support OGERIA on Ko-fi

The payment is processed by Ko-fi, not by this site. Open on ko-fi.com