18 August 2025, Geneva

Situation Overview

As militaries and governments integrate artificial intelligence (AI) into decision support systems, the risk of inadvertent escalation grows. Academic research warns that large language models (LLMs) and other AI systems can misinterpret context, amplify biases, or even generate first-strike rationales when tasked with simulating adversary intentions. This raises profound concerns about crisis stability in both military and diplomatic arenas.

Key Developments

• AI in Defense Planning: The U.S., China, Russia, and NATO allies are all actively exploring AI-enabled wargaming and decision-support tools. These systems are designed to accelerate scenario analysis and strategic options, but their outputs can be opaque and prone to hallucination.

• Research Warnings: Studies highlight that LLMs can adopt aggressive escalation logics, particularly in nuclear or crisis simulations, without clear grounding in human intent. This mirrors Cold War-era fears of automated decision loops, now amplified by AI’s unpredictability.

• Diplomatic Applications: Beyond the battlefield, AI is increasingly used for forecasting negotiations, drafting policy options, or analyzing adversary communications. A misframed model output in this context could encourage brinkmanship or misjudge an opponent’s red lines.

• Feedback Loops: When multiple states adopt AI-driven decision aids, there is a risk of AI-to-AI interaction across borders. This could create unanticipated escalation spirals, especially in high-tension environments.
  
  

Recent Case Studies & Evidence

• Wargame Simulations Escalating to Nuclear Use: A Stanford-led study found multiple LLMs escalating in multi-actor crisis scenarios, occasionally recommending nuclear first-strike rationales. 
  Further reading: arXiv → , ACM FAccT →  ACM →  

• Guardrails Work, But Are Uneven: An August 2025 preprint demonstrated that simple constraint framing significantly reduced escalation, though reproducibility remains a concern.
  Further reading: arXiv → 

• Nuclear Risk Pathways: SIPRI’s 2025 report warns AI use in intelligence, surveillance, and reconnaissance compresses decision timelines, heightening risks of miscalculation. 
  Further reading: SIPRI → 

• Normative Efforts Lagging: UNIDIR’s 2025 review of initiatives such as REAIM highlights a gap between political declarations and enforceable evaluation standards.
  Further reading: UNIDIR → 

• Operational Reality: In Gaza, AI-driven targeting systems were reported to improve efficiency but drew criticism for lowering thresholds on lethal decision-making. While proponents argue these tools accelerate targeting workflows, critics counter that they reduce transparency and increase the risk of overuse.
  Further reading: 972 Mag →, RUSI →  , Lieber Institute →  , Guardian →  , HRW →  

• PLA Adoption: Reports confirm PLA-linked institutes are adapting LLMs for military analysis, raising the likelihood of future AI-to-AI adversarial interactions. 
  Further reading: Reuters → 

• Diplomatic “Peace-Tech” Experiments: Startups are pitching AI digital twins of leaders to forecast negotiations, illustrating both promise and peril. 
  Further reading: The Times →,  Business Insider →
  
  

Risks

• Strategic Stability: Overreliance on AI may erode human judgment in nuclear or crisis decision-making.

• Diplomatic Missteps: AI-generated rationales could harden negotiating positions or normalize brinkmanship.

• Deterrence Misinterpretation: Adversaries may misread AI-recommended maneuvers as deliberate escalation.

ISRS Analysis

AI’s growing role in decision support systems demands urgent governance. Just as nuclear strategy developed protocols to prevent accidental launch, AI requires safeguards to ensure outputs remain advisory, transparent, and bounded by human oversight.

• Short-term: Mandate red-team testing of escalation risks of all AI systems in military/diplomatic contexts.

• Medium-term: Build shared allied evaluation packs (common “escalation score” metrics, scenario stress tests).

• Long-term: Establish a global framework for AI in security contexts, comparable to arms control regimes.
  
  

Conclusion

If left unchecked, AI-driven escalation risks could become the most destabilizing force in international security since the advent of nuclear weapons. Now is the narrow window to build guardrails before these systems are deeply embedded in command and diplomatic infrastructures.