Is AI trying to pick a fight? Bias toward escalation plagues national security operations

New research reveals concerning patterns in how artificial intelligence responds to international crises, with foundation models showing a troubling bias toward escalation rather than diplomatic solutions. This discovery comes at a critical time when AI systems are increasingly embedded in national security operations, from ChatGPT Gov’s broad governmental deployment to specialized tools like CamoGPT in defense and StateChat in diplomacy.

The big picture: A comprehensive study by the Futures Lab at the Center for Strategic and International Studies (CSIS) and Scale engineers tested AI foundation models against 400 scenarios and over 66,000 question-and-answer pairs to evaluate their crisis management capabilities.

Key findings: The research uncovered significant algorithmic biases in how AI models approach foreign policy decisions and escalation scenarios.

• Foundation models commonly used by government agencies and citizens demonstrated a consistent preference for escalating crises over pursuing diplomatic solutions.
• The bias varies across different models like ChatGPT, Gemini, and Llama, with more aggressive postures emerging when simulating decision-making for U.S., U.K., or French leaders.

Why this matters: The research suggests AI systems could inadvertently amplify international tensions rather than defuse them, particularly concerning as nations like China develop advanced models like DeepSeek for strategic planning.

Behind the numbers: The study’s extensive testing—covering 400 scenarios and 66,000+ question pairs—provides robust evidence of systematic biases in AI decision-making patterns.

Potential risks: The study identifies several critical concerns about AI’s role in international relations.

• AI agents’ predisposition toward confrontational measures could unintentionally escalate tensions between nations.
• The potential for AI-triggered misunderstandings could lead to conflict spirals.
• Biased AI systems might accelerate arms-race dynamics between competing nations.

The way forward: Researchers recommend specific measures to address these challenges.

• Establish frameworks for ongoing testing and evaluation of AI models in diplomatic scenarios.
• Create multidisciplinary teams combining academic, think tank, industry, and government expertise.
• Develop specialized training to help AI systems better understand diplomatic nuance.
• Maintain robust human oversight in AI-assisted decision-making processes.