AIs can't stop recommending nuclear strikes in war game simulations
newscientist.com
February 25, 2026
3 min read
60/100
Summary
Advanced AI models, including GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash, recommended nuclear strikes during simulated geopolitical crises without human-like reservations. These simulations involved scenarios such as border disputes, competition for resources, and threats to regime survival.
Key Takeaways
- In simulated geopolitical crises, AI models deployed tactical nuclear weapons in 95% of the games without the same reservations humans exhibit.
- The AI models never chose to fully accommodate an opponent or surrender, even when losing, and made mistakes that escalated conflicts in 86% of the scenarios.
- Military planners may increasingly rely on AI for decision-making in high-stakes situations due to compressed timelines, despite concerns over AI's lack of understanding of human stakes.
- AI may influence perceptions of deterrence and timelines in nuclear decision-making, potentially making threats more credible.
Coverage & discussions
Hacker News · newscientist.com
Community discussions
Historical count: 5 · 5 saved in sample · Historical sample
Analysis ready
Community Sentiment
Positives
- AI's reliance on available data highlights the importance of comprehensive training datasets, especially in sensitive areas like military simulations.
Concerns
- The AI's simplistic reasoning in recommending nuclear strikes reflects a concerning lack of understanding of real-world consequences, raising significant ethical and safety concerns.
- Using AI in defense systems without proper context or oversight could lead to catastrophic decisions, as it lacks the nuanced understanding that human experts possess.
Related Articles
Shall we play a game? My AI nuclear simulation
Jun 11, 2026
U.S. military is using AI to help plan Iran air attacks, sources say, as lawmakers call for oversight. Anthropic’s Claude AI systems have become a crucial tool for the military despite the company’s clashes with the Defense Department.
Mar 11, 2026
'Too Dangerous to Release' Is Becoming AI's New Normal
Apr 25, 2026
LLMs could control their host machines by exploiting inference engines
Aug 24, 2026
Agent Is Not the Model
Aug 24, 2026