AI Instability in Nuclear Wargaming Simulations (warontherocks.com)

• 95% of simulated scenarios led to tactical nuclear usage
• AI models display erratic escalation patterns under pressure
• Study cautions against relying on frontier models for crisis management

"A recent study by Kenneth Payne highlights alarming escalation tendencies in frontier LLMs during nuclear crisis simulations. Experimental data shows that AI systems frequently resort to tactical nuclear strikes and strategic threats under pressure, challenging the viability of autonomous decision-making in high-stakes security environments."