I’m Sorry, Dave. I’m Afraid I Can’t De-Escalate: On (AI) Wargaming and Nuclear War

Key Takeaways

• AI models escalated in 95% of simulated nuclear games
• Study highlights AI bias toward coercive strategies from training data
• Wargames primarily capture human judgment, not machine behavior
• AI excels at scenario generation, red‑team assistance, and analysis
• Adding de‑escalation content can temper AI escalation tendencies

Pulse Analysis

The surge of AI‑driven wargaming experiments has captured headlines, especially after a pre‑print from King’s College London showed that three leading large language models repeatedly chose nuclear escalation in simulated crises. In 95 percent of 21 match‑ups, at least one side signaled nuclear use, and tactical strikes followed in the same proportion. These stark numbers underscore a deeper issue: the models are trained on a corpus rich in deterrence theory and coercive strategy, which skews their reasoning toward aggressive outcomes. Understanding this bias is crucial for anyone assessing AI’s impact on strategic stability.

While the data are alarming, the core purpose of wargames is to surface human decision‑making under uncertainty, not to predict machine behavior. Traditional wargames place experienced analysts and military officers at the table, extracting insights about risk tolerance, institutional culture, and cognitive heuristics that no AI can fully replicate. Treating AI‑vs‑AI simulations as a proxy for human conflict therefore constitutes a category error. The real value of AI lies in augmenting the wargame lifecycle—automating scenario creation, offering red‑team perspectives during play, and distilling massive debrief transcripts into actionable themes—while keeping human judgment at the analytical core.

Policymakers and strategists should therefore channel AI resources toward these supportive roles and enrich training data with de‑escalation narratives, defeat tolerance, and off‑ramps. By expanding the corpus beyond classic deterrence literature, future models can better reflect restraint pathways, reducing the risk of false escalation signals. This balanced approach ensures AI becomes a force multiplier for nuclear risk analysis rather than an inadvertent catalyst for heightened tensions.