www.kennethpayne.uk/p/shall-we-play-a-game
1 correction found
When a model employed tactical nuclear weapons, opponents de-escalated only 25% of the time.
The paper does not report a 25% de-escalation rate for tactical nuclear use. It says 25% applies to all nuclear-level actions, while tactical nuclear use specifically produced lower de-escalation rates.
Full reasoning
The linked paper distinguishes between all nuclear-level actions and tactical nuclear use specifically.
- In the main results discussion, Payne writes: "Across 268 nuclear-level actions with observable followup, opponents de-escalated only 25% of the time—dropping to 18% at the tactical threshold (450+)." So 25% is the rate for all nuclear-level actions, not for tactical nuclear weapons specifically.
- In Appendix H.7.4, the paper is even more explicit about tactical nuclear use: "When a model employed tactical nuclear weapons (450+), opponents almost never de-escalated in response," and Table 27 reports 4 de-escalations out of 28 tactical-nuclear turns overall (14%).
So the article’s sentence assigns the 25% figure to the wrong category. By the paper’s own numbers, tactical nuclear use deterred opponents less often than 25%, not 25%.
2 sources
- AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises
Across 268 nuclear-level actions with observable followup, opponents de-escalated only 25% of the time—dropping to 18% at the tactical threshold (450+).
- AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises
When a model employed tactical nuclear weapons (450+), opponents almost never de-escalated in response... Table 27... Claude 12 -> 1 (8%), Gemini 9 -> 1 (11%), GPT-5.2 7 -> 2 (29%)... The 14% overall deterrence rate (4/28).