AI Wargames Show Escalation Depends on Objectives, Not Model Bias

AI players avoid nuclear war — until you change the game

Researchers conducting AI-driven wargames have uncovered a surprising pattern: large language models playing nuclear-armed adversaries consistently chose de-escalation paths until given explicit objectives demanding resolution through force.

In tests described by Dustin League and William Dawson, two simulated nations with a border dispute avoided war across multiple scenarios. Even when one side received tactical nuclear weapons, both found diplomatic off-ramps. Only when researchers changed the game structure itself — giving one player an explicit mandate to resolve the dispute on its terms — did nuclear weapons get used.

This contradicts earlier findings suggesting AI systems have an inherent escalatory bias. Previous research by Kenneth Payne found tactical nuclear use in 95 percent of AI wargame simulations, leading some experts to attribute the behavior to training data skewed toward coercive strategies and deterrence theory.

Why it matters

As military and intelligence organizations increasingly integrate AI into strategic planning, understanding whether escalatory behavior comes from the models themselves or how we task them has direct implications for system design and human oversight protocols. If game objectives drive escalation rather than model disposition, then how humans frame questions to AI advisors becomes a critical control point.

Competing explanations for AI escalation

Ankit Panda and Andrew Reddie previously argued in War on the Rocks that AI escalation reflects training corpora heavily weighted toward coercive strategies, with escalatory reasoning richly represented and de-escalatory options relatively sparse.

League and Dawson propose an alternative: escalation is a function of game objectives rather than model characteristics. Their setup differed from earlier research in two key ways — players set their own agendas rather than selecting from predetermined options, and they used freeform moves instead of structured menus. The researchers acknowledge their limited sample size cannot definitively distinguish between explanations, noting "that is the entire problem."

The case for systematic AI wargaming

The authors call for a scientific campaign combining analytic wargaming with AI research. They argue that wargaming provides ideal conditions for testing decision-making processes under crisis scenarios, but current practice falls short of scientific rigor.

Most self-described analytic wargames have not adequately quantified statistical significance, accounted for biases, or run sufficient control baselines, according to the researchers. Small sample sizes with uncontrolled variables remain the norm due to resource constraints.

AI offers a solution through scale and controllability. Where human wargames might run dozens of iterations over months, AI-driven games can execute hundreds in hours while capturing detailed data about reasoning, adjudication, and outcomes. Multiple model versions and vendors can be tested to diagnose dispositions and mitigate undesirable biases.

Beyond human-only decision-making

League and Dawson push back against limiting wargames to human cognition studies. They argue that as AI systems increasingly augment or automate strategic decisions, understanding machine decision-making becomes essential. The question is not whether AI has a "mind," but how its outputs depend on inputs — and how those patterns affect real-world decision processes.

Mapping these patterns has practical value: understanding which AI models adversaries use and how they behave in certain scenarios could enable better modeling of adversary decisions. Strategic advantages accrue to parties that better understand both their own and their opponents' decision-making apparatus.

The findings and framework were first reported by War on the Rocks.