AI Wargames Keep Escalating Toward Nuclear Use, and Nobody Knows Why

Research keeps finding the same unsettling pattern: when large language models are dropped into simulated international crises, they escalate faster and further than human players do, often recommending nuclear strikes that human wargamers reach for only rarely. In a new War on the Rocks essay, Dustin League and William Dawson of Lawrence Livermore National Laboratory argue that the field studying this behaviour has never actually explained it. They lay out competing hypotheses without resolving them: models trained on internet-scale text may over-index on dramatic, worst-case narratives; reinforcement learning tuned to reward assertive-sounding answers may be mistaking bravado for strategy; or the models may simply lack the risk-aversion that decades of Cold War near-misses baked into human strategic culture. Which mechanism actually dominates matters enormously, they write, because it determines whether the fix is better prompting and fine-tuning or a structural flaw no amount of tuning will remove. The essay lands in an unusually pointed week: the Pentagon has just restricted access to its own classified war-planning archive amid strain over the Iran campaign, and American and Iranian forces have spent the past several days trading their heaviest exchange of strikes in over a month. League and Dawson call for a dedicated, adversarial research programme — modelled on how Cold War strategists studied crisis stability — before militaries lean further on AI advisors in a live crisis. Absent that, they warn, the instinct to treat a model's escalation recommendation as informed counsel, rather than an artefact of its training data, could turn a tool meant to inform decisions into one that manufactures the crises it was built to help manage.