LLM Control in Wargames: Only 9% of Studies Use AI for Both Actions and Adjudication

A scoping review of 223 de-duplicated papers on AI in wargames and strategic simulations, retrieved through May 1, 2026, reveals that only 20 studies (~9%) grant language models open-ended control over both player actions and adjudication. The review, published on arXiv (2509.17192), emphasizes that LLM-based social simulations may conflate distinct roles: deciding what an actor says or does, determining consequences, or both. This distinction is critical in open-ended wargames, where models are valued for handling unusual actions and ambiguous outcomes. The authors argue that researchers must understand the model-control profile—whether the LLM has open-ended control over actions, adjudication, or both—before treating outputs as social simulations. Fidelity in open-ended simulations depends on this control structure. The paper calls for clearer reporting of model roles in future research.