AI-Supported Adjudication in Educational Wargames: Integrating LLMs for Strategic-Level Simulations

This paper examines the use of large language models to support adjudication in educational strategic wargames, arguing that AI can function as a constrained analytical layer rather than an autonomous decision-maker. Its central case is "Arctic Hybrid 2026," a February 2026 matrix wargame at the Bundeswehr Command and Staff College, in which eighteen participants organised into six teams managed a fictitious Arctic crisis. The game combined a DIME/PMESII analytical framework with a D20 dice mechanic, and adjudication was centred on NotebookLM instances grounded in curated source corpora.

The authors' core claim is that grounding contained hallucination risk to acceptable levels, allowing AI to scale complexity and reduce facilitator burden. A five-stage workflow was used in which AI-driven adjudication and non-player controllers were triangulated across three human White Cell reviewers. The AI non-player characters tended to act as outcome harmonisers, smoothing extreme or irrational behaviour, and the study found no significant difference between AI and human adjudicators when AI was deployed in this constrained role. Participants reported self-assessed gains in AI and security-policy competence, though these findings rest on self-reported data. Notable challenges included infrastructure inequality among teams and a contingency-encoding loophole in move submission that affected validity.

The authors conclude that the framework is transferable to other AI-enhanced strategic simulations but not yet institutionalisable. Generalisability is limited by the single case study and reliance on self-reported outcomes. The accompanying literature reflects a dual picture: simulation-based learning and serious games are associated with positive cognitive effects, while frontier models operating in simulated nuclear crises raise escalation-related concerns that reinforce the case for tightly constrained AI roles.