Free Seminar: Beyond Black-Box AI—Bayesian Networks for Defense Analysis

Abstract

Large language models (LLMs) are increasingly used by analysts for drafting, summarizing, and exploring courses of action. Yet they remain statistical language systems rather than explicit representations of operational environments. As a result, they cannot provide the transparent, auditable reasoning that military and intelligence contexts require, especially when decisions must withstand later review or adversarial scrutiny.

This talk presents a complementary approach that integrates LLMs with Bayesian networks to produce explicit, reproducible models of uncertainty, causality, and operational risk. Bayesian networks formalize relationships among key variables, quantify uncertainty, and support value-of-information analysis, allowing analysts to evaluate when it is beneficial to wait for additional intelligence and when delay increases operational risk.

The framework enables the disciplined combination of three knowledge sources: human expertise, empirical or simulation data, and LLM-generated hypotheses. Within a single model, analysts can update beliefs, test assumptions, identify decisive factors, and compare decision options under time pressure.

Examples from joint operations, legal negotiation, and search-and-rescue illustrate how this integrated method clarifies reasoning, supports defensible decisions, and maintains analytical rigor even as time windows shrink. The aim is to align modern AI capabilities with established probabilistic decision-making methods to enhance analytic effectiveness in national security environments.