The Department of War is placing a tremendous amount of faith in AI and Big Data approaches to complicated and wicked problems. However, in the rush to capitalize on AI in national security, what are the risks of incorporating new vulnerabilities, not unlike those seen in the creation and development of cyberspace itself? What assumptions currently pervade military culture about AI and Big Data that are inaccurate and counterproductive, and what are the short-term and long-term implications of these vulnerabilities for the Department of Defense? Specifically, what is the impact on generating reliable training data when answers are needed quickly and without prior curation, and what techniques could be employed to prevent response or confirmation bias from skewing analytical models? Furthermore, what are the limitations of AI and Big Data techniques when applied to sociopolitical exercises like irregular warfare, as opposed to fundamental engineering questions? Ultimately, by understanding these limitations and mitigating data bias, how can the military apply AI tools in a strategic, productive, and quick manner?
-
AU Library Libguide - Military Applications of Artificial Intelligence
-
Brode, Michael C., "Battle Management Automation: Balancing Technological Adoption and Trust with Risk," SAASS thesis, 2025, 95 pgs.
-
This question asks what inaccurate and counterproductive assumptions currently pervade military culture regarding AI and data. Brode answers this by identifying "automation bias"—the dangerous and inaccurate assumption that a computer "system can never be wrong"—as a major cultural flaw. He explains that automated systems suffer from programming "brittleness" because they cannot always adapt to novel or unexpected operational variables. Brode warns that the military's assumption that automated systems can substitute for experienced, critically thinking crews artificially inflates the perceived output quality of the technology, leading to tragic outcomes when the system misidentifies targets.
-
Chapman, Maj. Kyle, "Artificial Intelligence: Fact, Fiction or Marketing?" ACSC elective paper, 2024, 12 pgs.
-
Harding, Emily, Col. Matthew Strohmeyer and Mackenzie Richardson, "From Data to Insight: Making Sense out of Data Collected in the Gray Zone," CSIS brief, AFF paper, 2021, 17 pgs.
-
Holloway, Maj. E. Minnenne and Maj. Bridget K. Pantaleon, "Rethinking Data Protection: AI, Big Data and Privacy in National Security," AF Fellows paper, 2024, 5 pgs.
-
Klare, Capt. Christopher D., "Artificial Intelligence: Air Force Unprepared for 2025 Recommendations," SOS AUAR paper (AI) 2021, 12 pgs.
-
Keith, Andrew J., "Alignment: National Security Objectives in Cold War Computer Simulations," SAASS thesis, 2025, 117 pgs.
-
Keith tackles this by refuting the common assumption within military culture that AI and logic-based computer simulations cannot represent "soft" sociopolitical objectives or irrational human behavior due to technological limits. He notes that while strategists often assume that simulation technology restricts the representation of ambiguous political goals, early models like TEMPER actually actively coded for psychological characteristics, biases, and cultural motivations to successfully simulate international relations. However, he highlights a counterproductive reality where these sociopolitical factors were unilaterally quantified by private contractors based on their own intuition rather than explicit policymaker input. Keith warns that treating sociopolitical modeling as a purely engineering challenge leads to opaque systems where the AI's embedded values and objectives misalign with those of the actual government sponsors and users.
-
Lewczyk, Jonathan A., "Maintaining Human Authority while Utilizing Artificial Intelligence in Commercial Aviation," AF Fellows paper (Georgetown, The Policy Issues of Big Data and Artificial Intelligence), 2023, 9 pgs.
-
Nicholson, Capt. Jonathan, "LLM Use Case," SOS AUAR, 2025.
-
Nicholson answers by critiquing the cultural assumption that commercial AI software is easily adaptable to national security, showing that the commercial sector has a vastly higher risk tolerance because corporate errors do not risk human lives like military failures do. Operationally, he explains that feeding comprehensive DoD data into an LLM creates a critical security risk by centralizing distributed information into a single point of failure, giving adversaries deep insights into military processes and courses of action if compromised. Furthermore, he warns that model biases can manifest as institutional blind spots, while the tendency of deep learning models to output generalized data makes them highly unreliable for rare or unique real-world tactical scenarios.
-
North, Matthew, "Can Machines Have Ethics?" AFGC thesis, 2025, 44 pgs.
-
North warns that AI/ML systems are entirely dependent on the historical data they are trained on, which leaves them highly vulnerable to human bias. For instance, if an AI is trained on historical human conflicts—such as the rapid shift in World War II from condemning the bombing of civilians to actively targeting them in firestorms—the algorithm may learn the wrong ethical lessons. Because AI only seeks the most efficient mathematical path to success, North argues that using flawed or biased human data could create a self-feeding cycle of unethical automated decisions.
-
O'Connor, Maj. Michael, "The Right Stuff for AI: Hard-Won Safety Lessons from the World of Flight Testing," AF Fellows research, published in Breaking Defense, December 1, 2023.