“Major powers such as the United States, China, and Russia are experimenting with integrating AI DSS into their command structures, thereby decreasing their sensor-to-shooter timelines,” the researchers write — a sentence that frames both the promise and the peril behind a simple finding: 87.7 percent of West Point cadets saw strong beneficial applications for AI, compared with 72.5 percent of the public.
Cadets’ AI knowledge and performance versus the public
Researchers from Georgetown University, the University of Pennsylvania, and the U.S. Military Academy published a paper last week comparing 236 West Point cadets to a demographically similar sample of 702 members of the public. The study measured two related phenomena: automation bias — people’s tendency to over‑rely on automation — and algorithm aversion, the inclination to “prematurely distrust automated outputs in ways that increase the risk of accidents or mistakes.”
Across several measures, cadets outperformed the public. The paper reports that cadets’ “AI knowledge scores” were nearly twice as high as those of the general public, and cadets were less than half as likely to commit an automation bias error — meaning they were substantially less likely to accept a mistaken output from an AI decision-support system (DSS).
How cadets checked AI: confidence indicators and calibrated expectations
The study found cadets were more likely than average users to consult a tool’s own confidence indicators when assessing AI output — a behavior the researchers linked to efforts at the U.S. Military Academy to develop “justified confidence.” In the authors’ words, this training seeks to ensure cadets’ “expectations of an AI system’s accuracy match the reality of the accuracy of the system.”
Those behaviors map to a wider pattern: cadets expressed both greater enthusiasm for AI’s potential and a higher level of concern about its risks. About twice as many cadets expressed worry about the dangerous consequences of AI as did members of the public, yet cadets were much less likely to describe AI as “sinister.” That combination — elevated vigilance paired with optimism — is central to the paper’s argument that calibration, not blanket trust or distrust, is attainable through training.
Decision‑support systems and command structures
The authors link these psychological tendencies directly to military practice. They note that “Major powers such as the United States, China, and Russia are experimenting with integrating AI DSS into their command structures, thereby decreasing their sensor‑to‑shooter timelines.” That integration, the paper warns, could lead officers to delegate decision‑making authority to AI DSS if personnel do not maintain appropriate skepticism and calibration.
The arrival of large language models such as ChatGPT, Claude, and Gemini prompted West Point to name a special academic focus for the 2024‑25 school year: “The Human and the Machine: Leadership on the Emerging Battlefield.” The study treats these curricular efforts as part of the explanation for cadets’ comparatively higher knowledge and lower automation bias.
What this means for the U.S. military, the public, and technologists and leaders
- The U.S. military: The paper suggests training can produce “justified confidence” in AI users; that path is illustrated by cadets’ higher knowledge scores and lower automation bias. UPenn’s Michael C. Horowitz told Defense One the study “shows what some of the opportunities might look like for further training within the U.S. military at least.” He added such training “could be effective,” while noting cadet training “is probably not representative of the average military right now, but it shows the path forward, at least for the U.S. military.”
- The U.S. public: Because cadets demonstrate that instruction changes behavior and attitudes, the study raises the possibility that similar educational efforts might reduce the public’s tendency either to over‑trust or to prematurely distrust AI — an idea the paper links to broader national‑security concerns about declining public trust in AI compared with populations in China.
- Technologists and leaders: For designers of AI DSS and policymakers, the empirical link between training and calibrated use suggests that usability features like confidence indicators, and programs that teach how to interpret them, could materially affect operational risk when AI is placed into time‑sensitive decision chains.
Conclusion: training as a concrete lever
The paper offers a concrete, evidence‑based argument: with instruction, people can learn to be neither credulous nor reflexively fearful of AI. West Point’s special curriculum and the study’s findings point to training as a practical lever for reducing automation bias while preserving, even increasing, productive enthusiasm — a combination the authors and UPenn’s Michael C. Horowitz both identify as a pathway forward. The remaining question the study leaves on the table is whether that path scales beyond a military academy to the force at large and to the public — a test the authors imply will be important as AI DSS move deeper into command structures.
