The Ethics of Artificial Intelligence in Military Operations

Modern military operations increasingly rely on deep learning systems for tasks ranging from logistics to lethal targeting. The authors argue that the primary ethical challenge is not merely the question of whether machines should be permitted to use force, but whether the conditions for responsible human judgment can survive when critical decisions are delegated to opaque algorithms. This creates an epistemic crisis: decision-making processes become uninterpretable, evaluation practices are easily gamed, and accountability is fractured across designers, operators, and policymakers.

Rather than attempting to encode ethics directly into algorithmic loss functions—which the authors argue is brittle and prone to failure—they propose a procedural governance framework. This approach shifts the focus from the impossible task of assigning moral responsibility to a machine toward creating institutional infrastructure that ensures human oversight remains meaningful. The framework consists of four main pillars:

• Named Accountability Roles: Clearly defined responsibilities for System Owners, Operational Authorizers, and Users to prevent the diffusion of agency.
• Adversarial Auditing: Using undisclosed, regularly rotated benchmarks to prevent developers from gaming evaluation criteria.
• Tiered Deployment Thresholds: A classification system that scales oversight requirements based on the lethality, reversibility, and speed of the AI system.
• NATO Standardization: A proposal for a standardized evaluation protocol (STANAG) to ensure consistent certification across allied forces.

By analyzing eight operational cases from 1988 to 2025—including the USS Vincennes, Project Maven, and the use of AI in recent conflicts—the authors demonstrate that current governance regimes are insufficient. They show that failures in military AI are typically procedural rather than purely technical. This framework provides a concrete, actionable path for military organizations to integrate AI while preserving the capacity for human operators to understand, interrogate, and override automated recommendations.