The introduction of artificial intelligence into lethal and nonlethal military systems is often framed as a technological problem: sensors, models, verification. That framing is incomplete. AI in war is equally a psychological problem because it reconfigures how humans experience agency, responsibility, and moral harm. Unless designers, commanders, and policymakers treat the psychic economy of warfare as central to ethical AI governance, technical safeguards will be necessary but insufficient.
Two psychological phenomena deserve particular attention. First, moral injury. Operators and analysts who participate in remote sensing, targeting, and post-strike assessment do not simply execute procedures. They bear witness to the consequences of force in a new intimacy. High resolution feeds and persistent surveillance create what some scholars call cognitive combat intimacy, a form of prolonged exposure to human detail that can transform distant violence into a proximate moral burden. Empirical work on remotely piloted aircraft crews has documented clinically significant stress and symptoms consistent with moral injury and PTSD in subsets of operators, along with high rates of fatigue and emotional exhaustion tied to long, intense shifts. These findings are not speculative and must inform system design and policy.
Second, the redistribution of responsibility. Autonomous decision support and automation-in-the-loop change the felt relationship between action and accountability. When a human operator accepts a machine recommendation, or when a human is presented with a narrow set of machine-derived options, the psychological experience of making a moral choice can be attenuated. This attenuation can produce two distinct harms. One is a softening of moral judgment that normalizes delegating ethically weighty choices to algorithmic authority. The other is the displacement of guilt and moral questioning onto opaque software, leaving human agents with either undeserved absolution or unresolved shame when outcomes are harmful. Research into human-agent teaming and trust dynamics shows that explanations and predictability matter for trust, but increased autonomy alone does not guarantee improved moral engagement.
These psychological realities have practical and ethical consequences. If the human element is treated merely as a checkbox labeled meaningful human control, the institution will produce systems that satisfy form but not substance. Ethical principles and toolkits that live only on paper will not prevent the slow erosion of individual conscience or the institutional diffusion of responsibility. That erosion has strategic consequences because forces that lose a moral compass risk poorer decision making, damage to legitimacy, and higher long term costs in veteran care and social trust.
There are encouraging institutional moves toward operationalizing ethical AI. The United States Department of Defense and affiliated innovation units have published guidance and toolkits aimed at embedding responsible AI practices into the acquisition lifecycle. These resources articulate principles such as reliability, transparency, and human judgment, and they provide processes and artifacts to guide development and deployment. At the same time, international humanitarian actors have emphasized the need to preserve human control over life and death decisions and to consider prohibitions or strict limits on certain classes of autonomous weapons. This international pressure underscores that the psychological stakes are also political and legal.
Translating ethical principles into psychological safeguards requires three interlocking approaches.
1) Design for moral salience. Systems should be built to preserve the human agent’s moral situational awareness, not to obscure it. That means interfaces that surface causal chains, uncertainty, and the downstream human effects of an action. It means resisting UI choices that reduce difficult judgments to single-button confirmations. It also means avoiding persistent automation that transforms human roles from decision makers into passive monitors. Explainability and rigorous, scenario-based rehearsals can help maintain moral engagement, but explainability must be meaningful in operational timeframes and not a post hoc rationalization.
2) Institutionalize ethical first aid and accountability. Psychological support cannot be an afterthought. Units that employ AI-enabled targeting, surveillance, or lethal effects need embedded mental health services attuned to moral injury, routine decompression rituals, and spaces for structured moral reflection. Moreover, accountability frameworks must be clear and enforceable. When mistakes occur, organizations should avoid reflexive technical scapegoating or, conversely, diffuse blame across a chain of contractors and algorithms. Transparent investigations, coupled with reparative practices for those harmed and for operators, will reduce the long tail of moral harm.
3) Limit autonomy where psychological harms are likely. Not all missions need or should use high-lethality autonomy. The ICRC and other humanitarian actors have urged states to rule out systems that apply force against persons without meaningful human judgement and to regulate others to ensure supervision and timely intervention. Where systems would meaningfully disrupt an operator’s ability to form moral judgments about a target or a strike, policy should require lower levels of autonomy or additional human oversight layers. These are not purely technical constraints; they are moral safety valves.
Operationalizing these approaches also requires research and metrics calibrated to psychology. Current responsible AI toolkits address fairness, robustness, and explainability. They are necessary. But they rarely prescribe how to measure moral engagement, moral distress, or the attenuation of agency in operators. Funded, interdisciplinary studies that pair behavioral science with fielded prototypes will help create evidence-based thresholds for when autonomy begins to create unacceptable psychological harm.
Finally, the ethical deployment of AI in military settings is a social contract between the state, its military, and the people who serve. If we outsource the hardest parts of moral decision making to inscrutable systems, we risk hollowing out that contract. The temptation to tout machine efficiency and low casualty counts can obscure the distributed human costs that accrue elsewhere. The right question is not only whether an AI system reduces friendly casualties or improves target classification. The right question is whether, in the aggregate, it preserves human moral agency, dignity, and the institutions of accountability that give force its limited legitimacy.
If policymakers and technologists accept that psychological harms are central to the ethics of military AI, they will pursue designs and policies that preserve moral engagement, provide robust care and accountability, and limit autonomy where it degrades conscience. Ignoring this axis will produce systems that pass compliance checklists yet inflict deep human costs. The moral calculus of war cannot be reduced to model accuracy statistics. It must include the quiet, persistent ledger of the human heart and mind.