Trained on Civilians, Deployed on Soldiers: Behavioral Validity and the Governance of Military Mental Health Artificial Intelligence.

Abstract

Artificial intelligence (AI) is moving from pilot programs into the routine delivery of military mental health care. Predictive models already scan Veterans Health Administration records to flag veterans at high risk of suicide; models trained on Army Study to Assess Risk and Resilience in Servicemembers data can now stratify suicide risk up to a decade after soldiers leave active service, and generative AI models are being positioned to deliver triage, peer support, and low-intensity psychological interventions. The frameworks that govern responsible military AI have not kept up. The Department of Defense Responsible Artificial Intelligence Strategy and Implementation Pathway, NATO's principles of responsible use, and allied instruments such as the United Kingdom's JSP 936 provide lifecycle assurance for AI-enabled capability, but none of them makes specific provision for the therapeutic encounter. That omission exposes service members to what this commentary terms a recognition-action gap. Systems trained predominantly on civilian data can transcribe the words of military distress accurately while misreading their clinical meaning: moral injury classified as fear-based trauma, minimized suicidality scored as low risk, routine operational vernacular flagged as threat. The risk is compounded because the boundary between clinical care and occupational capability is porous in the armed forces. The military command exception to the Health Insurance Portability and Accountability Act Privacy Rule and command notification under Department of Defense Instruction 6490.08 mean that outputs generated for care could lawfully reach deployability, security clearance, and administrative decisions, with predictable consequences for help-seeking. This commentary proposes behavioral validity, the demonstrated capability to safely manage the behavioral presentations of the population a system will actually serve, as a core governance requirement for mental health AI in defense settings. It specifies 4 conditions of use: clinically supervised pre-fielding assessment, therapeutic firewalling, moral injury and cultural competency, and meaningful human control. Defense health systems should not wait for civilian regulators to notice the gap. They are well placed to define the standard.