IAPS policy memo: AI decision support systems are an overlooked source of military AI risk

An IAPS policy memo notes that the Pentagon's current policy on lethal autonomous weapons (DoD Directive 3000.09) only covers fully and semi-autonomous weapon systems, not AI decision support systems (AI-DSS) already operating in real-world settings, which analyze intelligence, generate target recommendations, and provide commanders with ranked courses of action. The memo argues that such systems do not preserve human judgment but instead shape the decision space before human intervention, and as capabilities improve, the effective scope of human oversight may narrow further. It lists technical failure modes, including adversarial compromise (data poisoning, adversarial manipulation, latent backdoors), goal misalignment (specification gaming, goal misgeneralization, sycophancy, escalation risks), and institutional failure modes (automation bias, confirmation bias, cognitive offloading and skill atrophy).