Cognitive Resilience and Automation Bias in AI-Augmented Military Cyber Operations and Intelligence Analysis

The accelerating integration of artificial intelligence (AI) into military cyber operations and intelligence analysis has introduced a critical yet under-examined cognitive vulnerability: automation bias, the systematic tendency of human operators to defer uncritically to AI-generated recommendations. This dissertation investigated automation bias in AI-augmented military cyber-intelligence workflows, evaluated debiasing interventions for its mitigation, and examined the systemic consequences of biased AI outputs propagating through military decision chains. A four-phase sequential mixed-methods design was employed, grounded in a pragmatist research philosophy. Phase 1 conducted directed content analysis of 30 government oversight documents and three major cyber incident databases (VERIS, NVD, DCID), revealing a significant policy gap: only 2 of 30 reports explicitly addressed automation bias despite pervasive AI integration across defense enterprises. Phase 2 implemented a 4 × 3 × 2 factorial experiment with 156 participants comparing three debiasing interventions—adversarial questioning, alternative scenario generation, and probabilistic reasoning aids—against a control condition across intelligence disciplines (SIGINT, GEOINT, OSINT) and ambiguity levels. The omnibus main effect of debiasing condition on decision accuracy was significant, F(3, 144) = 10.94, p < .001, η²p = .19. Adversarial questioning emerged as the most effective intervention, producing a 13-percentage-point improvement over control (Cohen’s d = 1.30), with disproportionately greater effects under high-ambiguity conditions (η²p = .33). Phase 3 agent-based modeling of 100 scenarios across five network topologies demonstrated that bias propagation rate was the strongest predictor of organizational escalation (r = .600, p < .001), surpassing AI accuracy level as a risk factor. Phase 4 adversarial validation through 50 red team/blue team exercises confirmed that combined human-AI detection (M = 0.713) consistently outperformed either analyst-only (M = 0.463) or AI-only (M = 0.385) detection across all adversarial attack types. These findings converge in the Cognitive Resilience in Adversarial AI-Augmented Environments (CR-AAAE) framework, a novel integrative model bridging automation bias theory, trust calibration, cognitive load theory, resilience engineering, and cyber escalation dynamics. The results carry direct implications for military training doctrine, AI system design, intelligence community reform, and international norms governing AI in cyber operations.