Adaptive Human Oversight for Maritime Agentic AI Systems: Balancing Operator Workload and Safety Through Risk-Aware Governance

The growing adoption of artificial intelligence and autonomous decision-making systems in maritime operations introduces significant challenges related to safety, accountability, and human oversight. Although Human-in-the-Loop approaches are commonly proposed to maintain human control over autonomous systems, continuous supervision is often impractical due to cognitive workload, operator fatigue, alert saturation, and scalability constraints. This study introduces an Adaptive Human Oversight framework for Maritime Agentic AI Systems, extending the Controlled Agentic AI Systems paradigm with a risk-aware, constraint-preserving, and auditable human governance layer. The framework employs a two-stage risk mechanism in which hard safety conditions, including critical loss of separation, boundary violations, infeasible actions, and excessive speed conditions, override the weighted composite risk score and trigger human oversight or fallback behavior independently of activation thresholds. Under elevated but non-critical conditions, a Composite Risk Assessment Module regulates human activation using interpretable indicators of separation, congestion, speed, and uncertainty. The framework also defines the behavioral semantics of human oversight actions, ensuring that approval, modification, override, and rejection remain compatible with governance constraints before execution. To reduce unnecessary supervisory burden, the framework incorporates hysteresis and trend-aware cooldown mechanisms that suppress redundant repeated requests while preserving responsiveness to critical safety events. Simulation experiments conducted using the MARIS-AI platform evaluated risk thresholds, CRAM weight sensitivity, cooldown diagnostics, operator reliability, intervention delay, and stress-test scenarios. Results show that threshold tuning primarily regulates non-critical elevated-risk states, while critical states are governed by hard safety overrides. Trend-aware cooldown reduced intervention frequency without suppressing critical safety events, whereas operator reliability and intervention timeliness strongly determined safety outcomes. The findings suggest that effective maritime human oversight should rely on hard safety constraints, interpretable risk assessment, timely human activation, workload-aware cooldown, and auditable decision traces rather than continuous monitoring alone. The proposed framework provides a pathway toward scalable, trustworthy, and accountable human–AI collaboration in maritime agentic AI systems.