Engineering Trust in AI-Driven Optical Networks: Challenges in Moving from Monitoring to Autonomy

Paper i proceeding, 2026

Artificial intelligence and machine learning (AI/ML) are widely recognized as primary enablers for next-generation optical network automation. However, despite significant algorithmic advances, the deployment of these technologies remains largely confined to monitoring tasks and supervised decisionmaking. In this paper, we explore recent achievements in applying AI/ML to optical networks, highlighting successes in quality-oftransmission estimation, failure management, and sensing. We then analyze the discrepancy between high algorithmic predictive performance and the limited adoption of fully autonomous control loops. Specifically, we investigate the technical factors that inhibit operator trust, including model opacity, lack of generalization to dynamic and evolving conditions, insufficient performance characterization, and limited performance guarantees. The paper concludes by identifying research efforts that would bring us close to transitioning AI/ML from a passive tool into a trusted, autonomous controller.