The discourse around trustworthy AI has largely focused on properties of individual models (fairness, robustness, and explainability) while overlooking a fundamental reality: AI does not operate in isolation. In practice, AI models are embedded within complex systems that include data pipelines, deployment infrastructure, human decision makers, and governance structures. Failures in these surrounding layers can render even a well-designed model harmful, as illustrated by documented incidents in autonomous vehicles, medical diagnosis, and hiring systems.

In this talk, I argue for a shift from model-centric to system-level trustworthiness, and discuss what this shift means in practice for the design and assessment of dependable autonomous systems. Drawing on a layered framework that spans data trustworthiness, infrastructure dependability, human-system interaction, and governance, I show how established empirical methods (including fault injection, robustness testing, and failure prediction) can be applied to assess and improve trust at the system level. I conclude with open challenges and a call to the dependability community to lead in operationalizing system-level trustworthiness for AI-enabled systems.