Title: The Geometry of Good Unreliability: How AI Can Balance Optimization with Meaningful Flexibility Abstract: Modern transportation optimization pursues an alluring goal: minimize uncertainty, maximize predictability, deliver the optimal route. Yet this framing conflates two fundamentally different types of uncertainty: the frustrating unpredictability that erodes trust, and the generative uncertainty that enables discovery, encounter, and adaptation. In this keynote I will (try to) argue that distinguishing "good" from "bad" unreliability is not just a philosophical exercise but might admit a tractable mathematical perspective and that AI may be uniquely positioned to operationalize this distinction at scale. We will motivate a geometric perspective on mobility optimization, where the feasible region represents not a constraint to be minimized against, but a space of possibilities to be thoughtfully navigated and even explored. Rather than collapsing this space to a single optimal point, AI systems can learn to balance competing objectives: efficiency and serendipity, reliability and exploration, predictability and the productive friction that fosters social connection. Similar to multi-objective optimization, robust planning under uncertainty, and recent advances in reinforcement learning, we outline how AI can serve as a balancing mechanism, by dynamically adjusting tradeoffs between, e.g., tight optimization and deliberate slack based on context, user needs, and system-wide considerations. The result is not optimization or unreliability, but "optimized unreliability". We finish with a few thoughts on the implications for the design of future mobility systems and the broader question of how AI can enhance rather than erode human agency.