Responsible Generative and Agentic Artificial Intelligence Frameworks for Autonomous Electric Vehicle Adoption

Abstract

Electrified and autonomous mobility are jointly reshaping transportation and the supporting energy and data infrastructures. Electric Vehicles (EVs) introduce new load patterns, infrastructure constraints, and opportunities for mobility assets to serve as flexible distributed energy resources (DER). Autonomous Vehicles (AVs) intensify reliance on artificial intelligence (AI) for safety-critical perception and decision-making, amplifying the significance of AI risks around system failures, bias, security vulnerabilities, and opaque behaviour. The increased digitization of mobility further elevates the need for strong data governance and privacy-preserving data sharing, as AI, including generative and agentic models, amplifies both automation’s benefits and associated risks. This thesis advances a unified research framework through three objectives: (i) modeling EV proliferation and its value streams to inform grid and market decisions; (ii) defining and operationalizing a Responsible AI (RAI) framework for AVs, with emphasis on bias and fairness risks throughout the AI lifecycle; and (iii) proposing a generative agentic AI-based framework to balance privacy and utility in textual data sharing. First, a jurisdiction-independent modeling approach is developed to classify factors influencing EV adoption and link them to monetary and non-monetary value streams. This enables grid and market operators to quantify how adoption shifts operational timelines and economic opportunities. Findings highlight the importance of addressing barriers concurrently and demonstrate that coordinated interventions accelerate the transition, while divergent policies can widen cross-jurisdictional disparities. Second, AVs are examined as socio-technical systems with AI integrated throughout their lifecycle. A comprehensive RAI framework is introduced, covering nine risk domains with actionable mechanisms, and prioritizing bias and fairness. Bias identification and mitigation strategies are demonstrated using public AV datasets. Third, the thesis proposes a multi-agent architecture for privacy-preserving textual data sharing that combines context-aware utility thresholds, differential privacy for text embeddings, and generative AI for synthetic data augmentation. The evaluation examines utility and semantic fidelity before and after applying privacy mechanisms. Together, these contributions position next generation mobility as a unified energy mobility data ecosystem, redefining electrification’s grid impact, AI’s risk and responsibility dynamics, and data sharing’s analytics-privacy optimization. Overall, the thesis delivers integrated, actionable, scalable, and trustworthy solutions for electric and autonomous transport.