Generative AI–driven no-code platforms can automatically produce user interfaces from natural language prompts, yet the resulting designs frequently contain accessibility violations, usability inconsistencies, and ethically problematic interaction patterns. Existing evaluation tools focus primarily on isolated accessibility checks and assume complete structural access to the interface, limiting their use for modern multi-modal AI-generated outputs. This thesis presents a framework for evaluating AI-generated user interfaces that unifies accessibility (Web Content Accessibility Guidelines), usability heuristics, and ethical design principles within a single evaluation taxonomy. The framework is modality-aware, adapting its evaluation strategy to different interface representations, including source code, repository artifacts, and rendered interface images with varying levels of structural observability.