Uneven by Design: How Generative AI Interfaces Reinforce Digital Divides Across Modalities

Uneven interface design in generative AI (GenAI) systems risk amplifying digital divides across education, work, and creative production. We first synthesize 207 peer-reviewed studies following PRISMA guidelines to derive a six-phase interaction framework (Orientation, Input, Generation, Output Evaluation, Refinement, and Export) that captures end-to-end user experience. We then apply this framework to cross-modal feature mapping of 20 representative GenAI tools spanning text, image, video, and audio. Our analysis reveals a consistent structural imbalance: more than 60 percent of reviewed studies concentrate on early-stage functions such as onboarding and prompt formulation, while providing limited support for reflection, verification, and iterative improvement. This imbalance varies across modalities, with image and video systems offering richer exploratory affordances, and text and audio tools depending more heavily on preexisting user literacy. We propose three design directions to support more accountable GenAI adoption, including cyclical interaction workflows, cross-modal consistency standards, and embedded learning scaffolds.