The Missing Layer: Why EdTech Needs Design-Time Generative UI, Not Just Runtime Personalization
This work addresses the problem of accessibility and representation diversity in adaptive educational technology for learners with diverse needs, but the proposal is conceptual and lacks empirical validation.
The paper argues that runtime personalization in EdTech is insufficient for equitable accessibility, proposing a design-time generative UI approach where content is authored as modality-agnostic semantic units and multiple representations are generated and verified by instructors before reaching learners.
The dominant paradigm in using generative UI (GenUI) for adaptive EdTech considers the use of AI as a runtime engine: content is authored once in a fixed form, and AI adapts delivery dynamically based on learner needs, behaviors, or profiles. We argue that this paradigm has an issue: it moves the burden of accessibility and representation diversity onto systems that see learners only after content has already been locked into particular details. For learners who might need audio-first, simplified text, interactive, or low-bandwidth representations, runtime adaptation is too late and too costly to be equitable at scale, and might lead to inaccurate learning content due to the inability to conduct verification at scale. We propose an alternative method: accessibility belongs in the authoring layer. Specifically, we advocate for a card-based GenUI paradigm, in which educational content is encoded as modality-agnostic semantic units, and GenAI produces multiple interface representations, such as interactive, audio, text-simplified, or low-bandwidth, at learning design time to be verified by the instructor before it reaches any learner. This shifts the AI intervention from delivery to creation, embeds Universal Design for Learning principles into the authoring workflow, and removed per-learner inference costs. We situate this idea against recent work on GenUI, multimodal content generation, adaptive authoring, and equitable delivery, and argue that realizing this goal requires closer integration of AI, HCI, and learning sciences than what either of those communities has so far provided.