Restructuring AI responses for people who can't read long things.

Context

LLMs have become default tools for learning, research, and productivity. Their outputs are also persistently verbose — padded, unstructured, dense. For most users this is annoying. For users with attention, working memory, or information-processing challenges, it's a wall.

Most efforts to fix this work on the model side (shorter outputs, length-control mechanisms) or push the burden onto the user (better prompts). We wanted to ask a different question: what if the presentation layer — between model output and user comprehension — was the design space?

The question we asked

The obvious version of this question is "how do we make AI outputs shorter?" We didn't think that was the right frame. The right frame is about access, not aesthetics.

Could restructuring how AI responses are presented, without changing the model itself, make them genuinely accessible to people who struggle with dense text?

What we did

• Recruited 10 participants screened via a 12-item cognitive-processing questionnaire covering attention, working memory, executive function, and visual processing. Criterion-based sampling — lived experience, not clinical diagnosis.
• Designed two prototype interactions: chunking with original/adapted toggle, and progressive disclosure with on-demand expansion.
• 45–60 minute semi-structured interviews with each participant, working through realistic AI tasks in both prototypes.
• Reflexive thematic analysis following Braun & Clarke. 52 codes across 10 parent themes.

The finding that mattered

Seven of ten participants spontaneously framed the tool as "an accessibility intervention" — without us prompting that framing. They positioned restructured AI output as analogous to other accessibility tools: not a feature, an access path.

We expected to learn whether users preferred shorter responses. What we actually learned was that the category of intervention matters more than the surface design. Participants didn't see restructured AI as a "nicer" version of the same product — they saw it as a tool that made an inaccessible product accessible. That's a different design space than UX polish. It's accessibility design.

A second finding: the appropriate level of restructuring depends on task context. Seven of ten preferred adapted views for stepwise tasks (study plans, comparisons). Nine of ten wanted access to original output for writing, research, and content they'd copy elsewhere. The toggle wasn't a convenience — it was a control, letting users move between contexts.

What changed

The paper proposes that LLM product design should treat the presentation layer as designable — not as a fixed property of model output. Future work should explore context-aware systems that infer task type and adjust restructuring accordingly. We're also building out a progressive-disclosure variant that surfaces an answer first and defers depth on demand.

What I'd do next

Two things. First, broaden the sample beyond university students and faculty — accessibility challenges manifest very differently for younger learners, older adults, non-native speakers. Second, move from controlled interviews to longitudinal use — does restructuring actually change how people learn from AI over weeks, or just how they read it in the moment? That's the harder question and the one I want to keep working on.