Explainable AI for Blind and Low-Vision Users: A Guide for the Agentic Era—Trust, Input Modality, and “Why It Decided That”

Why Read This from an HCI Perspective

This article frames XAI less as a problem of producing good explanations and more as a question of how users can verify and intervene in AI behavior. In particular, it explains why visual explanations become a barrier for blind and low-vision (BLV) users, and why conversational explanations and step-by-step checks are especially important. For HCI/UX practitioners, it offers concrete criteria for accessibility design; for researchers, it provides trust- and validation-related questions for the agentic AI era.

CIT's Commentary

The core of this article is “verifiable interaction,” not the form of the explanation. When AI becomes an agent that executes multiple steps on its own, a single wrong answer can grow into a much larger failure as it propagates backward and forward. In that context, what matters more than visual heatmaps is where the system state has progressed to, and whether users can stop, ask again, or roll back at specific points. Notably, the way BLV users verify through conversation goes beyond a simple accessibility requirement—it generalizes into safety interface principles that apply to all users. Also, since LLMs are not only explanation generators but tools for producing explanations, there is room to design LLM-based UX measurement questionnaires or audit procedures to evaluate the quality of these “verifiable explanations” together.

Questions to Consider While Reading

• Q.To make conversational explanations work in real products without imposing excessive cognitive load, what information should be shown first and what should be revealed later?
• Q.In agentic AI, what are the minimal interface patterns that allow users to undo, stop, and re-check?
• Q.How can the self-blame bias found among BLV users be generalized to AI trust design for other user groups?