Explainable AI for Blind and Low-Vision Users: Navigating Trust, Modality, and Interpretability in the Agentic Era
agents xai
| Source: Mastodon | Original article
A paper posted on arXiv this week — “Explainable AI for Blind and Low‑Vision Users: Navigating Trust, Modality, and Interpretability in the Agentic Era” — highlights a blind spot in the rapid rollout of autonomous AI assistants. The authors, led by Abu Noman Md Sakib of the University of Texas at San Antonio, argue that most explainable‑AI (XAI) tools are visual by design, leaving blind and low‑vision (BLV) users to rely on opaque audio cues or, worse, to mistake system errors for personal mistakes. Their study, accepted for presentation at the upcoming Human‑Centered Explainable AI (HCXAI) conference, maps the “modality gap” and proposes a framework that blends multimodal feedback—spoken, haptic and tactile cues—with “blame‑aware” designs that explicitly signal when an AI has failed.
The timing is significant. As large language models evolve from query‑based helpers into agentic systems that can schedule appointments, draft contracts or even control smart‑home devices, the stakes of misunderstanding grow. For BLV users, an unexplained mis‑action could jeopardize safety, privacy or financial outcomes. Moreover, the EU’s AI Act and emerging accessibility regulations are beginning to require demonstrable transparency for high‑risk AI, making inclusive XAI not just a moral imperative but a legal one.
What to watch next includes pilots that embed the proposed multimodal explanations into mainstream assistive platforms such as VoiceOver, TalkBack and emerging haptic‑feedback wearables. Industry players are already experimenting with “explain‑by‑example” audio snippets that describe a model’s reasoning path, while standards bodies are drafting guidelines for non‑visual XAI. Follow‑up studies will test whether these interventions improve trust scores and reduce the tendency of BLV users to internalize AI failures. If successful, the work could set a new baseline for accessible AI design in the agentic era.
Sources
Back to AIPULSEN