The Perceptual Gap: Why We Need Accessible XAI for Assistive Technologies
Shadab H. Choudhury · 2026 · Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA '26) · doi:10.1145/3772363.3799105
Summary
Choudhury (University of Maryland, Baltimore County) presents a position paper and targeted literature review arguing that explainable AI (XAI) — the body of methods that help users understand why a black-box model produced a particular output — is fundamentally inaccessible to the very users who most need it: people with sensory disabilities who rely on AI for perception tasks. Apps like Seeing AI, Lookout, Envision, and TapTapSee let blind and low-vision users identify objects, read text, and describe scenes; Otter and built-in captioning let deaf and hard-of-hearing users follow spoken language. These systems are regularly wrong, but existing XAI methods (saliency maps, Grad-CAM, SHAP, LIME, prototype explanations, heat maps, spectrograms) rely on exactly the perceptual channel the user cannot access. This is what the author calls the perceptual gap — explanations delivered in the same modality the user is already unable to perceive. Methodologically the paper surveys four core accessibility and HCI venues (ASSETS, TACCESS, CHI, TOCHI) by progressively filtering ACM Digital Library search results on accessibility + AI/ML + XAI keyword terms. Of 2,936 accessibility papers, 1,092 mention AI/ML, 147 mention XAI, and only 10 actually validate XAI methods with users after manual verification. The paper asks RQ1: to what extent does existing XAI work support users with disabilities, and RQ2: how can Accessibility and XAI research jointly improve explanation usability.
Key findings
The survey finds that the intersection of XAI and sensory-disability research is vanishingly small. Visual XAI dominates the literature (Grad-CAM and its variants, SmoothCAM, ScoreCAM, LayerCAM, prototypes, counterfactuals), and all of these presume the user can see both the input and the explanation — a requirement that excludes blind and low-vision users almost by definition. Audio XAI (spectrograms, phoneme-level confidence, feature-based explanations for speech captioning) is less well developed but still largely visual when surfaced to users. Language-based explanations are promising because they can be rendered as text, speech, or braille, but they rely on the same visual captioning models used in the underlying assistive app and therefore tend to be verbose, sometimes incorrect, and not verifiable by the user against ground truth. The author introduces the concept of verifiability — whether a component of an explanation can be cross-checked using another sense the user has (touch, hearing, smell) — and argues it should be the organising design goal for Accessible XAI. Five research directions are proposed: make verifiability explicit in multimodal explanations; improve vision-language models on datasets labelled by disabled users (e.g., VizWiz); make XAI visualisations themselves accessible (audio data narratives, tactile representations); integrate XAI into Teachable AI systems so personalised models can also be contested; and invite disabled users into every step of model training, not just data collection.
Relevance
This paper reframes explainability as an accessibility problem and should be required reading for accessibility practitioners evaluating AI-based assistive tools. The 'perceptual gap' framing names a failure mode that procurement and compliance checklists rarely capture: an AI tool can satisfy every technical accessibility requirement (screen-reader-compatible UI, proper labels, keyboard support) and still leave users unable to assess whether its outputs are correct. The verifiability principle maps cleanly onto real accessibility design: a cat-identification app should ideally tell a blind user not just 'cat' but 'tabby cat lying on a couch near a window, based on features at the centre of the image', so the user can cross-check against what they hear or touch. The survey evidence that only ten accessibility-venue papers actually validate XAI with disabled participants also supports advocacy for participatory-design mandates in assistive-AI funding calls. Limitations: the paper is a position piece with a focused survey rather than a systematic review, the search is limited to four venues and the ACM Digital Library, and it does not yet propose concrete evaluation metrics for accessible explanations — those remain open research problems.
Tags: explainable AI · XAI · artificial intelligence · blind and low vision · deaf and hard of hearing · assistive technology · human-centered AI · accessibility research · AI accessibility · position paper