Artificial Intelligence (AI) for Web Accessibility: Is Conformance Evaluation a Way Forward?
Shadi Abou-Zahra, Judy Brewer, Michael Cooper · 2018 · Proceedings of the 15th International Web for All Conference (W4A 2018) · doi:10.1145/3192714.3192834
Summary
This short paper from three W3C/WAI staff members explores the potential and limitations of applying artificial intelligence to improve web accessibility, proposing accessibility conformance evaluation as a strategic pathway to accelerate AI adoption in this domain. The authors survey four areas where AI is already being applied to web accessibility: image recognition (Facebook’s automatic alt-text, Microsoft’s CaptionBot), voice recognition (Google’s automatic YouTube captioning), text processing (automatic summarisation, reading level adaptation, key term extraction for cognitive accessibility), and affective computing (emotion recognition for adapting interfaces for people with anxiety or autism). While acknowledging these advances, the paper identifies three fundamental limitations of current AI for accessibility. First, accuracy: automatically generated text alternatives are often insufficient or misleading — Facebook’s system was criticised for emphasising beards and body parts while missing the actual purpose of images, producing descriptions like "two people standing, beard, feet, outdoor, water" that fail to convey the image’s communicative intent. WCAG requires text alternatives to serve the "equivalent purpose" of images, which demands understanding context and intent, not just identifying objects. Second, accountability: when AI replaces human interpretation (e.g., translating text to symbols for AAC users), the biases shift from human interpreters to algorithms and training data, and accountability becomes diffused. Third, sensitivity: AI accessibility services collect highly sensitive data about users’ disabilities and capabilities, posing privacy and security risks — even knowing that someone uses an assistive service reveals sensitive health information.
Key findings
The paper identifies several potential applications of AI for web accessibility beyond current uses: detecting and describing the role, behaviour, and relationships of web objects (extending WAI-ARIA semantics), describing virtual reality scenes at different zoom levels rather than individual objects, learning user preferences and adapting content for people with fluctuating conditions, supporting accessible web authoring by integrating AI into content management systems and code editors, and using predictive analytics to detect potential accessibility barriers during content creation. However, the authors argue that AI is not yet reliable enough to replace human judgement for web accessibility standards compliance. Instead, they propose using web accessibility conformance evaluation data — particularly human expert evaluations that assess whether content meets WCAG criteria — as training data for machine learning. This approach would improve AI accuracy specifically for accessibility tasks by training on data that evaluates not just what objects are in an image, but whether text alternatives serve the equivalent purpose. The key insight is that accessibility conformance evaluation produces exactly the kind of labelled, expert-verified data that machine learning algorithms need, and this data already exists in significant quantities from ongoing evaluation activities.
Relevance
Written by the W3C’s accessibility strategy lead, the WAI director, and a WAI staff member, this paper carries institutional authority in framing how the standards community views AI’s role in web accessibility. The central argument — that AI should be improved through accessibility-specific training data rather than being treated as a ready-made solution — remains highly relevant as AI-powered accessibility overlay products have proliferated since 2018, often making claims that echo the limitations this paper identifies. The distinction between identifying objects in images and serving the "equivalent purpose" of those images is critical: it explains why generic image recognition cannot substitute for human-authored alt text in most contexts. For accessibility practitioners, the paper provides a useful framework for evaluating AI accessibility claims: does the system understand context and purpose, not just content? Is accuracy sufficient for the specific use case? Who is accountable when AI-generated accessibility content is wrong? And what sensitive data is being collected? The proposal to leverage conformance evaluation data as training data offers a constructive path forward that connects the existing accessibility evaluation ecosystem with AI development.
Tags: artificial intelligence · web accessibility · WCAG · automated testing · machine learning · image recognition · alternative text · conformance evaluation · W3C · accessibility standards
Standards referenced: WCAG · WAI-ARIA