Human-Centered AI and Accessibility
summary
Claims that generic LLMs solve accessibility fail because models train on largely inaccessible web code; Deque reports early scanners produced 200% of issues (half false positives) and rebuilds to 11 rules with >98% no false positives, open sourcing Axe-core which reaches ~1 billion downloads and ~57% automated coverage. Deque ships AI features for object detection, image Q&A (Be My Eyes, NVDA plugins, AI Content Descriptor) and LLM-assisted code fixes (ChatGPT 3.5/4, Copilot) but keeps a human-in-the-loop to avoid hallucinations and unsafe automatic remediations.
AI-written summary · about
chapters (15)
- 0:00Accessibility awareness
- 1:05Global movement
- 3:36Automated testing history
- 5:32Introduction of Axe Core
- 6:24Barriers and user needs
- 7:26Identifying key issues
- 9:44AI in accessibility
- 10:33AI principles
- 14:26AI image description
- 21:55Life-changing AI impact
- 25:08AI image use cases
- 29:00AI for accessibility
- 32:20Object detection approach
- 35:55LLMs for accessibility
- 42:14Q&A and concluding thoughts
related talks
50mEnhancing Accessibility with AI and ML
43mHow to Use AI to Build Accessible Products
35mWrite inclusive, people-first language about AI
41mAI and Human Centred Accessibility
59mFixing Accessibility After the Fact Is Already Too Late
01h 26mInclusive & Ethical NLP Design Building AI for Everyone
32mAccessibility powered by AI
48mAI and Disability: Progress or Exclusion?