AI and Human Centred Accessibility
summary
Research tests show Wave finds ~30% of issues yet scanning the top 1M sites surfaces 94.8% of detectable WCAG failures, so LLMs learn broken HTML. Local models (MiniCPM v5, Meta LLaMA 3.2) misidentify images and gender, cloud models (ChatGPT, Claude, Gemini) improve but misreport exact WCAG criteria, scramble heading structures and produce incorrect fixes. Agentic chains suffer error compounding (95%^20 ≈ 36% end accuracy) and balloon cost (50 turns ≈ $2.50, 100+ turns > $100); use spec-driven prompting and small task phases.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:23AI's inevitability
- 3:51Accessibility concerns
- 4:32LLM response generation
- 7:59Bias in training data
- 10:31HTML/DOM accessibility
- 13:19Inaccessible web prevalence
- 15:47AI accessibility failures
- 16:32Alt text inclusivity
- 19:45WCAG standards misinterpretation
- 24:05Document restructuring limits
- 25:47Markup errors
- 28:58AI UI accessibility failures
- 30:20Error compounding in AI
- 33:52Token economics challenges
- 34:56Overpromising underdelivering
- 36:40AI accessibility best practices
- 39:08Human-centered vs AI limits
- 40:53Conclusion
related talks
59mFixing Accessibility After the Fact Is Already Too Late
35mAutomated Accessibility Testing
28mYou Don’t Need to Worry About Accessibility with AI (Or Do You?)
50mHuman-Centered AI and Accessibility
43mHow to Use AI to Build Accessible Products
16mWeb Accessibility is actually well-forgotten old
33mAutomated Accessibility Testing
50mMaking a Strong Case for Accessibility