Fixing Accessibility After the Fact Is Already Too Late
summary
AI creates more code faster and replicates human accessibility mistakes because models train on messy open source; WebAIM shows homepages averaging ~56 errors/page and model outputs remain nondeterministic, costly in tokens, and prone to noisy fixes. Practical fixes include embedding deterministic testing and accessibility checks into design and code generation, using Axe‑core‑based scanners and repo‑level custom instructions, feeding curated bug diffs back into models, and building harness engineering (agent + deterministic tools) so PRs become verification rather than the first gate.
AI-written summary · about
chapters (6)
- 0:00Shifting left on accessibility
- 3:40AI impact on accessibility defects
- 15:28Training AI for accessible code
- 29:31Automated systems in SDLC
- 41:45Prompting for accessibility
- 51:08Collaborative solutions
related talks
41mAccessibility is a Requirement
48mFind and Fix Your Accessibility Problems
28mYou Don’t Need to Worry About Accessibility with AI (Or Do You?)
50mMaking a Strong Case for Accessibility
26mHow (Not) to Fail at Accessibility
43mHow to Use AI to Build Accessible Products
50mShift Left on Accessibility: Beyond the Audit
25mPrevent, Catch and Fix Accessibility Issues