Re: [webai] AI and accessibility: opportunities and risks (#14)

## Opportunities

Academic studies have shown where LLMs have better search/retrieval capability when the source documents are well structured, markdown being preferred. Between this fact and the cost of context (measured in tokens), several browser automation projects have popped up to give agents better access to the web and all picked the axTree as the primary resource. When saved/cached, it can get converted to markdown. Edge providers, like Cloudflare, have created services to offload this repetitive task. Sites that do this well, get ranked higher in "SEO" when AI is layered on top. If only 3 citations appear at the top of Google, first page is no longer good enough.

All this to say, accessibility-first (which we've been hammering on for years) is also agent-first. They intersect on wanting well-formed, structured information in plain language. Robots are people too?

## Risks

- If anyone can vibe code a UI, but the agents pick the shortest (lazy) path, it won't be accessible. 
- UIs moving to the terminal (because agents like the CLI), causes more accessibility issues than web.
- Social engineering attacks are way more sophisticated. The speed at which someone can be tricked into divulging PII thinking they're talking to a trusted person OR a trusted AI is astounding. Some attacks even [target the AI](https://www.microsoft.com/en-us/security/blog/2026/02/10/ai-recommendation-poisoning/).

-- 
GitHub Notification of comment by AutoSponge
Please view or discuss this issue at https://github.com/w3c/webai/issues/14#issuecomment-4042061794 using your GitHub account


-- 
Sent via github-notify-ml as configured in https://github.com/w3c/github-notify-ml-config

Received on Wednesday, 11 March 2026 20:38:10 UTC