W3C

– DRAFT –
Demonstration of AI Powered Accessibility Auditing

25 September 2024

Attendees

Present
Ben_Tillyer_, CharlesL, cyns, duga, Fazio, hgarg, MJ, NehaJ
Regrets
-
Chair
Fazio
Scribe
hgarg

Meeting minutes

Traditionally accessibility audits have been manual. We're attempting use of AI to automate some of it.

Leveraging AI (LLM's) where they help.

Current set of tools have limited localization in local languages. We have used AI to localize the tool itself and also to analyze/audit pages in different languages.

<cyns> Can you share a link to the slides? It's hard to see in the room

One rule that the tool checks against is WCAG 3.1.4 which checks for use of uncommon abbreviations in text.

Another rule that the tool checks against is WCAG 3.1.4 which checks for use of unusual words in text.

<Ben_Tillyer_> @fazio, we have had requests for a link to the slides if possible

Use of LLM's allows analysis of different languages in the tool.

The tool would incorporate elements to allow end users to interact with AI to fetch more examples or possible solutions for reported issues.

The tool also uses LLM's to analyze where the Link text is inadequate and could more meaningfully convey the link target.

Question: `Click here` can be reported by non AI tools too. Do you look at the surrounding text too?

Answer: Yes, the context is shared with AI.

<Zakim> Ben_Tillyer_, you wanted to ask 3 questions 1) about links 2) about hallucinations 3) about sustainability

Question: How do you protect against hallucinations?

Answer: Human intervention is used.

Question: 10 clients submitting requests that are similar? Do you plan to offline some of the LLM responses?

Answer: Yes. Not yet. But a necessity.

Limits on number of tokens, requests are current challenges

<Fazio> ?

Question: How do you fix source files hosted on wordpress?

Answer: We can't, but the idea is to test it locally via the tool. And then fix it at source

Question: Is checking for use of correct semantic tags on the horizon?

Answer: Yes, though not implemented yet. Work in progress.

We will be at m-enabling in October also showcasing this.

Question: Existing LLM's vs fine tuning vs prompt engineering, which one?

Answer: Existing LLM's with appropriate input from the web page for analysis.

Email: dfazio@helixopp.com

<Fazio> dfazio@helixopp.com

<NehaJ> njadhav@helixopp.com

<Ben_Tillyer_> rssagent, create minutes

That worked ZoeBijl Ben_Tillyer_

Thanks

<ZoeBijl> Did that capture everything?

Yes most of what I put down as scribe.

Minutes manually created (not a transcript), formatted by scribe.perl version 229 (Thu Jul 25 08:38:54 2024 UTC).

Diagnostics

Maybe present: Answer, Email, Question

All speakers: Answer, Email, Question

Active on IRC: Ben_Tillyer_, CharlesL, chrisp, cyns, duga, Fazio, hgarg, Lionel_Wolberger, MJ, NehaJ, tpac-breakout-bot, ZoeBijl