Meeting minutes
Accessibility of machine learning and generative AI.
Jasonjgw: anything we can coordinate on this and keep it moving
janina: edits from Scott that I need to add to Github page. Need to put into the page. Scott- you marked sections off - are these additions or replacements?
scott: some places easier to replace others to note references. There are two sections around change - gen alt text for images. Added inner search in sections. The rest is where references align with what we wrote and put those in. I can change formatting if need to - let me know.
Janina: these are big chunks of text? (Yes). Image is the same but using different language? (Yes)
Janina: references, probably not the best candidate for that list of candidates. Are these additional?
Scott: no. Drawn from the wiki. So around how they slot in. Discussion to what we've written...Just highlighting where those line up.
Janina: Are we making sure we're referencing text in the doc to existing references in the ...what JS file in the respository?
Jason: dot bib at the end...
Scott: references are just where they are already noted in the wiki. There are places we wrote about AI, based on references
Janina: so we're making connections. What you're giving me is places in the text where those should be references. These are cross references.
Scott: yes
Janina: Second area, thoughts about the content here. GEt a little worried as I read. Excellent write up about what's happening and alt text changing/regenerated...But am I just introducing Scott's edits or introducing new edits. Think what we're seeing is not so useful to SR user. If alt text changes all the time how does this help provide context
to what is happening? Have to talk about the JPEG you're pointing to. Diff alt text in different SRs. DIfferent things under the hood. How to talk about this. How to faciliate this happening and make it useful, not just faciliate AI happening. Seeing the facilitation of AI. But it's not refined for usability, it's refined for promoting AI. Does
enough that's useful that we put up with these problems. If we talk about this more thoroughly...
Jason: there's a tradeoff there. A decision to make regarding (img descrviption ex) - generate description and everyone gets the same one forever and it;s baked in. BUt user can run an LLM themselves and get more details on the image. Then yes, they'll get the problem that it's not the same or the model this year will have different results from
the model from last year. How do you decide what solution is best? At authoring time has the advantage - open to editorial review and same with auto generated captions, for instance. Next year's model might give better results than this year's model. User will have to decide to use the captions or generate their own. If not human edited/reviewed it
will be subject to errors. parallel cases between different applications and open question to user to how they want to proceed and have access. Answer isn't obvious.
Scott: It's interesting parallel to captions and gen alt text. If captions are wrong (1 in 10 not arruate) live catpioning have a frame pf reference. Curated content on slides in a presentation. If quality of presentation is good...then still have the slides to say OK I understand what those words were meant to be. Generated, difficult to know if
the words are good or not.
Scott: Family photo and three generated options for alt text. In terms of trust, are all three options right? But they're all different. Some of the details are important.
Janina: to what degree is wcag 3 talking about AI? Do we know?
Janina: W3C is actively building standardization...but a full usability across the stack, that's what WAI and WCAG are. that's what APA trying to protect. Is what your'e generating usable and usable by people with disabilites. Trying to tool up correctly (whatever tools you use) to get access to content and content manipulation, authoring creation,
consumption...
Jason: Whether it's in the conformance statement related to a11y guidelines or some other way, is it helpful to user that it's subject to human intervention distinguished from material that's been human edited or not.
Janina: wuality may be good but with have charactersistcs
Jason: stronger quality expectation around it. People working on guidelines will need to think about this. Some kind of meta data associated with that? Issue in publishing world
Janina: kinds of questions that we need to pose in this document to be compelling first public working draft, if we don't answer them
Janina: I can put it in a branch
Jason: then we can add some of this conversation in relevant sections
Janina: as notes? put in markers so they don't get lost?
Jason: any preferences how people would like this documented?
Janina: Have a question for Raja...Scott - exciting what's happening in that section, but leaves Raja out. Anyone playing with a UI with an SL (sign language)? Anyone tried? Does it work? I'm curious. Is it AI generatable type of thing?
Jason: in principle. But subject to issues Microsoft found in 2019 of auto generated SL iamges. No large data sets for system to generate SL images. In principle if those can be addressed...
Janina: possibility if you create one that's generally available. Something not complicated like hotel or airline. If you do something simpler you start to create that database of what people do and how they respond. Have a camera looking at individuals who are signing to you that AI can learn from. I don't know recipes...something straight forward
Jason: if someone needs specific vocab or terminology it might be able to use that and modulate its language and meet the needs of the user. Possible today or at least not too far away.
Janina: interesting to talk about adaption to user needs in the world...
Jason: meet many needs that are currently not well met by existing tech, if you have the ML system set up to interact with them and the underlying interface that doesn't have to be adapted (scribe note: I might have missed some points/words here)
Janina: trying to encourage W3C to do both - encourage AI to be integrated to what you can do. Actual access...Refine and encourage what we think is useful and profitable to users in AT development
Janina: personalization of interface has been a goal
Jason: this is a potential vehicle to do that. Symbols for an AAC system...ore a suitably refined model as well. This hasn't been investigated...
Scott: for example, Inno search, could be that someone requires symbols to interact about booking flights, or an avatar for signing. How far off is AI to pull that off in an interface that works. Insightful point. How closely the lines in a structure...what is the audience or value in pulling this together...lots of cross over in accessible capture
of work. That doc - this is capture, what's going on and what people are using and implications. Similar doc- these are solutions that work and flip around, how can these things be beneficial and associate risk as AI dominates and ensure a11y incoirporate into their process. Noticed some parallels in the work.
jason: get the questions/issues noted in there...
Janina: a little bit more on this - think I should make two branches before making citation references.
janina: I'll get those done and make sure they're clean and correct, then do the citations.
Jason: keep notifying over email how it proceeds and develop ideas of issues as we work through this. Can bring back next week for consideration. Do that and bring conversations back? Anything else to cover?
Janina: there's going to be an AI demonstration in an hour and a half
Janina: Benetech to the MM model that's being published and working with it since the Vancouver TPAC. Now going forward with it in a more refined and elaborate form, they used AI to create a front end in hours. About two days with human intervention...
Jason: can we get a synopsis on this for next time?
Miscellaneous topics.
Janina: I can give the recording link to RQTF group
Janina: TPAC planning continues. Good time zone for AUS for virtual participation
Janina: break outs are wide open if anyone wants to do them