Meeting minutes
Artificial intelligence and people with disabilities.
jasonjgw: introduces topic and our presenter, Lionel
janina: Notes that OpenAI is now to be part of Microsoft Bing according to overnight news releases; It was mainstream news this morning in DC
Lionel: Provides a few definitions and notes the pervasiveness of A-I in everyone's day
Lionel: Discusses use in labeling vis a vis common labels in our environments; harms, identification of items, etc.
lionel: Reviews the bitcom taxonomy of A-I/ML; noting there's no overall taxanomic consensus
lionel: Reviews his company approach in web content and notes how IOS has now incorporated some similar ML procedures for a11y
lionel: Shows OpenAI generated code examples
lionel: And now an example of an "Accessible Summary"
<scott_h> great presentation Lionel, thank ou
janina: Asks whether Lionel recalls some gap related questions ...
lionel: were equity related, but not finding right now ...
janina: let's not digress for now
jasonjgw: asks about risks when a-i is asked to close a11y gaps; how to think about relationship between users and content creators?
lionel: notes traditional buckets of mainstream vs. AT; notes too much opportunity for problems, but also overwhelmingly helpful
Lionel_Wolberger: notes speech reco, and TTS is now essentially free; so an example of benefit that's now just generally available
Lionel_Wolberger: We should look at failure modes and assess how disasterous the failures are; we could get better at assessing low vs high cost failures
Raja: q from legal perspective
Raja: in communications there's push for a-i to substitute for human mediated access in the U.S.
Raja: sometimes a reasonable accomodation might be OK with ASR; but there are breakdowns, so in certain situations it would not function as a reasonable accomodation
Lionel_Wolberger: agrees there's much to discuss there
Raja: asking actually in general, e.g. emergency;
Raja: concerned about accountability
Lionel_Wolberger: have seen this addressed in criminal assessments;
Lionel_Wolberger: would need at least a warning
Lionel_Wolberger: believe it's positive that sentencing and parole support are aware that a-i does indeed skew due to dataset
Raja: agrees
Raja: asks for slide sharing?
Lionel_Wolberger: Yes, the Zero Project will be public, and I will share what I've presented here
janina: asks benefit of identifying topic area, e.g. a Dewey Decimal kind of topic identification
lionel: shares an ibm fact sheet that asks to have dataset contents identified
Lionel_Wolberger: i.e. what were you trained for?
Lionel_Wolberger: ex. mortgage evaluator
Lionel_Wolberger: not finding that info in a google? now looking at microsoft responsible ai -- in azure cloud has a dashboard with explainer
Lionel_Wolberger: sometimes the corporate work forgets what it has done!
Lionel_Wolberger: not trivial
<Joshue108> very interesting, good presentation Lionel!
<Lionel_Wolberger> Lionel thanks everyone for their generous listening.