So, I’ll try to keep this as short as possible but it’s gonna be long
anyway.
Okay, the first thing I used it for as an accessibility tool was getting
top (1) Linux output into a narrative format. I also do that with just
about any command output that’s in like tabular format. I’m not a
developer, but I try to use the command line because sometimes its
easier than trying to use a GUI.
The second thing is describing images. Be My Eyes is great on the
iPhone, but they don’t have a desktop app, and their Android version
doesn’t have the feature where I can send photos to it. So I have to use
the ChatGPT app for that. I’ll get to the accessibility of the ChatGPT
services in a while. So, describing images can be really great, but it’s
hard to tell if its accurate because I’m blind, so can’t judge the
accuracy myself. So I don’t rely on it all the time, but it is really
nice for pictures sent to me, or public social media stuff like Facebook
pictures.
The next thing is video games. And this is kinda where it fails
sometimes. So, in visual novel games, like the story mode of Blazblue,
it works really well, but in action games, like Chrono Trigger, or Final
Fantasy, it doesn’t know where things are, and can’t always lead me to
things. Also, I’ve tried sending the OpenAI API a batch of images to try
to describe them like a video, and it can’t handle more than maybe 50
images at a time. So, sending it videos of gameplay doesn’t seem like
it’s gonna work well.
Okay, so the ChatGPT apps and such themselves. iOS works best,
naturally. VoiceOver speaks new messages as they appear, and the app
works pretty well. The Android app is accessible, but messages aren’t
automatically spoken by TalkBack. On the web, it’s much worse. The place
where you choose a model is just text to screen readers, even though
when you press Enter, it opens a list to choose a model. So, a screen
reader user wouldn’t think to press Enter on it, since it’s not a
button, or link, or dropdown box (combo box as screen readers call
them). Messages aren’t spoken as they appear, so we have to keep
checking the “stop generating” button to see when it changes to
“regenerate” or whatever it says, to see when the message is done
generating. AI companies don’t really focus on accessibility, besides
Google and somewhat Microsoft, and they have accessibility teams that
have to work very hard to be heard over all the other teams.
So, we have to make our own interfaces to OpenAI. We have this NVDA
addon: GitHub - aaclause/nvda-OpenAI: Open AI NVDA add-on 1 which works
very well. I use it a lot. But, it uses the OpenAI API, even though I
already pay $20 a month for ChatGPT Plus. So, I’m essentially paying
extra for accessibility. Also, the Vision API that regular OpenAI API
can access is different than the API that Be My AI uses, so I’ve heard,
so we don’t get details like popular characters in comic books, or place
names and such that Be My AI happily gives us. So even when we use
accessibility tools, we still get left out is some cases, because Be My
Eyes is a small team, that prioritizes iOS over what a lot of blind
people use more, the web/desktop and Android.
Oh yes, I’ve also tried reading comic books with GPT4 Vision, and it
works well on Be My AI, the GPT4 Vision part of Be My Eyes, but since I
have an iPhone SE 2020, a tiny iPhone because I’m not rich, it can’t
read the speech bubbles. So no comic books. And on a PC where I can send
it big enough pictures, it’ll sometimes say it’s not allowed to read
stuff. So no BatMan: the Killing Joke, for me.
Overall though, it’s pretty helpful. I wish it had much more data within
the model about accessibility, blindness forums and accessibility blogs,
and and standards like WCAG, and Braille, but it’s gotten better since
GPT4 came out. I know a lot of people just use the free 3.5 and give it
a bad name because of that, but it has improved. I just wish access to
images were equally good across desktop, web, and mobile, that AI
companies paid tons more attention to accessibility (like those AI pins
and phones that have no accessibility info whatsoever), and that ChatGPT
was accessible everywhere.
https://community.openai.com/t/how-are-blind-people-using-openai-technology/591018
===========================================================
The fb-exchange mailing list
Manage account,
List Page: https://www.freelists.org/list/fb-exchange
Archive: https://www.freelists.org/archive/fb-exchange
To unsubscribe: log onto the List page and select "Unsubscribe".
Administrative contact: insight@xxxxxxxxxxxxxxxxxxxx
===========================================================