E-mail List Archives
Thread: How Useful is Gemini/Bard?
Number of posts in this thread: 6 (In chronological order)
From: David Russell
Date: Tue, Mar 05 2024 2:04PM
Subject: How Useful is Gemini/Bard?
No previous message | Next message →
Hello WebAIM forum,
I am an NVDA screen reader user, and to preserve some of the "how to"
I have grown to rely on, I have begun asking the Google A.I.
assistant, Gemini, how to do specific tasks with my screen reader?
I am experiencing something of concern in that directions given do not
always equal experience gained.
Most recent, trying to link my "pay account" to Fiverr. Their
accessibility for the vision impaired in my opinion is a joke.
There is a definite disconnect between what the tutorials intend with
what experience provides.
Is this occurrence on the minds of those who produce, administrate,
and roll out products for general use by the public?
Input appreciated!
--
David C. Russell, Author
= EMAIL ADDRESS REMOVED =
Human Being Human Doing
David C. Russell
https://books2read.com/u/mlqXG9
Available now at your favorite digital store!
books2read.com
From: Hayman, Douglass
Date: Tue, Mar 05 2024 2:25PM
Subject: Re: - How Useful is Gemini/Bard?
← Previous message | Next message →
David,
I'm curious to see what those AI tools are using to do their machine learning. Are they scrolling through a huge table of commands for JAWS or NVDA and calling that sufficient? Are they scouring some short youtube video of someone using a screen reader?
Doug Hayman
IT Accessibility Coordinator
Information Technology
Olympic College
= EMAIL ADDRESS REMOVED =
(360) 475-7632
From: Léonie Watson
Date: Tue, Mar 05 2024 2:27PM
Subject: Re: How Useful is Gemini/Bard?
← Previous message | Next message →
David,
The short answer is that you cannot entirely trust the responses you get
from any generative AI tool, including Gemini or ChatGPT.
The longer answer is that these tools have been trained on the data that
has been amassing on the web for the past 30 years. So they've learned
all the nonsense, all the good stuff, all the right ways to do things,
as well as all the wrong ways.
You need a measure of knowledge of the subject you're querying to be
able to verify the usefulness of the response you get. I wrote this post
recently that touches on one aspect of this:
https://tetralogical.com/blog/2024/02/12/can-generative-ai-help-write-accessible-code/
That said, sometimes that isn't possible. I use these tools to describe
images (that I can't see and have no way of verifying), but I'm aware of
the limitations of these tools, their habit of hallucinating, and can
take that into account when judging whether or how much I trust the
description I've been given. A somewhat light-hearted look at this is here:
https://tink.uk/adventures-with-bemyai/
It sounds like you've verified the trustworthiness of the responses
you've received by measuring the gains you've made (or not) by following
the advice given.
HTH
On 05/03/2024 21:04, David Russell wrote:
> Hello WebAIM forum,
>
> I am an NVDA screen reader user, and to preserve some of the "how to"
> I have grown to rely on, I have begun asking the Google A.I.
> assistant, Gemini, how to do specific tasks with my screen reader?
> I am experiencing something of concern in that directions given do not
> always equal experience gained.
> Most recent, trying to link my "pay account" to Fiverr. Their
> accessibility for the vision impaired in my opinion is a joke.
> There is a definite disconnect between what the tutorials intend with
> what experience provides.
> Is this occurrence on the minds of those who produce, administrate,
> and roll out products for general use by the public?
>
> Input appreciated!
>
--
Léonie Watson (she/her)
Director
https://tetralogical.com
From: Dean.Vasile
Date: Tue, Mar 05 2024 3:18PM
Subject: Re: How Useful is Gemini/Bard?
← Previous message | Next message →
Hello David
Another thing to keep in mind when using generative AI
It is usually a year and a half to two years out of date.
I usually ask whatever version I'm using, how recent it's last update was
so keep that in mind as well.
Good luck
Dean Vasile
617-799-1162
> On Mar 5, 2024, at 4:04 PM, David Russell < = EMAIL ADDRESS REMOVED = > wrote:
>
> Hello WebAIM forum,
>
> I am an NVDA screen reader user, and to preserve some of the "how to"
> I have grown to rely on, I have begun asking the Google A.I.
> assistant, Gemini, how to do specific tasks with my screen reader?
> I am experiencing something of concern in that directions given do not
> always equal experience gained.
> Most recent, trying to link my "pay account" to Fiverr. Their
> accessibility for the vision impaired in my opinion is a joke.
> There is a definite disconnect between what the tutorials intend with
> what experience provides.
> Is this occurrence on the minds of those who produce, administrate,
> and roll out products for general use by the public?
>
> Input appreciated!
>
> --
> David C. Russell, Author
> = EMAIL ADDRESS REMOVED =
> Human Being Human Doing
> David C. Russell
> https://books2read.com/u/mlqXG9
>
> Available now at your favorite digital store!
> books2read.com
> > > >
From: Jeremy Echols
Date: Tue, Mar 05 2024 3:32PM
Subject: Re: How Useful is Gemini/Bard?
← Previous message | Next message →
I've found that ChatGPT Plus is often useful, as it will search the web for information that might not be built into the AI. I don't know if Gemini does this as well. But even so, ChatGPT Plus absolutely still makes mistakes.
There are three scenarios where I find that it's particularly bad:
1. Asking about something niche, like a tool that is only in use by university libraries, and/or is just sparsely documented.
2. Asking Something that changes a lot, but has vast articles and documentation for all its versions. Drupal can be a major source of conflicting information online.
3. Asking about something controversial in any way - even if the AI is amazing, it can have trouble deciding who is telling the truth online.
While I've found ChatGPT Plus to be really useful at times, I always double-check its responses when it's an important situation, and in any of the above three scenarios I expect that I'll have to ask a few follow-up questions in addition to double-checking the information.
From: Jeremy Echols
Date: Tue, Mar 05 2024 3:42PM
Subject: Re: How Useful is Gemini/Bard?
← Previous message | No next message
I just realized I didn't closely read your actual question. Let me try again.
If your question is "can AI figure out how to do something Fiverr documented wrong?", probably not. If Fiverr's documentation is incorrect, the AI will almost certain fail. Large language models (LLMs), the basis of the current AI trend, are incapable of learning things. They're trained on a huge corpus of data and then "frozen" for all intents and purposes. If you're hoping for AI to tell how to do something that's not documented, it's very unlikely to succeed. It might give you suggestions or ideas to try based on whatever data it's crunched in about NVDA and Fiverr, but it will just be general guesses, it might totally lie (i.e., hallucinate), and it will likely waste your time.
If your question is "can AI help me learn screen reader commands and tricks that might help with this task, but aren't specific questions about Fiverr?", that's possible. The massive LLMs are trained on gobs of free data, which often includes open source manuals and other information. It could take some back-and-forth, but you might find a really high-power AI pretty helpful here. I don't know about Gemini, I haven't used it, but ChatGPT Plus and Claude Pro both seem pretty capable, especially if you spoon-feed them a PDF manual for something like NVDA.