Re: Reflow

You are quite write to check if what ChatGPT is true before using it. AI
tools like ChatGPT are very questionable at the best of times. They are
very useful as tools, but all they do is draw information from a range of
other sources, and the information they gather are just a reflection of
those unknown sources. So I would not class anything obtained that way as
"definitive"; much better to track down the original sources and (if they
are reputable), quote them instead.
AI tools can pick up wrong information. Perhaps more dangerously, they are
being found to keep many of the same prejudices and preconceived ideas that
the majority of people do - I have seen comments to that effect in the
world of research. That is worrying if they promote anti-equality views. (I
think it was in Nature magazine where I recently saw a comment by
researchers to that effect.)
On that note, I saw this from a researcher last week:

"Bias too, as well is always going to be a really overriding factor and
something to really consider around accessibility because we do know that
these tools can be prone to reinforcing the bias from their datasets.Those
datasets, particularly when they've been scraped from the Internet, such as
in the creation of things like ChatGPT, we know they're from primarily the
Western English side of the internet, and they'd be dominated by largely
neurotypical writing. I know neurodiversity specialist Liz Chart Hall, who
has done some excellent talks on using GenAI tools with neurodivergent
students. They emphasised that AI is not a neurodivergent thinker. I think
that's really important to keep in mind that the outputs can reinforce
those biases...."


*Using AI for accessibility*
Only yesterday I came across a fascinating case where an accessibility
consultant, Joe Watkins, did a random test to find out how reliable an AI
tool might be on our subject. His original question elicited several points
of bad advice from it. He had to poke it and prod it, using several
supplementary questions, before he could beat it into submission to give
the right answers!
ChatGPT recommended a placeholder for an unlabelled field, for instance,
instead of a label. In fact, it disputed the need for a label. Twice it
actually apologised (which seems a very human touch, I have to say!), when
challenged on things, for giving wrong information. But in both cases it
still refused to back down completely and insisted it's solution should be
used in addition to the correct solution! It also displayed a pronounced
reliance on ARIA rather than HTML, and argued the toss when Joe quoted the
First Rule of ARIA.
Joe Watkins' conversation with ChatGPT can be seen at
https://intopia.digital/articles/using-chatgpt-to-make-chatgpts-experience-more-accessible/.
Have a read of it, it's very interesting!
The problem with this kind of thing, of course, is that Joe is an expert
who knew the supplementary questions to ask to eventually get the right
answers. The average user would not know the questions to ask and would
probably take ChatGPT's misleading information as definitive.
Regards,
Guy Hickling