[NFB-Science] AI and STEM Access

[Louis Maher]

Hello,

A friend of mine, Robert Shelton, sent me this note.


Regards
Louis Maher
Phone: 713-444-7838
Email: ljmaher03 at outlook.com

From: rshelton1 at gmail.com<mailto:rshelton1 at gmail.com> <rshelton1 at gmail.com<mailto:rshelton1 at gmail.com>>
Sent: Thursday, May 28, 2026 10:37 AM
To: ljmaher03 at outlook.com<mailto:ljmaher03 at outlook.com>
Subject: AI and STEM Access

Good morning. I recently saw the agenda for the CS Division meeting of the NFB, and I wanted to mention something I think could be significant. Over the last few months, I've been experimenting with ways to use AI models to interpret math. There are a few gaps in my knowledge of special functions, and I've always been interested in filling them, but most higher math is found in inaccessible sources. What I discovered is that models like Gemini and ChatGPT will not only find the content for you, but upon request, they will translate it into a plane text representation. Now, it's not exactly LaTeX, but it's actually just plane text. For example, I was looking for a proof of the duplication theorem for Carlson Symmetric Forms (https://en.wikipedia.org/wiki/Carlson_symmetric_form), a nontrivial example, and it was easily readable  with a screen reader, and I suspect would have been laid out well for a Braille display. The most exciting part is that you can always ask for clarification if the model uses notation you don't understand.

This is not quite ready for prime time yet. I spent quite a bit of time debunking ChatGPT's ideas about how the proof should go, but the translation part, i.e. the capability to translate from PDF, or graphics, or (shudder) MathML into plane text is solid.

It is hard for me to overstate the importance of this development. To get plane text, all I had to do was explain that I was a totally blind mathematician trying to learn a new area. I had to remind ChatGPT when it "forgot" and used MathML, which I detest, but it was quick to fix the issue. If memory serves, most of my research was with ChatGPt 3.x, so by no means the most powerful model available. I tried Gemini, and that worked great, but I don't have nearly as much experience as with ChatGPT.

I think we need to start using this capability aggressively so the model developers will realize there is a market for it. I believe that including a translation capability for math content may have been a deliberate decision, and could easily disappear if not used.

Thanks for listening.


--Bob