Blind person here, can confirm this. Reading PDFs with a screen reader is bad, reading PDFs that come from LaTeX is worse, reading LaTeX math is pretty much impossible. All the semantic info you need is just thrown away.

You can make decently accessible PDFs but it's lots of work, you need Acrobat on the producer' side and might also need it on the consumer's side. Free tools don't even come close. There's also the fact that the process of making accessible PDFs in Acrobat isn't itself accessible.

With that said, the way screen readers treat HTML math certainly isn't perfect, it's geared more towards school children than anything above calculus. I'm probably going to stay with my LaTeX source files for now. At least ArXiv offers those, not many sites do. To be fair, that approach also has its own set of problems (particularly when people use some extra fancy formatting in their math equations, making the markup hard to read), but I find this to be the best approach for me so far, at least on AI/ML papers.