> How well do LLMs handle handwriting recognition?

Pretty well for neat modern handwriting, but much worse for cursive or messier writing. They also really struggle if the text is at an angle. I have some recent experience with a project where we tried to use LLMs to digitise handwritten specimen labels from the 19th and early 20th century, and the success rate was far too low to proceed with that approach.

Hallucination was also a common problem, with the output often replaced by a similar (but more common) name or word.

I’d assume you could improve the results by using a model trained specifically on handwriting data sets, grounding the model, or using existing purpose-built OCR tools - but frankly that’s above my pay grade.