You underestimate how hard the problem is. Things that in other fields would be minor typographical differences are often semantically meaningful in traditional math notation. Formula OCR would need to distinguish chi from x, lowercase delta from o with an overbar, a dot-product dot from a dot intended as multiplication, parentheses from brackets, v from the or sign. It needs to be able to determine whether your fraction "a / b / c" (by / I mean horizontal lines) was intended as "a / (b / c)" or "(a / b) / c" from the size of your letters, or from the relative widths of the fraction lines.
Of course, anything a human can do, ML can learn to do eventually. Which invites the question: is reading arbitrary hand-written math actually something any human can do?
Of course, anything a human can do, ML can learn to do eventually. Which invites the question: is reading arbitrary hand-written math actually something any human can do?