One cool ML application was examining medieval manuscripts and being able to tell which passages were written by different scribes. That in turn yielded information about how scriptoria were organized, and also showed that some texts (books of hours, religious texts) were assigned to single highly skilled scribes, while other texts (usually more mercantile) were more cheaply done using multiple less-skilled scribes.

There's also been some interesting work on guessing how to fill in lacunae (gaps in a manuscript due to damage or compounded transcription errors).

It’s technically just guessing, but so is a person in that scenario and the person will have less data to use to guess from.

There are other contextual clues that aid the guessing: in particular, the size of the space where the characters are missing, and sometimes grammatical inflections in surrounding words that imply what kind of word is missing (say, a verb or an adjective). That's where experienced human experts can often perform well.