It's surprising that no one has applied ML techniques to historical documents like these. Scanning the documents to uncover some text that was overwritten would be a very useful technique for historical research. Written documents have very distinct statistical properties and if there is a big enough corpus of text then it would be easy to train a classifier to uncover the statistical properties that are not clearly visible but still have the distinct statistical signature of written text.
Presumably we are on the verge of AGI so this seems like a very easy application.