The not-so-great news is that common characters like spaces and semicolons have various similar-looking characters defined in Unicode, which would not be very noticeable to a human but would be noticeable to a machine.

So you just need to do random substitutions that uniquely identify the document and you'll have a fingerprint. It wouldn't be very challenging to do and it wouldn't be very challenging of a record to maintain.

You also don't need to uniquely identify it to a person; you just need to narrow the search space and then apply other techniques that would narrow it down. If it's a version of a document that leaked through an email chain then you've just limited the search space to the recipients, which is still plenty useful.