Whoops. Turns out the JBIG2 compression algorithm substitutes fuzzy parts of the document with similar parts from the rest of the document.

Sounds like a good idea in theory, except a fuzzy 6 looks awfully like a fuzzy 8. And some more examples too!