I don't have a source for this, so take it with a huge grain of salt... but for some reason I have a memory of someone telling me that the older versions of Word saved and loaded documents by writing the bytes of in-memory data structures directly to files on disk, with not much in the way of marshalling or validation in the middle. Because it was fast, or something. You can imagine the kind of edge cases and oopsies that might result.

The new versions at least serialise to some kind of monstrous XML representation of Word's internal state, so while it's not going to win any awards for world's most elegant document format, it should be slightly harder to corrupt in subtle ways.