The problem is that I can author "foo bar" and "moo bar" constructed to hash the same, and then assert retroactively which one I meant. While collisions are likely to have random binary garbage in them, it may not matter - I am on phone now, but IIRC I have two PDFs on my desktop that have the same MD5, one of which viewed is an airbus pamphlet, the other being a Boeing one. (created by Dan Kaminski, IIRC)