One of the goals of a good hash is a uniform distribution. The source files don't have to be random at all. I wouldn't put any money on the odds of by accident hitting a systematic flaw that security researchers haven't yet found in analyzing the hash.

That said, I think it will be possible to deliberately create a collision in SHA256 at some point in the future.