For starters, there are four nucleotides in DNA (A, C, G, and T). Simplifying our model of mutation to "a flipped bit" where mutation takes takes a nucleotide to another one with equal likelihood, and only one nucleotide in that position makes it functional, there is a 3/3 chance that a mutation breaks it but only a 1/3 chance that a subsequent mutation fixes it.

Next, there's the fact that every gene is a sequence of many nucleotides. The odds that the next mutation of the same gene would happen in the same place is low.

Finally, many mutations don't take the "flipped bit" form we assumed above. Repetitions, insertions and deletions also happen, and may be even less reversible (deletion of a meaningful portion would be exceedingly unlikely to be recreated by chance - what provides the information about what should have been there?).

Consider this digital signal:

0000000000

Let's say there's a 1% chance of a bit flipping. But any one of these 10 bits could flip. So for, say, the third bit to flip there's a 0.1% chance to become 0010000000.

Now, to go back there's still a 0.1% chance - but there's a 0.9% chance that _another_ bit will flip. At which point you need two bit flips to go back to your original value.

As you can see, since errors accumulate, the probability of wandering back to a working version is much lower than that of producing increasingly broken versions instead.