I is finite, so it's a discrete probability space. The elementary event in this probability space is "hash h, h in 2^I, has been generated". The probability of every elementary event is 2^-I regardless of what hash h corresponds to this elementary event.
I think this is the key point. The amount of work needed to produce a block depends on the target value and not on the value of its hash. Therefore, using the hash value to determine the canonical chain could be considered as incorrect.
However, I don't see a problem with the distortion in practice, and it does provide the benefits of a reduction in orphaned blocks and a deterrence of selfish mining.