I don’t believe this is true. The usage of the birthday problem to find collisions doesn’t require introspection of the hash function. While yes, the map from preimage to output is different, two quality hash functions should both have the same distribution: one effectively uniform.
Mind linking the evidence/theory behind this claim?