I suspect it has something to do with the perceptions around always returning a new thing from a function rather than the mutated input to the function. For example, if you need to mutate the property of an object based on other inputs, the perception is you would clone that object, modify the property appropriately, and then return the cloned object.
If you really need your type to be backed by a contiguous block of memory, you batch updates (stencils, SIMD, etc.)
Not that there aren’t ways to represent mutability in Haskell, just that the de facto use of immutability causes excess allocation.
Consider if you "duplicate" a Data.Sequence Seq (finger tree) for modification. You're not actually copying the whole structure, you are creating a new root node and re-using as much as possible of the common structure.
The end result is that a bit more memory is used in the simplest case, but not due to duplication I think.
The benefit is that a thread can make a modified value at cheaper cost without affecting another thread that is still using the original value. I also think it's easier for the programmer to understand the code.
If it's ok to support only a single concurrent user of the value, then a mutable structure is more efficient. Even in Haskell we have mutable structures that can be used for such purposes.
The interesting question to me is, what should be the default? I think there is a good argument that it should be the safer, simpler immutable structures.
If you're multi-threaded, you have to choose: Do I go with immutable, which means that the other thread gets something safe? It also means that the other thread gets something stale (out of date), which can have its own consequences.
Or do I use a lock or mutex, which means that I have to place guards in all the right places, which is error-prone and can break in weird ways if I miss one?
Or do I use mutable data without guards, and just let it crash? (Because it's probably going to crash if I do that...)
With strong immutability like in Haskell, you can share values even between threads and can avoid defensive copying. Two versions of the same immutable tree-like data structure can also share part of their representation in memory.
(Haskell has other problems causing increased memory usage, but not related to data duplication in my mind)