Memory use: Unicode scalar values go up to 0x10ffff, which on most machines means a 32-bit value for each character. A UTF-8 representation can be less than 30% the size. And that's not even counting the fact that many languages (Haskell included) represent lists as a linked data structure, with the overhead of a pointer per list entry.

Correctness: you often don't want to operate on individual Unicode scalar values. Extended grapheme clusters can combine multiple scalar values to form a single human-readable character, and that's usually the unit you care about. Representing a string directly as a list of extended grapheme clusters would use even more memory.

Fundamentally, a string has more structure than a list representation gives you (encoded bytes vs. scalar values vs. grapheme clusters). I think it's better to expose this structure than it is to pretend a string is just a list of characters.