You obviously can hear the bottom bits when the music is quiet, and ears are really good at picking up "wrong" even when it's very quiet. Dither exists to make the errors unobtrusive. Besides, I'm not arguing that 16 bits is more than you need. I'm arguing that 24 bits is more than you need, that 16 is enough. Plus, just because it's possible to master correctly for 16 bits, doesn't mean that everything being converted from 24 to 16 was mastered correctly.