The diversity of the outputs is actively penalized in many applications.
If you're super creative in RL, it's not good. Yeah. Or maybe if you're doing a lot of writing, help from LLMs and stuff like that, it's probably bad because the models will silently give you all the same stuff.