If you roleplay or do fanfic or write any story—original or some fanfic of an already existing story—the dialogue sounds awful, and that’s true for every AI: ChatGPT, Claude, Grok, etc. Mainly, the dialogue is unbelievably repetitive, and then characters either have an overreaction to something or an underreaction to something. The dialogue sounds truly, truly awful. No one really has a distinct personality; they all sound the same, repeat the same lines, and the dialogue is cringelord-tier. It doesn’t sound like real, actual people talking.
It doesn’t sound like real actual people talking because it’s not real actual people talking.
Try something for me. Type a sentence, any sentence, on your phone. Then mash the middle suggested next word button about fifty times.
Congratulations, you’re an AI! Well, that’s a very simplified version of what generative AI does, it takes your text and then predicts the likely “next” text again and again until it either runs out of room or predicts “text ends here”. Yes there’s some preamble, and it’s not using the same dictionary, and it’s not predicting word-by-word but token-by-token, and there’s some randomness thrown in for spice, but that’s about the size of it.
When you’re texting, do you ever type a word rather than selecting the middle word? Or even the top three? That’s the fundamental reason why LLMs don’t talk like humans. They literally cannot do that.