This is just a way to prove, statistically, that one model is better than another as part of its validation. It's not collected from normal people using ChatGPT, you don't ever get shown two responses from different models at once.