Remember that everytime you're interacting with ChatGPT you're sampling from a distribution, so there's some degree of variation to the responses you get. That's all you need to have regression to the mean.

If the results are really getting worse monotonically then that's a different matter, but the evidence for that is, as far as I can tell, in the form of impressions and feelings, rather than systematic testing, like the sibling comment by ChatGTP says, so it's not very strong evidence.