Well, they've been using the whole internet and anything else they can get their hands on to train their models.
If you are using LLMs and want to continue seeing them get better and more useful, you really have to accept that they train on data collected from their users.
Unless there is another technology breakthrough somehow, this is going to be the best they can do to increase user satisfaction and make their product more useful.
TBH, I'd love to have a model which was specifically trained on conversation which I had with an earlier iteration. That would make it adapt to me and be less frustrating. Right now I'm relying only on instruction files to somewhat tune a model to my needs.