Sort of. There are biases introduced during training/post training and there are the additional runtime / inference safeguards.

I’m referring more to the runtime safeguards, but also the post-training biases.

Yes we are talking about degree, but the degree matters .