"Their performance depends solely on the model training before release and how well you curate the context you feed it".

Wrong. The face-saving backtracking doesn't change that.