I agree with this. And also think that we should train and select for trustworthy models. I also agree that these models may never truly be trustworthy.