> We think most foreseeable cases in which AI models are unsafe or insufficiently beneficial can be attributed to a model that has explicitly or subtly wrong values

Unstated major premise: whereas our (Anthropic's) values are correct and good.