Anthropic thinks that their models are smart, but not so smart so as to be able to fake responses to their alignment tests. How would they know again?