Plus, given that sampling is probabilistic, one would have to prompt the model many times and evaluate all the responses. When I see examples as in the post, I always assume that the result was completely cherry-picked from a larger samples that contained all kinds of stuff.