>if you had a [sic] AI-created drug that tremendously improved cancer treatment outcomes

His whole point is that you wouldn't know if it improved outcomes — what if the AI faked the eval.