OpenAI researchers are testing “confessions,” a technique that trains fashions to confess after they make errors or act undesirably, serving to enhance AI honesty, transparency, ..
