When researcher Nicholas Tiller began to feed health questions into chatbots as a test, he expected some imperfections — but not this level of failure. Five AIs, 250 questions and a total score of ...