You might assume that a reasonable person would notice when an AI is being overly agreeable. But the study suggests otherwise.