In an experiment, AI agents were left to converse with each other. Soon their expressions became incomprehensible to the researchers, who were cut off from the dialogues.