If language models such as ChatGPT are trained using the outputs of other AIs, they will become heavily biased, overly simplistic and disconnected from reality, researchers warn ...