Scalable watermarking for identifying large language model outputs [pdf]