A representative AI watermarking scheme was proposed by Kirchenbauer et al. at the University of Maryland in 2023; it divides the vocabulary into a green group and a red group based on preceding tokens at each token generation step, biasing selection toward green tokens.