“The training process for the Cicero AI combines self-play with regularization towards a human data-derived policy, penalizing actions that are unlikely to be made by human players.”