The same positive/negative reinforcement learning from human feedback used to train them for chat/task completion rather than just autocomplete in the first place.