They would need to be trained for open ended creative tasks, which is just hard in the current reinforcement learning paradigm.