One thing that I would consider quite nice about an AI politician is that you can run them through simulations and find out what they would do under various scenarios. Put the AI in a scenario where it’s offered a bribe, put it in a scenario where it has to cast a vote on a particular bill, and so forth, and you’ll find out exactly what sort of AI it really is.

Of course, consistency will depend on its operators. So there’d need to be some careful oversight.