“A significant risk of preference aggregation methods for AI alignment is that they will reflect existing societal power structures rather than ideal ones.”