Would an advanced AI try to take over the world? If it uses a common approach called ‘reinforcement learning’, the answer is almost certainly yes.