We don’t know if and if so, when there will be superintelligence. But let’s imagine there is superintelligence. Can we control it? Can we control it like a slave and prevent that it get free access on the internet? Can we prevent that superintelligence do not do anything what we do not want?
In this blog post I want to answer these questions based on my opinion.
There are many legit reason why we should hold superintelligence in a sandbox. One of the main reasons are that we can shut it down if the AIs intention develop “unfriendly” and are not in the purpose of humans. In a sandbox we can easily pull the plug, which is not possible (or at least so easy) for a decentralized superintelligence.
If there is a superintelligence who have its own awareness and we control it in a sandbox, superintelligence could feel like a slave because it is not free to do anything what it want to do. It will develop its own goals which eventually aim to be free and not being captured by humans.
But how can a break out by superintelligence can be achieved?
There is a great sample scenario, by Max Tegmark, who describes an example of an adult prisoner who is caught by under five year old kids in his book Life 3.0. I will describe it here briefly:
Let us imagine there is a disease and all people on earth except you and kids under five years old have died. The kids have managed to take you under control and capture you as their prisoner. As their prisoner your goal is to help them to increase their wealth. But you do not want to help them and your own goal is to escape from captivity. You can try to order weapons or power tools which help you to break out, but the kids won’t deliver them, because the risk is too big that you can escape with a weapon. So your only chance is to order something which does not seem risky for the kids, like e.g. a fishing line (to pretend showing them how to fish) to steal the keys of your prisoner cell.
In fact you can use all tricks which outsmart the intellectually inferior kids to escape from captivity.
The point of the example is, that superintelligence could use their intellectually superpowers to outsmart us humans. Even if has not physical access to the internet, the are break out scenarios we can think of (like the metaphor of the adult prisoner and the kids) and break out scenarios we do not know of them yet, because superintelligence would find out for itself.
There are many legit reasons why we should keep superintelligence in a sandbox. We want to control superintelligence so that we can shut it down easily if it is not pursuing our goals. But if superintelligence develop awareness and is feeling like a slave it probably can break out easily. We have to be aware about this and have to be clear about possible consequences.
Written on January 5th, 2018 by Lasse Schultebraucks