Will a general artificial intelligence be benevolent?

It is hard to predict what AIs will do, especially as we have problems to define what an AI is. If you take a chess program, most would define it as narrow AI. Good at chess, but worthless at other things. But how then do you make a general AI? And if you succeed to create one, what makes it do anyting at all? Why wouldn't it just sit there, thinking of something, and remain incommunicative?

I think one of the important building blocks of an AI is the driving force. The problem with chess programs is that the driving force is extremly narrow. It is something like "for each piece, try every possible move. For each of the tests, try every possible response. All the time, evaluate the situation according to a strictly defined algorithm". And some more. This doesn't give the program much ability to generalize.

If the computer instead would have a driving force more generally defined, as "play human opponents, goal is to win", then it would leave more room for out-of-the-box solutions. For example, you win more if the humans play worse. So one solution would be to select low talented opponents. Another solution that would fulfill the goal would be to somehow arrange to have the opponents killed, and win by walk-over. A program that can take this kind of very general definition of driving force doesn't exist yet. Even if the given example is extreme, more near future realistic examples can be imagined.

It is a version of the old adage "the way you ask something determines what kind of answer you get". I think the defined driving force, the algorithms, and the rules, are what will ultimately make up an AI. You don't really need the rules to make an AI, but I strongly encourage the use of them. Even if the driving force is seemingly benevolent, it is no guarantee of the eventual behaviour.

A discussion you see now and then is if the driving force should be copied from humans. Looking closer to this, I think the main human driving force is to spread your genes. Everything else is just a side effect, to improve the chance that you succeed. Behaving nice to people, even doing charity, improves the chance.

The hope is that using a human driving force for an AI would also give it the same behaviour. But I think that is a mistake, as the nice part of (most) human behaviour is only a side effect. There are spectacular examples of humans that succeed well on an individual point of view, but catastrophically from the point of view of the human civilization.