What would be an argument to restrict AI instead of interacting with it (or he/she)? [closed]

Because you don't know what the AI has learned.

Advanced General Intelligence at human level is actually much more capable than humans. The problem lies in the speed they think. Current computers which are "dumber" than humans far exceed us in certain tasks, like many computational powers. An AGI at human level intelligence would actually be as capable of hundreds if not thousands of perfectly synchronous working humans of that intelligence.

To create such an intelligence you do not code every single bit, but you write learning programs and give it tasks that will teach it something. The problem is that you aren't sure what it really learns, and how new information might change that information. Your AI might be incapable of killing any human, but after it was tasked with disaster relief and had learned it needed to make choices between which groups it would offered aid and which would be condemned to high death rates it could apply it elsewhere... And you'll never see it coming. It could decide that kicking a few hundredthousand people out of their homes and relocating them where a large portion will likely die just to build something where they lived.

This is one of the reasons why I think AGI's should all be put in a box and have a few more next to it. You feed the problem to the box, then first compare the answers, take the most common answer and analyse the hell out of it. Then either execute the answers manually or through a more crude AI that you can trust.