Even if it is possible to 'tether' AIs, won't we just remove those limits? Look what we do to fake persons today - corporations. They are like AIs in that small calculations made throughout the entity guide the decisions as a whole.
We cyclicly add boundaries to what they're allowed, then the next administration removes those boundaries in the interest of unbridled growth.
We'll do the same with AIs. They'll calculate a 'best outcome' of letting millions die in Indonesian drought, because average standard of living will increase for everyone else. In one administration we'll want to temper such decisions with humanity; in the next we'll just want more beer.