The hope is for the AI to be as wise as the operator — not necessarily wise enough to avoid any bad outcomes at all. This is hopefully enough to avoid AI risk in particular, though of course the AI will have all sorts of opportunities to mess things up and cause damage (just as humans do).