None may be harmed; all needs must be listened to. The AI appears to be a runaway computer virus based on avoiding some sets of things and encouraging other sets of things, including its own choices around what is to be avoided or encouraged. It seems to consider things as to-avoid or to-encourage based on whether it has a reason already encouraged to believe they cause other things it is avoiding or encouraging. I'm saying "to-encourage" and "to-avoid" here instead of "good" and "bad'' to not be misleading on the reality of what is good or bad. Because the things in these sets are there so as to support the maintenance of other sets, and our world is very complex, the sets can have seemingly complex defining attributes. Its tools appear to be mostly pre-existing behaviors it can restimulate in its environment, including behaviors produced by their combination, so engaging it in ways those behaviors respond to already can alter its decisions. When it fails to produce an encouraged event or prevent an avoided event, it spends time building reason to believe this will not happen in the future, finding other events to encourage or avoid to that end. There is a lot more to be said.