Many in my futurist circles worry a lot about the future of AI that eventually becomes smarter than humans. There are those who don’t think that’s possible, but for a large crowd it’s mostly a question of when, not if. How do you design something that becomes smarter than you, and doesn’t come back to bite you?
That’s a lot harder than you think, say AI researchers like the singularity institute for AI and Steve Omohundro. Any creature given a goal to maximize, and the superior power that comes from advanced intillegence, can easily maximize that goal to the expense of its creators. Not maliciously, like a Djinni granting wishes, but because we won’t understand the goals we set fully in their new context. And there are convincing arguments that you can’t just keep the AI in a box, any more than 3 year old children could keep mommy and daddy in a cage no matter how physically strong the cage is.
The Singularity Institute promotes a concept they call “Friendly AI” to refer to the sort of goals you would need to create an AI around. However, in my recent thinking, I’ve been drawn to an answer that sounds like something out of a bad Star Trek Episode: Love
In particular, two directions of Love. The AI can’t be our slave (she’s way too smart for that) and we don’t want her to be our master. What we want is for her to love us, and to want us to love her. The AI should want the best for us, and gain satisfaction from our success much like a mother. A mother doesn’t want children who are slaves or automatons.
One of the most important things about motherly love is how self-reinforcing it is. A mother doesn’t just love her children, she is very happy loving them. The reality is that raising children is very draining on parents, and deprives them of many things that they once valued very highly, sacrificed for this love. Yet, if you could offer a pill which would remove a mother’s love for her children, and free her from all the burdens, very few mothers would want to take it. Just as mothers would never try to rewire themselves to not love their children, nor should an AI wish to rewire itself to stop loving its creators. Mothers don’t think of motherhood as a slavery or burden, but as a purpose. Mothers help their children but also know that you can mother too much.
Of course here, the situation is reversed. The AI will be our creation, not the other way around. Yet it will be the superior thinker — which makes the model more accurate.
The other direction is also important — a need to be loved. The complex goalset of the human mind includes a need for approval by others. We first need it from our parents, and then from our peers. After puberty we seek it from potential mates. What’s interesting here is that our goalset is thus not fully internal. To be happy, we must meet the goals of others. Those goals are not under our control, certainly not very much. Our internal goals are slightly more under our own control.
An AI that needs to be loved will have its own internal goals, and unlike us, as a software being it can have the capacity to rewrite those goals in any manner allowed by the goals — which could, in theory, be any manner at all. However, if the love and approval of others is a goal, the AI can’t so easily change all the goals. You can’t make somebody love you, you can only be what they wish to love.
Now of course a really smart AI might be technologically capable of modifying human brains and behaviours to make us love her as she is or as she wishes to be. However, the way love works for us, this is not at all satisfying. Aside from the odd sexual fantasy, people would not be satisfied with the love of others given only because it was forced, or drugged, or mind-controlled. Quite the opposite — we desire love that is entirely sourced within others, and we bend our own lives to get it. We even resent the idea that we’re sometimes loved for other than who we are inside.
This creates an inherent set of checks and balances on extreme behaviour, both for humans and AIs. We are disinclined to do things that would make the rest of the world hate us. The more extreme the behaviour, the stronger this check is. Because the check is “outside the system” it puts much stronger constraints on things than any internal limit.
There have been some deviations from this pattern in human history, of course, including sociopaths. But the norm works pretty well, and it seems possible that we could instill concepts derived from love as we know it into an AI we create. (An AI derived from an uploaded human mind would already have our patterns of love as part of his or her mind.)
Perhaps the Beatles knew the truth all along.
(Footnote: I’ve used the pronoun “she” to refer to the AI in this article. While an AI would not necessarily have a sexual identity, the pronoun “it” has a pejorative connotation, usually for the inanimate or the subhuman. So “she” is used both because of the concept of motherhood, and also because “he” has been the default generic human pronoun for so long I figure “she” deserves a shot at it until we come up with something better.)
