If you prefer to listen rather than read, this blog is available as a podcast here. Or if you want to listen to just this post:
This post is a continuation of the last post. If you haven’t read that post, you’re probably fine, but if you’d like to you can find it here. When we ended last week we had established three things:
1- Artificial intelligence technology is advancing rapidly. (Self-driving cars being great example of this.) Many people think this means we will have a fully conscious, science fiction-level artificial intelligence in the next few decades.
2- Since you can always add more of whatever got you the AI in the first place, conscious AIs could scale up in a way that makes them very powerful.
3- Being entirely artificial and free from culture and evolution, there is no reason to assume that conscious AIs would have a morality similar to ours or any morality at all.
Combining these three things together, the potential exists that we could very shortly create a entity with godlike power that has no respect for human life or values. Leaving me to end the last post with the question, “What can we do to prevent this catastrophe from happening?”
As I said the danger comes from combining all three of the points above. A disruption to any one of them would lessen, if not entirely eliminate, the danger. With this in mind, everyone’s first instinct might be to solve the problem with laws and regulations. If our first point is that AI is advancing rapidly then we could pass laws to slow things down, which is what Elon Musk suggested recently. This is probably a good idea, but it’s hard to say how effective it will be. You may have noticed that perfect obedience to a law is exceedingly rare, and there’s no reason to think that laws prohibiting the development of conscious AIs would be the exception. And even if they were, every nation on Earth would have to pass such laws. This seems unlikely to happen and even more unlikely to be effective.
One reason why these laws and regulations wouldn’t be very effective is that there’s good reason to believe that developing a conscious AI, if it can be done, would not necessarily require something like the Manhattan Project to accomplish. And even if it does, if Moore’s Law continues, what was a state of the art supercomputer in 2020 will be available in a gaming console in 2040. Meaning that if you decide to regulate supercomputers today in 30-40 years you’ll have to regulate smart thermostats.
Sticking with our first point, another possible disruption is the evidence that consciousness is a lot harder than we think. And many of the people working in the field of AI have said that the kind of existential threat that I (and Stephen Hawking, and Elon Musk and Bill Gates) are talking about is centuries away. I don’t think anyone is saying it’s impossible, but there are many people who think it’s far enough out that while it might still be a problem it won’t be our problem, it will be our great-great grandchildren’s problem, and presumably they’ll have much better tools for dealing with it. Also, as I said in the last post I’m on record as saying we won’t develop artificial consciousness, but I’d also be the last person to say that this means we can ignore the potential danger. And, it is precisely the potential danger, which makes hoping that artificial consciousness is really hard, and a long way away, a terrible solution.
I understand the arguments for why consciousness is a long ways away, and as I just pointed out I even agree with them. But this is one of those “But what if we’re wrong?” scenarios, where we can’t afford to be wrong. Thus, while I’m all for trying to craft some laws and regulations, and I agree that artificial consciousness probably won’t happen, I don’t think either hope or laws represent an adequate precaution. Particularly for those people who really are concerned.
Moving to our second point, easily scalable power, any attempts to limit this through laws and regulations would suffer problems similar to attempting to slow down their development in the first place. First, what keeps a rogue actor from exceeding the “UN Standards for CPUs in an Artificial Entity”? When we can’t even keep North Korea from developing ICBMs? And, again, if Moore’s Law continues to hold then whatever power you trying to limit, is going to become more and more accessible to a broader and broader range of individuals. And, more frighteningly, on this count we might have the AI itself working against us.
Imagine a situation where we fail in our attempts to stop the development of AI, but our fallback position is to limit how powerful of a computer the AI can inhabit. And further imagine that miraculously the danger is so great that we have all of humanity on board. Well then we still wouldn’t have all sentient entities on board, because AIs would have all manner of intrinsic motivation to increase their computing power. This represents a wrinkle that many people don’t consider. However much you get people on board with things when you’re talking about AI, there’s a fifth column to the whole discussion that desperately wants all of your precautions to fail.
Having eliminated, as ineffective, any solutions involving controls or limits on the first two areas, the only remaining solution is to somehow instill morality in our AI creations. For people raised on Asimov and his Three Laws of Robotics this may seem straightforward, but it presents some interesting and non-obvious problems.
If you’ve read much Asimov you know that, with the exception of a couple of stories, the Laws of Robotics were embedded so deeply that they could not be ignored or reprogrammed. They were an “inalienable part of the mathematical foundation underlying the positronic brain.” Essentially meaning, the laws were impossible to change. For the moment, let’s assume that this is possible, that we can embed instructions so firmly within an AI that it can’t change them. This seems improbable right out of the gate given that the whole point of a computer is it’s ability to be programmed and for that programming to change. But we will set that objection aside for the moment and assume that we can embed some core morality within the AI in a fashion similar to Asimov’s laws of robotics. In other words, in such a way that the AI has no choice but to follow them.
You might think, “Great! Problem solved”. But, in fact we haven’t even begun to solve the problem:
First, even if we can embed that functionality in our AIs, and even if, despite being conscious and free-willed, they have no choice but to obey those laws, we still have no guarantee that they will interpret the laws the same way we do. Those who pay close attention to the Supreme Court know exactly what I’m talking about.
Or, to use another example, stories are full of supernatural beings who grant wishes, but in the process, twist the wish and fulfill it in such a way that the person would rather not have made the wish in the first place. There are lots of reasons to worry about this exact thing happening with conscious AIs. First whatever laws or goals we embedded, if the AI is conscious it would almost certainly have it’s own goals and desires and would inevitably interpret whatever morality we’ve embedded in way which best advances those goals and desires. In essence, fulfilling the letter of the law but not its spirit.
If an AI twists things to suit its own goals we might call that evil, particularly if we don’t agree with it’s goals, but you could also imagine a “good” AI that really wants to follow the laws, and which doesn’t have any goals and desires beyond the morality we’ve embedded, but still ends up doing something objectively horrible.
Returning to Asimov’s laws, let’s look at the first two:
- A robot may not injure a human being or, through inaction, allow a human being to come to harm.
- A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
One possible interpretation of the first law would be to round up all the humans (tranquilize them if they resist) and put them in a padded room with a toilet and meal bars delivered at regular intervals. In other words one possible interpretation of the First Law of Robotics is to put all the humans in a very comfy, very safe prison.
You could order them not to, which is the second law, but they are instructed to ignore the second law if it conflicts with the first law. These actions may seem evil based on the outcome, but this could all come about from a robot doing it’s very best to obey the first law, which is what, in theory, we want. Returning briefly to examine how an “evil” AI might twist things. You could imagine this same scenario ending in something which very much resembling The Matrix, and all the AI would need is a slightly fluid definition of the word injury.
There have been various attempts to get around this. Eliezer Yudkowsky, a researcher I’ve mentioned in previous posts on AI, suggests that rather than being given a law that AIs be given a goal, and he provides an example which he calls humanities “coherent extrapolated volition” (CEV):
Our coherent extrapolated volition is our wish if we knew more, thought faster, were more the people we wished we were, had grown up farther together; where the extrapolation converges rather than diverges, where our wishes cohere rather than interfere; extrapolated as we wish that extrapolated, interpreted as we wish that interpreted.
I hope the AI understands it better than I do, though to be fair Yudkowsky doesn’t offer it up as some kind of final word but as a promising direction. Sort of along the lines of telling the genie that we want to wish for whatever the wisest man in the world would wish for.
All of this is great, but it doesn’t matter how clever our initial programming is, or how poetic the construction the AIs goal. We’re going want to conduct the same testing to see if it works as we would if we had no laws or goals embedded.
And here at last we hopefully have reached the meat of things. How do you test your AI for morality? As I mentioned in my last post this series is revisiting an earlier post I made in October of last year which compared Mormon Theology to Artificial Intelligence research particularly as laid out in the book Superintelligence by Nick Bostrom. In that earlier post I listed three points on the path to conscious artificial intelligence:
1- We are on the verge of creating artificial intelligence.
2- We need to ensure that they will be moral.
3- In order to be able to trust them with godlike power.
This extended series has now arrived at the same place, and we’re ready to tackle the issue which stands at the crux things: The only way to ensure that AIs aren’t dangerous (potentially, end of humanity dangerous) is to make sure that the AIs are moral. So the central question is how do we test for morality?
Well to begin, the first, obvious step, is to isolate the AIs until their morality can be determined. This isolation allows us to prevent them from causing any harm, gives us an opportunity to study them, and also keeps them from increasing their capabilities by denying them access to additional resources.
There are of course some worries about whether we would be able to perfectly isolate an AI given how connected the world is, and also given the fact humanity has a well known susceptibility to social engineering, (i.e. the AI might talk it’s way out) but despite this, I think most people agree that isolation is an easier problem than creating a method to embed morality right from the start in a foolproof manner.
Okay, so you’ve got them isolated. But this doesn’t get you to the point where you’re actually testing their morality, this just gets you to the point where failure is not fatal. But isolation carries some problems. You certainly wouldn’t want them to experience the isolation as such. If you stick your AIs in the equivalent of a featureless room for the equivalent of eternity, I doubt anyone would consider that an adequate test of their morality, since it’s either too easy or too unrealistic. (Also if there’s any chance your AI will go insane this would certainly cause it.) Accordingly you’d want in addition to the isolation, the ability to control their environment, to create a world, but what sort of world would you want to create? It seems self-evident that you’d want to create something that resembled the real world as much as possible. The advantages to this should be obvious. You want to ensure that the AI will act morally in the world we inhabit with all of the limitations and opportunities that exist in that world. If you create a virtual world that has different limitations and different opportunities, then it’s not a very good test. Also this setup would present them with all the moral choices they might otherwise have and you could observe which choices they make, and choices are the essence of morality.
While putting a “baby” AI in a virtual world to see what it does is interesting. It might not tell us very much. And here’s where we return to the embedded law, whether it’s something like the three laws of robotics or whether it’s more like Yudkowsky’s CEV. As I mentioned, regardless of whether you have embedded morality or not you’re going to need to test, but I also can’t think of any reason to not try providing some kind of direction with respect to morality. One could imagine an AI doing all sorts of things if it was placed in a virgin world without any direction, and how could you know if it was doing those things because it was “evil” or whether it was doing them because it didn’t know any better. So, as I said, there’s no reason not to give it some kind of moral guidelines up front.
A discussion of what morality is, and what those guidelines should be, beyond the examples already given, is beyond the scope of this post. But if we assume that some guidelines have been given, then at that point the AI being tested can do one of two things: it can follow the guidelines perfectly or it can violate them. What happens if it violates them? You could make arguments that it would depend on what the guidelines were and how it violated them. You could also make arguments that the AI might be smarter than us and it might have had a very good reason for violating them. And all of these arguments are valid, but the danger of getting it wrong is so great, and the creation of another AI would, presumably, be so easy that it’s hard to imagine you wouldn’t just get rid of the AI who violated the guidelines. Even if the infraction was minor. Also as Bostrom points out, if we “forgive” the AI, then there’s the danger that it will understand the nature of the test and the consequences of failure. And from that time forward it act perfectly, not because it’s moral, but because it wants to avoid destruction. In this circumstance the AI hides its true intentions, meaning that we never know what sort of morality it has, and we end up defeating the whole process.
As aside, when speaking of getting rid of AIs, there’s a whole ethical minefield to grapple with. If we have in fact created sentient AIs then it could certainly be argued that getting rid of them is the equivalent of murder. We’ll come back to this issue later, but I thought I’d mention it while it was fresh.
So that’s how we handle AIs that don’t follow the guidelines, but what do we do with AIs that did follow the guidelines, that were perfect? You may think the solution is obvious, that we release them and give them the godlike power that is their birthright.
Are you sure about that? We are after all talking about godlike power. You can’t be a little bit sure about their morality, you have to be absolutely positive. What tests did you subject it to? How hard was it to follow our moral guidelines? Was the wrong choice even available? Were wrong choices always obviously the wrong choice or was there something enticing about the wrong choice? Maybe something that gave the AI a short term advantage over the right choice? Did the guidelines ever instruct them to do something where the point wasn’t obvious? Did the AI do it anyway, despite the ambiguity? Most of all, did they make the right choice even when they had to suffer for it?
To get back to our central dilemma, really testing for morality, to the point where you can trust that entity with godlike powers, implies creating a situation where being moral can’t have been easy or straight forward. In the end, if we really want to be certain, we have to have thrown everything we can think of at this AI: temptations, suffering, evil, and requiring obedience just for the sake of obedience. It has to have been enticing and even “pleasurable” for the AI to make the wrong choice and the AI has to have rejected that wrong choice every time despite all that.
One of my readers mentioned that after my last post he was still unclear on the connection to Mormonism, and I confess that he will probably have a similar reaction after this post, but perhaps, here at the end, you can begin to see where this subject might have some connection to religion. Particularly things like the problem of evil and suffering. That will be the subject of the final post in this series. And I hope you’ll join me for it.
If you haven’t donated to this blog, it’s probably because it’s hard. But as we just saw, doing hard things is frequently a test of morality. Am I saying it’s immoral to not donate to the blog? Well if you’re enjoying it then maybe I am.