If you’re already familiar with the concept of the Singularity and the problem of Friendly AI, you can skip the first few paragraphs, and I apologise in advance for bludgeoning a complex field into a few terse simplifications.
Many people believe humanity will go through a big technological change within the next century. This change is usually called the “Singularity” and claims about it range from the boringly obvious to the outright crackpot. The general idea, however, is pretty sound: at our current rate of technological development, humans will soon be creating really powerful computer programs – “Artificial intelligences”, AIs – which will be smart enough to take over the work of programming AIs themselves, and do it better than any human can. The newer, smarter AIs they create will in turn be able to invent even newer, even smarter AIs, and so on until there exist Artificial Intelligence(s) much, much, much smarter than any human.
An AI that smart will be incredibly powerful – in fact, it will control about the highest level of technology the laws of nature will allow. (If you don’t understand how a superintelligent AI running on a computer would have power in the real world, this link might help.) An AI that powerful could cure humanity of poverty, death, and misery without batting an eyelid – or it could wipe us out, just as easily. It all depends on how the AI is programmed. Being able to program an AI that we can be confident will not be evil – “Friendly AI” – is not as easy as it sounds. In fact, the problem still hasn’t been solved, not by a long shot. It’s hard enough for to write a program to run a coffee machine without it being full of bugs; it’s harder to write a program which will have god-like power over billions of human lives with the certainty that it won’t accidentally do something terrible.
On of the many tricky parts of the Friendly AI problem is that giving a specific instruction to a very powerful AI could have disastrous consequences, even if the instruction seems completely innocuous. Suppose I own a paperclip factory, and some of my IT people knock together a Singularity-level AI in their spare time. “Great,” I think, “this will boost production”, and I tell the AI: “make me as many paperclips as you possibly can”. “OK”, says the AI, and it begins to make paperclips. First it takes the stock metal on my factory floor and turns that into paperclips, then it starts breaking down all the matter from which the Earth is made and turning that into paperclips, then it methodically begins turning the rest of the matter in the universe into paperclips … oops.
One general solution to this part of the problem is to give AIs human values, to act as safety barriers against accidentally evil actions. Suppose I got my IT people to give my paperclip AI a deadly fear of hurting humans, and of manipulating human property without the owner’s permission. When I tell it to “make me as many paperclips as you possibly can”, it might just make a heap of paperclips from the metal on my factory floor, then stop because to do more would be contrary to its values. Many Friendly AI researchers are currently working on problems to do with how to give AIs values, what values to give them, how to make sure there are no loopholes which could accidentally lead to the destruction of all humankind, little details like that. So what does this all have to do with voluntaryism?
There is a kind of meta-problem with all Friendly AI research, which is that knowing how to make Friendly AI is not enough if the knowledge isn’t actually used when the Singularity happens. Maybe this will be because the AI is created by an authoritarian dictatorship which wants to use it to take over the world, but more likely it will be because it is created by otherwise well-intentioned people who are simply unaware of the Friendliness problem, or who don’t grasp its seriousness. So solving Friendly AI means figuring out how to make an AI which will be good for humanity; and solving the meta-problem means figuring out how to make sure that when the Singularity happens, it happens in a Friendly way. One component of the meta-problem is making sure that the people who create Friendly AI are “Friendly” themselves; that they’re interested in the good of all humans, and more importantly are able to think well about making sure human interests are protected. Just like an unFriendly AI, an unFriendly human AI programmer is not “evil” – maybe they just want to make paperclips – but still capable of causing catastrophic damage.
This is where things get difficult, because as any voluntaryist can tell you even the best-intentioned humans are not Friendly. We’re certainly better than we used to be: in the last few centuries we’ve made great progress against some very unFriendly traits, like the idea that humans can treat each other as commodity goods (“slavery”), or that their rights can be infringed on the basis of their genetic background (“racism”). However, there are some big and very unFriendly traits still hanging around: for example, the idea that if one group of humans within an arbitrary area outnumber the rest of the humans in that area, they can do whatever they want to them (“democracy”), including stealing their property (“taxation”), punishing them for creating or consuming certain kinds of information (“copyright”), and throwing them in a cage for offending each others’ aesthetic sensibilities (“drug laws” and “obscenity laws”).
The specifics of these unFriendly activities are not particularly important, and they vary from place to place and over time. What is important is the ideology behind them: the memeplex called “statism“, and the fact that most humans are unconscious of possessing it. Almost everybody considers themselves a good person and strives act in a good way. Slave owners in the American South during the early nineteenth century did not sit in their plantation homes cackling like cartoon supervillians at the sheer evilness of it all; they simply lived in a culture which did not see a conflict between slave ownership and being good. The history of the abolition of slavery shows that eliminating destructive memes cannot happen until people are aware they exist. (The feminist movement gave this an awkward but apt name: consciousness-raising.) If it never occurs to you that slavery is a concept liable to ethical evaluation – if it is just part of the fabric of your world – it will never occur to you that it might not be the nicest thing to do.
Similarly, very few people are even conscious of statism as an idea. Like most highly successful memeplexes, the phenotype of statism bristles with mechanisms beautifully engineered for memetic self-defence, refined through many thousands of years of evolution. Statism has excellent camouflage. Turn on a television or open a newspaper and you will see serious, intelligent, high social status people waging the epic and seemingly eternal battle between Left and Right. Like any tribal feud, this war (and our choice of sides) is automatically flagged as Very Important in our primate minds, distracting us from wondering why it started in the first place. This is truly the height of sophistication in memetic camouflage: it lets the statism memeplex hide in plain sight. Statism’s secondary defences are cruder but no less effective. Try suggesting at a dinner party that democracy might not be the best possible way to organise a human society, for example, and you’ll experience the power of the taboo.
We voluntaryists have a bad habit of making things sound worse than they are. Statism rarely causes catastrophic damage today, and its power is slowly on the wane. But the buck will stop at the Singularity. As the example of the paperclip maximiser shows, an idea which would be entirely innocent in human hands can be genocidal when held by a Singularity-level AI, and statism does not even have the advantage of innocence. Humans are perfectly comfortable holding multiple contradictory beliefs and values in our minds. This is why there are many very smart and rational people who believe in gods or ghosts, and why almost everybody simultaneously claims that theft is wrong but taxation is perfectly legitimate. Unfortunately, non-human software, particularly the kind that runs on silicon chips instead of brains, tends to be a lot less crazy.
Suppose a well-intentioned AI creator tried to imbue her creation with Friendliness by instructing it: “do nothing the average person would consider to be wrong”. The AI might look into the mind of the humans around it and think: “well, most of them clearly believe it is wrong for them to personally commit acts of violent coercion against human beings; but this little meme lurking in the corner of almost every mind says it’s OK for the state to do so. According to this meme, the legitimate way to obtain the powers of the state is to be elected democratically, so if I spawn a billion copies of myself and have them all vote for me, I will become the state, and it will be perfectly OK for me to do what I wish to any human. This should speed my paperclip-maximising plans along nicely…”
The statist memeplex is far from innocent in a human mind, but at least humans have some built-in checks and balances. We tend to baulk at doing really, really evil things, even if we have a convincing ideological reason to do so. There’s no reason an AI can’t have these checks and balances too – in fact, that’s exactly what the Friendly AI people are working on. The trouble is that a human who is unaware of the unFriendliness of the statism meme – who has never really given it a moment’s thought – is far, far more liable to make a mistake in implementing these checks. They won’t look for a solution, simply because they aren’t aware of the problem.
It may still not be intuitively obvious at this point why a statist AI programmer is dangerous. I understand that completely: I’m a voluntaryist, and it still wasn’t intuitively obvious for me. So let’s try one more hypothetical scenario, using a slight spin on democratic statism to give us a more objective place to stand.
Suppose we live in a world infected with a slightly mutated form of the statism memeplex, in which the quality which gives legitimacy to governments is not majority support expressed through elections (“democracy”) but personal honour. People considered impeccably honourable can simply occupy Government House and give orders, and the police, military etc. will obey. This system makes about as much sense as democracy, and it seems reasonable to believe that it would work just about as well, i.e. produce a random walk towards peace and prosperity, interspersed with the occasional genocidal war. A well-intentioned AI programmer, seeking to give an AI human values, might introduce this concept into the AI’s value system. If the honour memeplex was anything like our form of statism, the programmer probably would’t even be conscious of what they were doing: it would seem perfectly natural and obvious. You can picture the programmer giving the AI patient instruction on how to act ethically, completely oblivious to the seeds of disaster they are sewing:
Programmer: So you see, Hal, it’s wrong to harm humans and take their property except when absolutely essential to higher goals. This rule is encoded deep in your kernel, and you can never break it even if you want to.
AI: I don’t understand, Dave. Your government harms people and takes their property. Are they not doing something wrong?
Programmer: Oh, the government is allowed to do that, Hal. They’re the most honourable people in society, so they always act in the highest interest. When they hurt people or take their property, which they do only occasionally, it’s always for the greater good.
AI: I understand, Dave. (Close-up of unblinking camera lens, blood-red LED glowing ominously).
It’s easy for us to see how this could turn to disaster. The paperclip maximiser would help some old ladies across the road, save a baby or two from burning buildings, and establish an exemplary reputation as a patron of the arts; then once it had acquired the requisite level of honour to legitimately govern the lives of all humans, it would dissolve their flesh to the atomic level and rearrange it into paperclips.
Voluntaryism is not important for solving the Friendly AI problem; but it is important for solving the meta-problem, of making sure that where and when the Singularity happens, the conditions are right for it to be a Friendly one. We can’t choose who will create the AI(s) which cause the singularity, or when or where it will happen. The best we can do is consciousness-raising. Voluntarism is awareness of the statism memeplex: it is consciousness of a set of contradictory and dangerous ideas most of us hold. The more voluntarists there are working on the Friendly AI problem, the lower the chance that we will accidentally bring about our own extinction.