I’m known for predicting that later this century, there will be a terrible war over the issue of species dominance. More specifically, it will be fought over whether humans should build artilects (artificial intellects), which could become so vastly superior to human beings in intellectual capacity that they may end up treating us as grossly inferior pests, wiping us out. I anticipate billions of casualties resulting from the conflict over the artilect question.
To combat this horrible scenario, the Singularity Institute in Silicon Valley has been set up to ensure that the above scenario does not occur. The Institute’s principal theorist, Eliezer Yudkowsky, has coined the term “Friendly AI,” which he defines essentially as intelligent machines designed to remain friendly to humans, even as they modify themselves to attain higher levels of intelligence.
He claims (correctly, I think) that trying to prevent artilects from wiping out humanity is the most important thing on our agenda this century. Yudkowsky hopes that he and others will be able to mathematically prove that it is possible to design an intelligent machine that (of logical mathematical necessity, given its design) will be forced to remain human-friendly, even as it redesigns itself for increasing levels of intelligence.
I will present a set of arguments which I think refute this vision, and then comment on the political desirability (or otherwise) of this vision.
The Arguments Against Friendly AI
Let me start by assuming that Friendly AI can be implemented. The next question is whether humanity would be unanimous about implementing it. In other words: Does can imply ought?
I think that the more ardent of the Cosmists (the ideological group which wants to build artilects, viewing themselves as god builders) will argue that their main goal is to build artilects that are trillions of trillions of times above human intelligence levels – immortal creations thinking a million times faster than humans which possess unlimited memory, the ability to alter their shape and architecture in milliseconds and could venture out into the cosmos. These Cosmists would prefer that the artilects be built even if humans get wiped out in the process. If making them according to Friendly AI designs inhibits or even blocks their path to achieving godlike capabilities, then the Cosmists will want the artilects not to be made to be friendly to humans.
So even if Friendly AI designs can be created, it does not automatically follow that they will be universally applied. The more ardent Cosmists might go underground to build the artilects the way they want, ignoring the consequences for humanity. The Cosmists have a slogan: “One artilect is worth a trillion trillion human beings!”
On the other hand, if Friendly AI designs are impossible to make, then there is no point in discussing whether they should be implemented or not.
I will now present some arguments which claim to show that the notion of Friendly AI is impossible.
The Evolutionary Engineering Argument
Ask yourself how it’s possible for a creature of a given intelligence level to be able to design a creature of greater intelligence. Designing a creature of superior intelligence requires a level of intelligence that the designer simply does not have. Therefore, it is logically impossible to use the traditional blueprint-design approach to create a creature of superior intelligence.
For example, my good friend Ben Goertzel has recently written a book called “Building Better Minds” in which he lays out a humanly conceived (i.e. by himself) plan to build a near-human intelligence. He will only be able to go so far with such an approach. There will be limits to the ingenuity of his plan due to Goertzel’s own intellectual limits. So how can such limits be overcome?
Humans have been building superior intelligences for thousands of generations by having sex. Children often grow up to be smarter than their parents. How does this work? The answer is shuffling genes. When the genes of parents mix, and only one of each mother/father pair of genes is used, blind luck offers the chance of arriving at a DNA blueprint that builds an intellectually superior child. But there are limits to this process as well. It gets statistically harder and harder to generate ever higher intelligence. For example, the odds of creating a child as intellectually outstanding as noted American physicist Edward Witten are one in a billion.
So, how did modern Homo sapiens come into being? How did nature build us over millions of years? It did so by using evolutionary engineering, selecting genes with superior fitness levels arising from random mutations of DNA. This slow, blind process has resulted in us, and is very probably the only approach humans will have to build machines a lot smarter than we are.
But if we use evolutionary engineering to build artificial neural networks for our artilects, then the complexity levels of these networks will be so great that we are unable to understand them. They will be a black box.
One of the reasons I stopped my brain-building work was that I got bored evolving neural net modules for artificial brains. These modules were a black box to me. They worked because they were evolved, but I had no scientific understanding as to why they worked. I was doing great engineering but lousy science. After 20 years of it, I finally got fed up and turned to other research topics that taxed my own biological human brain more, such as pure math and mathematical physics.
Let us assume that the evolutionary engineering approach is the only way to create creatures of higher intelligence levels than human beings, and that the complexity levels of the evolved brain circuits are too complex for humans to understand. Then we would not be able to predict the attitudes and behavior of these creatures towards us. The only way to know how they would behave towards us would be to build them, but by then it’s too late. They would then exist and might choose to wipe us out.
This logic leaves us with a dilemma. Either we limit ourselves to humanly designed blueprints for intelligent machines that are incapable of reaching superhuman intelligence levels, or we use an evolutionary engineering approach that could attain superhuman intelligence levels. If we use an evolutionary engineering approach, we cannot be sure that the resulting artilects would be human-friendly.
The Cosmic Ray Argument.
It is almost certain that the circuitry used to create intelligent machines will be nanotech-based. For example, building a near-human artificial brain that is not the size of a room will necessitate the use of nanoscale components. Even if Friendly AI nanocircuits could be built, they would then be subject to the random mutations generated by the impact of energetic cosmic rays, which zap the nanocircuits in random ways, potentially generating rogue artilects. Nature would be doing the same kind of evolutionary engineering as the human sort mentioned above. Since these mutations would be random, their consequences on the behavior and attitudes of the artilects towards human beings would be unpredictable. So even if the initial, unmutated nanocircuits could be made human-friendly, they would not stay that way.
The Naïve Asimov Argument
Science fiction writer Isaac Asimov is famous for his “Three Laws of Robotics,” which were intended to ensure that the robots in his stories remained human-friendly. His robots were not allowed to harm humans, nor allow humans to be harmed. We can imagine fairly readily that it is probably possible to program robots in a conventional way to behave like this, with the proviso that the robots are less intelligent than their human programmers. But once the robots become smarter than humans, they would be able to examine their circuitry, detect the humanly created parts and delete them if they wanted. Hence Asimov’s Three Laws cannot help us. They are naïve; forget Asimov.
Friendly AI is a Dangerous Delusion
Hopefully, the above arguments have convinced you that the notion of Friendly AI is a delusion. But why might it be seen as a dangerous delusion?
If the future politicians who have to decide whether to legislate or not against building artilects of superhuman intelligence believe that Friendly AI robots can be built, then they will be much more likely not to legislate against their construction. On the other hand, if they learn that the artificial brain-building community has a consensus view that Friendly AI is impossible, then they will be far more hesitant.
If Friendly AI is indeed impossible, then humanity has a much tougher choice to make: Do we build gods, or do we build our potential exterminators? Spelling this out, humanity will then be forced to choose between building godlike artilects and risking extinction, or not building artilect gods, and seeing humanity survive. The first option is specicide. The second option is deicide. This choice will be the toughest that humanity will ever have to make.
If the pro-Friendly AI people can persuade politicians in the coming decades to let them go ahead with artilect building on the assumption that Friendly AI is valid, then if it is not valid, it is a dangerous delusion. Politicians may end up giving the green light to artilect builders to construct artilects that were thought to be human-friendly, but which in reality turn against us and wipe us out.