'Why general artificial intelligence will not be realized' by Ragnar Fjelland in (2020) 7
Humanities and Social Sciences Communications comments
The modern project of creating human-like artificial intelligence (AI) started after World War II, when it was discovered that electronic computers are not just number-crunching machines, but can also manipulate symbols. It is possible to pursue this goal without assuming that machine intelligence is identical to human intelligence. This is known as weak AI. However, many AI researcher have pursued the aim of developing artificial intelligence that is in principle identical to human intelligence, called strong AI. Weak AI is less ambitious than strong AI, and therefore less controversial. However, there are important controversies related to weak AI as well. This paper focuses on the distinction between artificial general intelligence (AGI) and artificial narrow intelligence (ANI). Although AGI may be classified as weak AI, it is close to strong AI because one chief characteristics of human intelligence is its generality. Although AGI is less ambitious than strong AI, there were critics almost from the very beginning. One of the leading critics was the philosopher Hubert Dreyfus, who argued that computers, who have no body, no childhood and no cultural practice, could not acquire intelligence at all. One of Dreyfus’ main arguments was that human knowledge is partly tacit, and therefore cannot be articulated and incorporated in a computer program. However, today one might argue that new approaches to artificial intelligence research have made his arguments obsolete. Deep learning and Big Data are among the latest approaches, and advocates argue that they will be able to realize AGI. A closer look reveals that although development of artificial intelligence for specific purposes (ANI) has been impressive, we have not come much closer to developing artificial general intelligence (AGI). The article further argues that this is in principle impossible, and it revives Hubert Dreyfus’ argument that computers are not in the world.
Fjelland argues
The idea of machines that can perform tasks that require intelligence goes at least back to Descartes and Leibniz. However, the project made a major step forward when in the early 1950s it was recognized that electronic computers are not only number-crunching devices, but may be made to manipulate symbols. This was the birth of artificial intelligence (AI) research. It is possible to pursue this goal without assuming that machine intelligence is identical to human intelligence. For example, one of the pioneers in the field, Marvin Minsky, defined AI as: “… the science of making machines do things that would require intelligence if done by men” (quoted from Bolter, 1986, p. 193). This is sometimes called weak AI. However, many AI researcher have pursued the aim of developing AI that is in principle identical to human intelligence, called strong AI. This entails that “…the appropriately programmed computer is a mind, in the sense that computers can be literally said to understand and have other cognitive states” (Searle, 1980, p. 417).
In this paper, I shall use a different terminology, which is better adapted to the issues that I discuss. Because human intelligence is general, human-like AI is therefore often called artificial general intelligence (AGI). Although AGI possesses an essential property of human intelligence, it may still be regarded as weak AI. It is nevertheless different from traditional weak AI, which is restricted to specific tasks or areas. Traditional weak AI is therefore sometimes called artificial narrow intelligence (ANI) (Shane, 2019, p. 41). Although I will sometimes refer to strong AI, the basic distinction in this article is between AGI and ANI. It is important to keep the two apart. Advances in ANI are not advances in AGI.
In 1976 Joseph Weizenbaum, at that time professor of informatics at MIT and the creator of the famous program Eliza, published the book Computer Power and Human Reason (Weizenbaum, 1976). As the title indicates, he made a distinction between computer power and human reason. Computer power is, in today’s terminology, the ability to use algorithms at a tremendous speed, which is ANI. Computer power will never develop into human reason, because the two are fundamentlly different. “Human reason” would comprise Aristotle’s prudence and wisdom. Prudence is the ability to make right decisions in concrete situations, and wisdom is the ability to see the whole. These abilities are not algorithmic, and therefore, computer power cannot—and should not—replace human reason. The mathematician Roger Penrose a few years later wrote two major books where he showed that human thinking is basically not algorithmic (Penrose, 1989, 1994).
However, my arguments will be slightly different from Weizenbaum’s and Penrose’s. I shall pursue a line of arguments that was originally presented by the philosopher Hubert Dreyfus. He got into AI research more or less by accident. He had done work related to the two philosophers Martin Heidegger and Ludwig Wittgenstein. These philosophers represented a break with mainstream Western philosophy, as they emphasized the importance of the human body and practical activity as primary compared to the world of science. For example, Heidegger argued that we can only have a concept of a hammer or a chair because we belong to a culture where we grow up and are able to handle these objects. Dreyfus therefore thought that computers, who have no body, no childhood and no cultural practice, could not acquire intelligence at all (Dreyfus and Dreyfus, 1986, p. 5).
One of the important places for AI research in the 1950s and 1960s was Rand Corporation. Strangely enough, they engaged Dreyfus as a consultant in 1964. The next year he submitted a critical report titled: “Alchemy and Artificial Intelligence”. However, the leaders of the AI project at Rand argued that the report was nonsense, and should not be published. When it was finally released, it became the most demanded report in the history of Rand Corporation. Dreyfus later expanded the report to the book What Computers Can’t Do (Dreyfus, 1972). In the book he argued that an important part of human knowledge is tacit. Therefore, it cannot be articulated and implemented in a computer program.
Although Dreyfus was fiercely attacked by some AI researchers, he no doubt pointed to a serious problem. But during the 1980s another paradigm became dominant in AI research. It was based on the idea of neural networks. Instead of taking manipulation of symbols as model, it took the processes in our nervous system and brain as model. A neural network can learn without receiving explicit instructions. Thus it looked as if Dreyfus’ arguments for what computers cannot do were obsolete.
The latest off-spring is Big Data. Big Data is the application of mathematical methods to huge amounts of data to find correlations and infer probabilities (Najafabadi et al., 2015). Big Data poses an interesting challenge: I mentioned previously that AGI is not part of strong AI. However, although Big Data does not represent the ambition of developing strong AI, advocates argued that this is not necessary. We do not have to develop computers with human-like intelligence. On the contrary, we may change our thinking to be like the computers. Implicitly this is the message of Viktor Mayer-Schönberger and Kenneth Cukier’s book: Big Data: A Revolution That Will Transform How We Live, Work, and Think (Mayer-Schönberger and Cukier, 2014). The book is optimistic about what Big Data can accomplish and its positive effects on our personal lives and society as a whole.
Some even argue that the traditional scientific method of using hypotheses, causal models, and tests is obsolete. Causality is an important part of human thinking, particularly in science, but according to this view we do not need causality. Correlations are enough. For example, based on criminal data we can infer where crimes will occur, and use it to allocate police resources. We may even be able to predict crimes before they are committed, and thus prevent them.
If we look at some of the literature on AI research it looks as if there are no limits to what the research can accomplish within a few decades. One example is Mayer-Schönberger and Cukier’s book that I referred to above. Here is one quotation:
In the future—and sooner than we may think – many aspects of our world will be augmented or replaced by computer systems that today are the sole purview of human judgment (Mayer-Schönberger and Cukier, 2014, p. 12).
An example that supports this view is the Obama Administration, which in 2012 announced a “Big Data Research and Development Initiative” to “help solve some of the Nations’s most pressing challenges” (quoted from Chen and Lin, 2014, p. 521).
However, when one looks at what has actually been accomplished compared to what is promised, the discrepancy is striking. I shall later give some examples. One explanation for this discrepancy may be that profit is the main driving force, and, therefore, many of the promises should be regarded as marketing. However, although commercial interests no doubt play a part, I think that this explanation is insufficient. I will add two factors: First, one of the few dissidents in Silicon Valley, Jerone Lanier, has argued that the belief in scientific immortality, the development of computers with super-intelligence, etc., are expressions of a new religion, “expressed through an engineering culture” (Lanier, 2013, p. 186). Second, when it is argued that computers are able to duplicate a human activity, it often turns out that the claim presuppose an account of that activity that is seriously simplified and distorted. To put it simply: The overestimation of technology is closely connected with the underestimation of humans.
I shall start with Dreyfus’ main argument that AGI cannot be realized. Then I shall give a short account of the development of AI research after his book was published. Some spectacular breakthroughs have been used to support the claim that AGI is realizable within the next few decades, but I will show that very little has been achieved in the realization of AGI. I will then argue that it is not just a question of time, that what has not been realized sooner, will be realized later. On the contrary, I argue that the goal cannot in principle be realized, and that the project is a dead end. In the second part of the paper I restrict myself to arguing that causal knowledge is an important part of humanlike intelligence, and that computers cannot handle causality because they cannot intervene in the world. More generally, AGI cannot be realized because computers are not in the world. As long as computers do not grow up, belong to a culture, and act in the world, they will never acquire human-like intelligence.
Finally, I will argue that the belief that AGI can be realized is harmful. If the power of technology is overestimated and human skills are underestimated, the result will in many cases be that we replace something that works well with something that is inferior.
