The Weakness of Strong AI
From WouterBeek.com
The notions of strong and weak AI have, since their introduction by Searle in 1980, been used to characterize two decidedly different stances towards AI research. This article will show the notion of strong AI to be quite problematic in its own respect, but this does not – as will subsequently be shown – make the concept of weak AI any more probable. For both ideas seem to be based upon the same problematic presupposition, namely that of a sufficiently clear notion of human intelligence and human action, that consequently (mis)guides the AI researcher's efforts.
Contents |
Strong and weak AI
The distinction between the strong and the weak program in AI was first defined in Searle's well-known article on the so-called `Chinese-room experiment', in which the strong program was described thus: “In strong AI, because the programmed computer has cognitive states, the programs are not mere tools that enable us to test psychological explanations; rather, the programs are themselves the explanations.” In contrast to this strong notion, the weak notion of AI would then amount to the supposition that an artificially intelligent agent can only act as if it is intelligent, emotional, or ethical; but it does not embody these very notions. An artificial agent will, according to the weak stance, always be a mere simulation of intelligent behavior.
So the traditional concept of 'strong AI' can be characterized as the ambition of the AI-researcher to not only provide the public with certain applications addressing a certain task, that in their design processes were brought about – at least partially – by a primal source of inspiration whose most basic impetus of motivation lay in an apt observation of the way in which human beings would have addressed the task at hand. Instead, the claim of the strong project in AI reaches much further, in their claiming to exist an (eventual) perfect similitude between machine and man. According to this latter view, knowledge of manhood does not only act as the primal source of ideas that allows one to think of the process of machine construction as somehow guided by a faint notion of how the human 'machine' would, under similar circumstances, have acted; but instead, to posit the eventually constructed machine as somehow identical – not necessarily on a material, but at least on a conceptual level – to the human individual (or, in multi-agent research efforts: identical to the gathering of human individuals within a community).
The idea of the strong program is thus that the construction of machinery provides us, in the end at least, with a full specification of the way in which the human mind is structured and comes to operate within the world.
The clearity of the focal notions of AI
That a clear specification of those concepts that guide AI research (like 'intelligence', 'emotion', etc.), is – scientifically speaking – still very far off, is something that is probably agreed on by the vast majority of AI-researches. This opinion was aptly expressed by John McCarthy as: “We cannot yet characterize in general what kinds of computational procedures we want to call intelligent.”
The problems of AI are casually formulated as the effort to reach artificial solutions that achieve the same level of complexity, promptitude, flexibility and veracity as human behavior exhibits. But although it might not be very difficult to reach an agreement on whether some artificial form of behavior coincides with some human form of behavior, for these things all lie at the very surface of our knowledge, it is quite hard to reach an agreement over the way in which the inner functioning of the system, that eventually comes to exhibit such humane behavior, is to be specifically realized.
For although it might be quite easy to model a computational agent that adheres to the behavioral pattern that we would like to call 'random', and it is equally quite easy to come up with an implementation of behavior that we would call entirely 'deterministic', it is – in sharp contrast to both of these easy efforts – not at all clear how something we would (casually) call 'free will', and the behavior that runs parallel to it, should be modeled. For it is clear that the notion of 'free will', as it can be observed in human individuals, hovers somewhere in between complete determinism and full randomness. But where exactly, i.e. at what position on the shifting scale between the extremes of determinism and randomness, this notion of determined randomness and random determination (which the human free will certainly is) lies, is still unknown.
It is thus clear that in saying that artificial agents should adhere to a steady and sacrosanct notion of human intelligence, we are positing an unspecified and completely fictional abstraction of the human mind. In our AI research we are not at all sure what it exactly is that a machine should eventually adhere to in order to sufficiently mimic a human agent. And thus the question of 'whether machines will ever be intelligent' can mean nothing to us, for the notion of human intelligence is not clear to us at all.
Technological determination
Let us investigate, for a moment, the feeling of the modern cognitive researcher, i.e. the feeling that the electric circuit is the ultimate metaphor and the ultimate realization of the human brain. For is this not just due to the determinateness of our times? We think of ourselves as machines, in the sense that we think of our minds as beings modeled in a similar way as a computational device is. Thus we think of our memories in terms of storage capacities (and of either longer- or shorter-term memories), of our perception system as an input device, and of our powers of ratiocination as the processes of a giant CPU. A thought in the head must be a certain representation that manifests itself in a certain regularity of electrical current, running on the wetware of the computational brain. And even the way in which we view our personal life's embedding within the fabric of social life is (partly) shaped after the image of machine intercommunication.
Technology, if not dictates, then at least heavily influences man's conception of the physical and metaphysical word, and also – as a consequence, or subset of this – technology comes to shape the conception of the self. So the coldness of the Renaissance universe, bereft of a immanent God or any other teleological purposeful entity, was shaped after the image of the mechanical clocks that were the epitome of the engineering efforts of the period. The steam engine proved to be a useful metaphor in Freud's psychiatric investigations, in order to characterize the various drifts that gather and mingle within the human soul and that must be let out in a certain way at a certain moment in order to maintain equilibrium and prevent the soul's 'kettle' from exploding. Also the transportation devices usually come to characterize the culture of an age. So the Dutch tugboat (or `trekschuit') seems paradigmatic for the pace of life in the eighteenth century. In a similar vein the time-spirit of the Industrial Revolution seems to be characterized most aptly by the steam-train. And so the modern age has come to be characterized by its automobiles and its planes, they somehow seem to express the empowerment of the individual, the hastiness of our lives, and the overcrowdedness of our habitat.
These technological images are of course used as metaphors in order to convey a certain idea. But at the same time these technologies are not mere metaphorical images that lie around in the writer's desk somewhere, in wait to be used for one or another apt characterization to be made. For these technologies surround us in our daily lives, they come to inhabit our worlds, and it is through these tools and transportation devices that we come to determine the way in which we live.
The modern scientist has to his or her disposal a computer for writing and searching through, for reading and comparing, research literature, along with various other tools that are used for retrieving, storing, and comparing various forms of data and the testing of hypothesis in simulated environments. This modern scientific habitat, then, is decidedly different from that of the ancient physicist who simply wandered through the woods in order to look at nature in a most direct manner. The only way in which the world fulfills a direct presence in the modern researcher's habitat, i.e. his or her university office, besides the steep once-were-trees staples of paper-work, is the potential presence of a dried-out plant on the desk. This is not to say that modern science has lost touch with the world when compared to earlier times, not at all, but it does hint at the way in which the concept of what the world is at all like has considerably changed through the course of time. This change is not a completely arbitrary and unguided one, but is to a very considerable extent determined by the technologies that have become the chief vehicles through which the scientist touches, and is touched by, the world.
Returning to the weak/strong-distinction
The problems of AI are in general being entirely misrepresented by the positing of a sacrosanct notion of 'humanity', in which concepts like 'intelligence', 'ethical behavior', etc. are thought to be absolute, steady, and unproblematic. Due to this entirely fictional notion, the success of future research is thought to consist of a more or less successful partial implementation of the mimicking of the thus posited absolutist concepts of human intelligence into artificial agents.
The question is not where on the measuring rod of the thought to be steady notion of the humane, ranging from being decidedly inhumane, unintelligent, exhibiting no sufficiently complex form of ethical behavior – via various stages of partial implementations of certain humane characteristics (weak AI) – to the other extreme, namely being a complete reproduction of the sacrosanct notion of humanity (the strong stance in AI), the machine model will fit in. For although the proponents ('strongists') and opponents ('weakists') to the fancied futurism of AI differ on the measure of compliance between machine and man that might eventually be reached, they both coincide in the presupposition of this stable, absolute and steady marked set of characteristics that are thought to constitute the humane.
But the more historically aware person might already feel a bit uneasy about this apparently universal presupposition, since do not these very notions, that are claimed to constitute the unshakable image of man, themselves come to change considerably? And are these changes not, to a very high degree at least, brought about by technological developments? But if technological development constantly changes the notion of humanity, then no notion of humanity (e.g. 'rationality', 'free will', 'ethical behavior') can ever be reached by using technological means. Not because technological means would not suffice (as many opponents of AI would have it), for they might as well do, but because the act of implementing a human characteristic makes us change the notion of humanity.
The teleology of AI research
What is the alternative view to both weak and strong AI then? For it is easy to show the deficits of certain ways of looking at our research field, but no scientist can hope to advance his or her scientific practice based upon negative knowledge only. I therefore would like to end this article by briefly sketching the outskirts of what should be the positive theorizing that would naturally arise from the negative critique that was given in the above.
We are continually confronted with the fact that we do not know exactly what is meant by words like 'intelligence', 'will', 'emotion', etc. And yet these notions do belong to the very core of that which is supposed to guide our research efforts. This critique, however justified it may be, should not drive us into the overreacting state of skepticism, in which we simply stop making any claims whatsoever about the teleology of our research program. Not only is such a brusque reaction not needed, such a skeptical stance is bound to lead to practical problems also. For it does not suffice to say “we make certain technologies possible” or “we develop programs that do certain things”. For one might ask “well, what things does it make possible?” and “in what way do these programs operate?”, i.e. “with respect to what measure can these solutions be claimed to be efficient, reliable, etc.?” Without a clear understanding of the intended end-product of scientific investigation (or engineering efforts), the process loses its measure of success. We therefore do need certain notions in order to drive our shared research project, otherwise it becomes an overcrowded forest of ideas in which everything, however outlandish, is allowed to grow.
The goal of AI might be formulated as follows: to build a useful agent just means to build an agent that does things the way we do them, whenever what we do is called 'useful' or 'intelligent'. But it must be reminded that by the very implementation of a certain portion of intelligent action, that very same action is rendered a lot less intelligent than it used to be, just because of the newly aquired possibility of its mechanical implementation. This can be seen in the downgraded esteem toward memorization since the invention of the bookpress, and – more recently – in the defeat of Kasparov in a game of chess.
The rationale here is that although we always seem to need, and indeed always seem to have, a certain knowledge of what we mean by notions such as 'intelligence', there is nothing that guarantees these notions to be steady. Moreover, there are clearly many different views of what characteristics humanity should consist of. And this is just to be expected, since decidedly different researchers are operating in decidedly different cultures and in decidedly different time periods.
And this is where things become rather different then they are according to the traditional weak/strong-distinction. For these central notions (like 'intelligence', etc.) that we indeed do use in our research, are not steady and absolute notions at all. They come to change together with the changes within our perception of the world and of the self. Thus the world that surrounds us, and for the scientist this also means the scientific world, determines to a very vast extent the goals that govern our own research effort.
Any full characterization of the situation in which our research field, when seen from a bird's eye view, unfolds itself, would therefore need to describe the way in which the theories we make are shaped by our ideas of humanness, but also – working at the same time – there is a profound influence the other way round, namely the one in which the technologies that we have at our disposal come to shape our ideas of what it means to be human. The schemata according to which the human psyche is explained in the research field of cognitive psychology can be taken as emblematic for this stance, for the abstractions in terms of which the human brain is described are almost exclusively based upon the machine model (a well-known example of this is the distinction between short and long term memory).
Conclusion
Some might say this conclusion is rather lame, in that it stays in the middle of the road, saying “this and that, both are the case”. But this is not at all to be interpreted as a safe or empty statement, for it is perfectly clear that without any technology at our disposal we would not be able to come to hold any idea of our own humanity at all. The notion of humanity itself cannot be directly observed. It, therefore, must always be mediated through some kind of a model, indicating (in a metaphorical manner perhaps) the way in which our selves can be understood with the help of the modeling machinery that we – at a certain moment in time – happen to have at our disposal. This influence, from machinery upon humanity, is thus far greater than the influence working the other way round.
There is no steady ground upon which the notion of humanity rests. But then any notion of humanity must rest upon some notion of technology, allowing some specific way of modeling. So that in the end there are just different technologies, the one succeeding the other, the latter always being based upon the teleological vision that the former sets forth, etc. There will thus be no final Turing test, since the notion of humanity that is tested against will depend upon the technology employed. And this is not a pale statement anymore, but constitutes a manifest change in the way in which we look at our research field in general. The central question of AI is not how long it will take before computers have become intelligent, but – instead – how long it will take before our notion of intelligence will be defined in computational terms exclusively.
Publication history
Adapted version of an article that was published in the Proceedings of the 2008 NSVKI Conference (June 2008).
