Artificial Intelligence and Ethics. thinking machines, constructs capable of acting under their own free will. With the advent of

Transcription

1 Artificial Intelligence and Ethics Michael Seaholm PHIL For millennia, mankind has envisioned in myth and fantasy the idea of autonomous thinking machines, constructs capable of acting under their own free will. With the advent of modern computers in the 1940's, the theoretical aspects of a true thinking machine were gradually compounded into the concept that we now call artificial intelligence, or AI. Over the last several decades, numerous strides have been taken in the field of computer science, including the specialized area of artificial intelligence. Along with these advances, there has been considerable ethical evaluation regarding the issues that come up when discussing the idea of machine sentience, which in turn have triggered a number of questions regarding humanity, intelligence, and consciousness. In order to resolve such concerns as AI continues to improve, an understanding of the key issues at stake is necessary. Artificial intelligence refers to the concept of computing machinery being able to replicate or originate what we as humans perceive to be intelligence, or more accurately the computational ability to achieve goals in the world (McCarthy 1). Given that computers are capable of tackling complex problems that are within the capabilities of human intelligence, such as solving mathematical equations, playing chess at an advanced level, and using deductive reasoning, it seems to follow that a computer could be considered intelligent at least to a small degree. It should be noted that artificial intelligence does not necessarily need to mimic the workings of human intelligence, but act in an intelligent fashion in order to solve problems using methods that a computer can perform (McCarthy 1). The primary goal of artificial intelligence research is to create machines that meet this definition of intelligence, as determined by various criteria considered important in the field. The modern field of artificial intelligence was first put forth by British mathematician

2 Seaholm 2 Alan Turing, who discussed the subject at length in his Computing Machinery and Intelligence, published in 1950 (McCarthy 2). At the time, however, the term 'artificial intelligence' had not been invented; this and other details of the field, such as its basic premises and goals, were first outlined in a conference at Dartmouth in 1956 (Cohn 1). Since then, researchers have been successful at building systems capable of tremendous specialized intelligence, such as in chess programs, statistical forecasting, and quantitative analysis. However, researchers have had less dramatic success in general or commonsense intelligence; as an example, while a human can quickly recognize objects in a photograph, a computer presently is unable to do so (Cohn 1). Development of systems that are capable of performing tasks that are elementary for ordinary people would be a leap forward for the field of artificial intelligence and is a matter under intense consideration by current research. Even before the concept of creating an artificial intelligence was considered plausible, literature and other popular media played with the idea, primarily in the form of sentient constructs. Amongst the first of these in modern fiction was Mary Shelley's Frankenstein, a novel published in 1818 that depicted artificial life as a thing that turns against its creator, a thing to be feared (t0ta11ed 1). The idea that intelligent machines or systems would possess a serious moral dissonance with humanity has since become pervasive in the public consciousness. One of the most iconic cultural portrayals of artificial intelligence can be found in 2001: A Space Odyssey's HAL 9000, a malevolent computer that attempts to murder the crew of the spaceship Discovery during a mission (Dirks 1). Conversely, the works of science fiction author Isaac Asimov typically depict robots not as malicious agents, but as helpful tools to be used for the good of humanity, as outlined in his now-famous Three Laws of Robotics (HalJor 3). Popular culture, then, provides a mixed bag of views on artificial intelligence, although more pervasive is

3 Seaholm 3 the idea that artificial intelligence is something to beware of, due to the sheer amount of time this viewpoint has existed. Once machine intelligence came into the forefront of possibility, issues regarding machine versus human intelligence were first put forth. Specifically, the definition of intelligence as it was applied to machines is in question. The issue of machine intelligence was tackled early on in Alan Turing's article Computing Machinery and Intelligence, in which the author tries to answer the question, Can machines think? (Turing 433). In this paper, Turing describes a situation in which an interrogator communicates a series of arbitrary questions by teleprinter to a machine and a human in turn, both of whom he cannot see; if the machine is able to successfully masquerade as human, then it is considered to be intelligent (Turing ). In this way, Turing suggests that a machine that is indistinguishable in conversation from a human should be thought of as intelligent, either because it acts intelligently or because in seeking to act human it does, in fact, possess intelligence. Although the test described in Turing's paper attempts to describe a metric by which machine intelligence can be measured, it is not without its weaknesses. First, the test determines intelligence based on whether or not a computing machine can pass as a human, which is not necessarily the goal of artificial intelligence study; as mentioned previously, the field of artificial intelligence is meant to produce machines that exhibit intelligent behaviors not specifically human intelligence. In the Turing test, it is safe to say that the interrogator would be able to pick out the machine easily if it were to answer something that a human could not, or answer something in a manner that a human would not. By placing an emphasis on behaving like a human, artificial intelligence research limits itself. Such criticisms were illustrated in the aftermath of the 1996 Loebner Prize competition, a contest to see which of a series of programs

4 Seaholm 4 comes closest to passing the Turing test; a contestant named Jason Hutchens won with his program HeX, which he revealed in an Internet publication to be a hoax (Dunn 2). Hutchens had purposely written a program that did not utilize any artificial intelligence yet would fool enough judges to win. He did so in order to make a statement about how the Loebner Prize competition and, by extension, the Turing test were doing nothing for the field of artificial intelligence (Dunn 2). This incident demonstrates how mimicry of humans, although important in some respects, is not the ultimate end in the study of artificial intelligence. Although the aim of artificial intelligence may not be to replicate the way that humans think, considerable research has been done in the area of brain simulation in order to gain a broader understanding of how human intelligence works. The current body of research has been pooled into what is called the Blue Brain Project, which touts itself as the first comprehensive attempt to reverse-engineer the mammalian brain toward the end of producing data on the workings of the human brain (Debono 1). The project took laboratory data garnered from neuroscientific studies and was able to model the neocortex at the molecular level, which, when coupled with a simulated animal, exhibited behaviors of learning and memory (Palmer 1). If a human brain or even an animal brain were to be fully reproduced in simulation, we could not only make neurological breakthroughs, but also come closer to understanding the phenomenon of intelligence. Artificial intelligence stands to benefit from such research. Displays of intelligent behavior have been longstanding in computer systems, due in part to the efforts of artificial intelligence researchers. However, given the inadequacies of the Turing test, alternative questions to Can machines think? have been considered. A more interesting question would be Can machines feel? or, perhaps more accurately, Can machines possess consciousness? If one considers this line of inquiry, one can be quick to say that machines

5 Seaholm 5 cannot be conscious because they lack the mechanisms that you and I possess, and further that these mechanisms if they can be qualified as such cannot be reproduced mechanically. Turing acknowledged this argument in his paper, stating that ultimately according to [the argument from consciousness] the only way to know that a man thinks is to be that particular man. It is in fact the solipsist point of view. It may be the most logical view to hold but it makes communication of ideas difficult... Instead of arguing continually over this point it is usual to have the polite convention that everyone thinks (Turing 446). Put more succinctly, he suggests that, for all intents and purposes, the appearance of consciousness denotes consciousness, making the question of machine sentience more straightforward from his perspective. However, Turing's argument does not seem wholly convincing, since the mere convention of consciousness does not mean that the opposite viewpoint could not be true, especially in the context of man-made constructs. One's view regarding the legitimacy of creating truly conscious machines stems largely from whether or not one would consider the human consciousness to be a wholly material process versus one that is intrinsically linked to the concept of the immaterial, or soul. These two contrasting ideas materialism, in which everything consists of matter and is explained in terms of physical processes, and idealism, in which reality exists within the context of ideas mirror the debate over machine consciousness. From a materialist standpoint, the workings of the human brain, consciousness, and indeed the biological mechanisms of life are all based on physical interactions of matter, meaning that they could theoretically be produced artificially. Conversely, idealism would argue that humans and other life forms possess an innate quality that cannot be replicated mechanically. There exists also a position that rests somewhere between materialism and idealism regarding the concept of life and the soul, allowing for some flexibility in interpretation. Regardless of the stance taken in this regard, artificial intelligence

6 Seaholm 6 research will doubtless continue to advance and produce functional examples of intelligent machines, although the extent of such intelligence depends both upon one's epistemological and metaphysical perspective as well as the actual upper bound of capability for computing machinery. Even within the bounds of strict non-materialistic thought, the advance of modern artificial intelligence systems is difficult to deny. Not only has the efficacy of artificial intelligence systems in specialized applications increased along with the standard increases in computer processing capability, but so too has the development of more general machine intelligence; by processing data obtained from environmental inputs, machines can garner more of what is called embodied intelligence, which allows computers to perform basic cognitive functions rather than their usual mathematics-heavy computation (Henig 1). Progress on both of these fronts makes human-level artificial intelligence at least a vaguely feasible prospect sometime in the future. It is this thought that has led to the development of an idea that someday computing machinery will become able to continually improve upon itself and eventually exceed the capabilities of human intelligence. The resulting paradigm shift in how technology affects the world is referred to as a technological singularity (SIAI, Singularity 1). This singularity, however, requires future computing models to support machines with a capacity for something distinctly associated with animal intelligence: learning. Computer programs that exhibit learning behaviors have already been created. The most simple example of these would be a program that plays a simple game against an opponent knowing only the rules of the game not those moves which would result in a win for the computer and eventually determines how best to play the game (Anderson 2). In a respect that is more related to the technological singularity concept discussed earlier, an intelligent program

7 Seaholm 7 could modify itself to act more intelligently in a recursive and, theoretically, exponential fashion to spur an overall increase in machine intelligence (SIAI, Paradigm 1). Such a program is deemed a seed AI, so named because like a seed, the program grows and becomes more complex until it eventually dwarfs its previous form both in size and in function. Although no generalpurpose seed AI has been created thus far, the concept of a self-improving system has sweeping ramifications for machine learning, intelligence, and life in general. It is difficult to imagine a world under the auspices of advanced machine learning and singularity-level intelligence. Since the very beginning humankind has used technology as a means to survive and prosper, and ultimately as an aid in the pursuit of knowledge and understanding; however, when confronted with self-improving artificial intelligence, such pursuits may be achieved in full, and the current aims of humanity may be fundamentally altered as a result. Mathematician and cryptographer I. J. Good mentioned this concept briefly in the 1960's, stating that with the advent of an ultraintelligent machine [that] could design even better machines... there would then unquestionably be an intelligence explosion, and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make (Farber 1). In the face of such radical change, one must wonder, what would be humanity's calling without the need for discovery? Upon what exploits would we embark, and by what impetus? Would such machines ultimately benefit us or harm us, and by what metric would this be considered a benefit or harm? All of these are considerations when one is trying to envision a world in which technology has reached a pinnacle, and all are difficult to answer because these questions deal with a paradigm that has never before existed in human history. A more immediate and relevant question would be, what threats could an advanced

8 Seaholm 8 artificial intelligence pose to humans when a great disparity of intelligence exists between them? Given the possibilities of artificial intelligence and the hand it could have in our future, this question is important to consider. Although the dominant negative view of artificial intelligence in popular culture is founded primarily on an aversion to the unnaturalness of artificial life and the perceived folly of playing God, there is some credence to the argument that a true artificial intelligence system may not be in humankind's best interests; there exists the possibility that a machine with intelligence several orders of magnitude higher than the average human would view us in the same light that we view wild animals. In order to ensure that intelligent machines do not adversely affect the people for whom they are intended, they must be designed specifically with the interests of humans in mind. Consequently, numerous concepts including friendly AI, which will be discussed shortly were borne out as possible solutions to this potential problem. The Singularity Institute for Artificial Intelligence in San Francisco defines friendly AI as an artificial intelligence whose purpose is one that is aligned with that of humanity; that is to say, it acts in a manner that benefits rather than harms people (SIAI, Friendly 1). It seems almost trivial to suggest that the way to create a friendly artificial intelligence system is to program it that way, but with a system capable of independent learning it could be difficult to come up with rules and regulations that could not be circumvented in some fashion. The concept of friendly AI stipulates that the entity in question, rather than being constrained by safeguards to prevent malicious activity, is designed to want to be friendly as a fundamental part of its being (SIAI, Friendly 2). The Institute suggests that the general concept of intelligence, i.e., intelligence beyond the scope of humans, is not bound by the metrics of phenomena inherent in human intelligence, and as such, intelligent machines would not lean toward either benevolent or

9 Seaholm 9 malevolent action on their own; they would act upon what they had learned, whether good or bad, as would be the case with any autonomous intelligent entity (SIAI, Friendly 2). By ensuring that artificial intelligence systems are designed with friendly AI in mind, the purpose of such systems can be safely realized. Another interesting question with regard to the eventual creation of machines in possession of human-level intelligence is how humanity intends to deal with such entities. Specifically, when dealing with a sentient party that has, thus far, never existed, what role could we as humans justifiably give them? Assuming feelings, emotions, and consciousness are mechanistic, what moral basis would we have in putting such machines in a position of subservience? Of course, technology is created with the intention of using it for our own gain; indeed, the concept of giving consideration to tools themselves is laughable at this stage. However, creating something that can think or feel goes beyond our current view of how technology works. It is unlikely that artificial intelligence researchers will try to simulate such human mental phenomena as feelings, but general intelligence may not be achievable without some degree of consciousness or self-awareness. The result of creating this sort of machine intelligence could spur an alarming social response to the blurred line separating humanity from technology. In a sense, the post-singularity age would be typified by a prevailing sense of confusion. In the presence of an advanced and embodied artificial intelligence that acts human not necessarily able to emulate human behavior perfectly, but enough to hold a rational dialogue with a human being our initial response would likely be sympathetic; when confronted with something that comes close to but doesn't quite reach our view of what an intelligent being should be, we tend to consider it nonthreatening, even possessing a certain charm of the same

10 Seaholm 10 kind that one could attribute to a pet animal. With regard to sentient machines, however, our perception would become torn by the disparity between the logical viewpoint of such entities as artificial constructs and the notion that they are in some sense human based on the way that they interact with us. Should such machines come to move and even look like human beings, the rights of such machines may come into question, reflecting the resulting change in our way of life. Perhaps the most progressive stance would be to respect all entities possessing a minimum level of sentience; however, throughout the course of human history, this has been difficult to enforce even without the possibility of artificial intelligence to complicate things. Regardless of what direction artificial intelligence ultimately takes, progress in the field will continue at a steady pace and substantive answers to the ethical issues that were discussed previously will undoubtedly be explored with more purpose than before. Although the future of artificial intelligence is at this point mere speculation, the conceptual side of things has been fleshed out over the decades since artificial intelligence systems moved into the realm of possibility. However, there is still much work to be done both in concept and in practice; considerable exploration has to be made in the areas of machine learning and embodied intelligence before a qualitative shift can be made. In the immediate future, artificial intelligence will undoubtedly continue to advance in specific niches even as researchers try to develop more general artificial intelligence solutions. Even as we move forward with artificial intelligence research in these respects, however, it is necessary to consider how to responsibly usher in a new intelligence into the world. Perhaps by learning about intelligence in a detached, non-human sense, we can in turn learn more about our own neurological network and the associated concepts of human intelligence and consciousness. With such tools in hand, artificial intelligence promises to fundamentally alter perceptions of technology, life, and the world in general.

11 Seaholm 11 Works Cited Anderson, David. "Artificial Intelligence: Can a Machine Think?" The Mind Project. The Mind Project, Web. 12 May < Cohn, David. "AI Reaches the Golden Years." Wired. Condé Nast Digital, 17 Jul Web. 15 Apr < Debono, Christiane. "Blue Brain Project." Blue Brain Project. Ecole Polytechnique Fédérale de Lausanne, Web. 15 Apr < Dirks, Tim. "2001: A Space Odyssey (1968)." filmsite. AMC, Web. 15 Apr < Dunn, Ashley. "Machine Intelligence, Part I: The Turing Test and Loebner Prize." Rutgers. Rutgers, 29 May Web. 15 Apr < MachineIntelligence1.html>. Farber, Dan. "Can 'friendly' AI save humans from irrelevance or extinction?" ZDNet. CBS Interactive, 24 Aug Web. 12 May < HalJor. "Three Laws of Robotics." Wikipedia. Wikimedia Foundation, 8 Aug Web. 15 Apr < Henig, Robin. "The Real Transformers." New York Times (2007). Web. 12 May < McCarthy, John. "Basic Questions." Formal Reasoning Group. Stanford University, 12 Nov Web. 15 Apr <

12 Seaholm 12 Palmer, Jason. "Simulated brain closer to thought." BBC News. BBC, 22 Apr Web. 15 Apr < SIAI. "Paradigm: Seed AI." The Singularity Institute for Artificial Intelligence. Singularity Institute for Artificial Intelligence, Inc., Web. 12 May < SIAI. "What is Friendly AI?" The Singularity Institute for Artificial Intelligence. Singularity Institute for Artificial Intelligence, Inc., Web. 12 May < SIAI. "What is the Singularity?" The Singularity Institute for Artificial Intelligence. Singularity Institute for Artificial Intelligence, Inc., Web. 12 May < t0ta11ed. "Pop Culture and A.I." Marzopolis. Invasion Plans, 23 Jul Web. 15 Apr < Turing, Alan. "Computing Machinery and Intelligence." Mind (1950): Web. 15 Apr <