Machine Intelligence and Explication

TL;DR: This report is an MA ("doctoraal") thesis submitted to the department of philosophy, university of Amsterdam, and attempts to answer the question whether machines can think by conceptual analysis, which follows a different route than the ideal argument.

Abstract: This report is an MA ("doctoraal") thesis submitted to the department of philosophy, university of Amsterdam. It attempts to answer the question whether machines can think by conceptual analysis. Ideally. a conceptual analysis should give plausible explications of the concepts of "machine" and "intelli­gence" and then investigate the intersection of the sets of entities defined b) these explications. If the intersection is empty and the a priori argument is correct (or plausible), then empirical research into machine intelligence will {plausibly) not result in an intelligent machine. On the other hand, if conceptual analysis cannot show the intersection to be empty. it remains an empirical (or rather, technical) question whether such machines can actually be constructed. Such a neat argument cannot be produced, however, due to the vagueness of the concept of intelli­gence. It is quite possible to provide a rather uncontroversial explication of the concept of machine. Exist­ing controversy about the possibility of machine intelligence is about the nature of intelligence, not about the nature of machines. Indeed, if intelligence could be unambiguously defined, we could (in principle) build a machine to implement it. Those who believe that intelligence cannot be realized in a machine, can­ not base their arguments on an explicit and uncontroversial analysis of the concept of intelligence. The argument in this essay therefore follows a different route than the ideal argument. After a defin­ition of machine which combines the important characteristics of that concept in computer science and sys­tems theory, try to explicate why we think this definition captures our informal intuitions about the nature of machine-like, mechanical processes adequately. This leads to an explication of what explicitly described processes are. Chapter 2 then replaces the question whether machines can think by the simpler question whether machines can explicate. Using the explication of the concept of explicit descriptions given in chapter 1, I argue that the process of explication cannot be explicitly described. If that argument is correct (or plausi­ble), then no machine can (plausibly) be built which explicates a situation , for to build a machine is to implement an explicit description. The bearing on the original question of machine intelligence is this: If human intelligence presupposes the ability to explicate, an entity which cannot explicate cannot have human intelligence. Chapter 2 contains some arguments why we do not attribute human intelligence to a being which cannot explicate. In chapter 3 the argument is defended against some possible counterarguments and compared with two well-known criticisms of artificial intelligence, those by Dreyfus and Searle. Finally, chapter 4 explores some practical as well as metaphysical consequences of the thesis. This short overview of the argument should already have made clear that I do not believe that an uncontr­oversial, explicit proof of the impossibility of machine intelligence exists. If such a proof existed, it could be automated, which would be close to a refutation of what the proof would establish. It follows that holes can be shot in the argument. It won't execute without errors in all environments. Therefore, in the interest of (among other things) brevity, I stopped explicating when further explication would backfire and merely expose the emptiness of the argument. That -the empty argument- would have been, as Isshuu Miura said, closer to the truth than the essay I wrote now. But then, I wouldn't have passed the exam by handing in an empty paper. Working on this thesis made me painfully aware that the semantic network we live in is essentially fluid and unbounded. Thanks are due to my supervisor Hans Swart, who followed me on my wanderings in various interesting directions and who suggested I stay with one topic and work that out. Thanks are also due to Dick de Jongh and Loet Leydesdorff, who gave constructive criticisms of the thesis.