Tuesday, 31 July 2012

Is Artificial Intelligence a Fallacy?

The quest for a working model of artificial intelligence began in earnest with Alan Turing's cracking of the Enigma Code towards the end of WW2. The enlisting of a mechanical apparatus – the Enigma Machine – facilitating complex and numerous logical operations in a fraction of the time it would take a human, or team of humans, to perform the same operations. But Turing's experimental concept of "thinking machines"1 went further than this, in theory. Turing's idea was that all human intellectual operations could ultimately be broken down into series of logical procedural steps, and the implication of this idea (in the 1940s) was that what limited machines from approximating closely to human intellectual activity was not a limitation in principle, but only one of the means available. By forward projecting advancements in technological materials and apparatus, such limitations might be overcome, at least in theory.

Turing opens his 1950 paper Computing Machinery and Intelligence with the question: "Can machines think?"2 The question implicitly begs a working definition of the concepts machine and thinking. With regard to the latter definition Turing anticipates a degree of resistance from common consent that so customarily human an activity as thinking might be conceived as an attribute of inanimate objects, and, rather than engaging in absurdity, abandons any further effort over the difficulty of the definition. Rather than risking a possibly fruitless meditation on the nature of thinking, he devises instead an experimental scenario, since known as the Turing Test, in which he proposes that an imaginary computer (one which in Turing's era was, and still remains, technologically unachievable) plays an "imitation game" against a human competitor, responding to a series of questions from a remote interrogator. If the computer succeeds in responding 'naturalistically' enough to the interrogator's questions that the interrogator is unable to reliably distinguish between human and machine, then, in Turing's terms, it has come to deserve the attribute of a "thinking machine":

"May not machines carry out something which ought to be described as thinking but which is very different from what a man does? This objection is a very strong one, but at least we can say that if, nevertheless, a machine can be constructed to play the imitation game satisfactorily, we need not be troubled by this objection."3

Turing's "imitation game" (which might just as well be characterised as a 'deception game') was intended to explore the capacity of machines to think, but was designed from a premise which eluded the very question of the nature of thinking. What is more, it perceived no practical limit on the design and operational capacity of those machines, which tends to place the imagined scenario within the category of Science Fiction – that is, as a pure thought-experiment. Framed in such terms it becomes a self-fulfilling prophecy. If instead we allowed ourselves to be troubled by a precautionary objection over the differences between the imagined machine process, which manages to deceive of the appearance of thought, and actual human thinking, we might elucidate, rather than elude, the enquiry over the question of what it is to think. Turing's motivation for not being troubled by this objection seems to be an anthropomorphic one – which is to say it equates the imaginary appearance of thinking in machine behaviour with its human analogue unproblematically, with as little concern for reality as the way in which Thomas the Tank Engine, for instance, is depicted with the imaginary features of a human facial expression.

Thomas the Tank Engine
As a matter of common consent, the term thinking implies what we as individuals experience as embodied consciousness. This includes a variety of faculties, including: reasoning; recollection; imaging; desiring; calculation; reflection; association, etc. There is an experience, common to us all, of what it is to be a 'thinking being', but which also appertains to all our experience, including what we witness as evidence of thought in others. Therefore, we are unable to achieve a clinical distance independent of our own processes of thought to be able to consider 'thinking' (or its corollary 'the mind') as a discrete phenomenon in isolation. The question requires a self-reflexive approach, i.e., one which refers its conclusions back upon the limits of its observations. This is possibly the reason for Turing's unwillingness to approach the difficult question of the meaning of what it is to think, as there is no 'outside' of the mind from which one could approach a definition with any prospect of empirical certainty.

The mind receives sense-data as 'inputs' in the form of physical, bodily processes; the expressions or 'outputs' of thought-processes in speech and gesture are also physical motor processes. The processes of thought, as well as the appearance of thinking, are therefore necessarily structured somatically. In a human context this is what gives to thinking its authenticity. This is perhaps what Spinoza meant when he wrote that: "the object of the idea constituting the human mind is the body"4. We do not (normally) have direct access to another's intimate thought processes, but we receive the other's motor outputs as our own sensory inputs, usually involving a degree of linguistic codification. However, we can anticipate reciprocally the cognitive processes which have given rise to these expressions, by reverse analysis and identification with our own familiar processes of cognition. This depends upon our awareness of a fairly restrictive array of parameters defining cognition, but which ultimately make it possible for individuals to cooperate and interact as a species.

Independently of this shared capacity there would be no grounds for establishing the norms, morals, codes of behaviour and understanding, necessary to the formation of any form of social contract. In order for such an ethics to be meaningful, and binding, thinking as the basis for action must take the form of embodied consciousness as a limiting principle, i.e., as one that is universally comprehended through the shared physical (somatic) structure which gives rise to our ability to think. The transfer, anthropomorphically and in science-fictional terms, of an analogue of human thinking to inanimate objects, suggests no conceivable executive role for such an 'intelligence' in which we could ever have any confidence in its ability to make an ethical decision. The question posed by Turing in his opening gambit: "Can machines think?" is therefore one for which there is no ethical legitimacy.

In the decades following Turing's influential publication, there was a significant impetus towards the development of a model for 'artificial intelligence', and its academic discipline, cognitive science, encouraged by its association with technological advancements in solid-state electronics and semi-conductors. The academic origins of this and of Turing's own theoretical and technological programme arose out of the largely Anglo-American empiricist philosophical tradition of logical-positivism, which had emerged prominently in the years following the end of WW1. Logical-positivism, or in Bertrand Russell's definition, logical atomism5, was fundamental to the development of information science, and was an attempt to analyse language into its simplest logical and factually 'positive' components – a point at which analysis could go no further. Its premise was that everything of meaning in language should be reducible to what is directly given in sensory experience. Natural language is typically ambiguous because it treats as 'things' concepts not derivable directly from positive experience (i.e., from direct sense-impressions): concepts like 'honour', 'friendship', 'life' etc.; in other words, concepts which represent qualities, relationships, and processes, and which are apprehended by intuition rather than as sensible objects. In informational terms, the substantive use of such concepts in natural language is 'metaphysical' and hence meaningless. Logical-positivism sought a method of making language speak without recourse to what it understood as metaphysical 'pseudo-statements'. Such statements were incongruous to empirical reality – in informational terms, effectively 'noise' – they could not be rendered as informational certainties. The correspondence of language with reality was to be found at the level of its simplest irreducible elements – the 'bare-bones' of subjects and predicates – in Russell's terminology, its atoms. In this fundamentalism of logic applied to language, logical-positivism prefigured a method for rendering natural language into machine-readable informational units.

The supposition by proponents of Artificial Intelligence was that the rubric of human intellectual operations could be functionally equated with this process of the empirical extraction of logical elements, together with the efficient processing, or 'computation', of the derived informational content. This construed an ideal form for 'intelligence' in which rational, logical procedures be conducted in their pure state, as isolable informational routines (algorithms), unimpeded by logical inconsistency. In this view, human intelligence actually sets the goal as it is perceived to be, at least potentially, superlative. However, in its everyday operations, human intelligence is usually flawed or hampered in some way – it is typically prone to distraction, and hence to error, or at least impedance (for instance, in the need for error-checking revision); it may also be disrupted by emotive impulse or fatigue. It is tempting therefore to project that a refined from of machine-intelligence might eradicate human unreliability in information processing, by the reduction of intelligence to its purely logical, machine-readable, constituents.

Within a philosophical framework which generally elevates human rationality above other organically occurring systems, forms of animal intelligence are generally disavowed – animal behaviour usually being construed in terms of the effects of instinctual impulses rather than rational thought processes properly speaking. Instinctual animal impulses could be perceived as the obverse of rational thought processes. To some extent conversely then, perhaps human animals might be understood as being affected by residual instinctual drives, which interfere with the straightforward operation of logical processes, so that the rational human animal becomes tainted, in practical terms, in comparison with its ideal machine analogue. As a counterpoint to the characteristic irrationalities and foibles of the human cognitive condition, imagine a race of Vulcans, including Mr. Spock, whose sole existential purpose seems to be to remind us mere humans of just how supremely rational we might be if we weren't so fatally quirky. It is a somewhat tormented perspective which projects a mode of intelligence from within us, so as to establish a measure for intelligence against which we must characteristically always come up short. I believe it is partly from this kind of tormented perspective that the impulse to fashion a form of 'artificial intelligence' achieved widespread popular and commercial support.

Partly, that is, because there were also clear instrumental reasons in support of such a project; not least because, even in their embryonic forms, these ideas had contributed in no small way to the allied victory over the Nazis, in 1945. In the high-stakes game of international conflict, the requirement for the smooth, fast, and efficient performance of information processing tasks would henceforth prioritise the quest for machine intelligence.

Key factors in these developments were therefore the influence for the pursuit of military intelligence, the development of sophisticated weapons technology, including nuclear weapons, and the associated drive towards the exploration of space; all situated against the backdrop of cold-war stratagems which so dominated the geopolitical landscape of the 1960s and '70s.

At no point in this process does there seem to have been much, if any, reflective analysis of what constitutes, or what might differentiate, what we commonly experience as 'human intelligence' beyond the logical-positivist reduction of cognitive processes to their functional logical components. What of the differences between perception and cognition, or between reason and intuition? How do we explain the faculties of association, or of judgement, or the intuitive categories of the understanding, in terms of a mere logical sequence of processing steps?

Take the following scenario as an example: if the sun emerges from behind a cloud, I feel the warmth of the sun on my face. The logical components of this scenario are: the sun shines; I experience warmth on my skin; but the relation of causality between these two events entailed by my cognition: the sun warms my face, is not inferable purely on the basis of their logical content – it requires an intuitive association to be provided by my understanding. A system which operates independently of intuitive understanding must in addition ask a series of questions in order to eliminate any of the possible alternative explanations for why my skin got warm (I experienced sudden embarrassment; someone fired up a barbeque in my vicinity, etc.), in order to ascertain with only relative certainty that it was due to the sun's rays. A system operating on the basis of machine rationality is intuitively inert, and therefore its certainty over questions of causality such as the one described can only be relative to the degree to which it is efficiently pre-programmed to assess all of the possible conditions pertaining to any possible scenario in which these two effects may be conjoined, and also to take account of the fact that the sun's emergence will not always, by necessity, have the effect of warming my skin – it depends on other factors contingent upon the unique situation. Such a pre-programming procedure might take a potentially infinite length of time, and still never achieve the degree of absolute certainty entailed by my response: the sun warms my face, which was available to me in an instant.

Or consider the difference between intuitive judgements of universality and particularity. For instance, imagine a situation where someone expresses hope (in the universal), without expressing optimism (in the particular). A man might say "I hope to see my children", and by this statement may be indicating that he expects (with a particular intention) to see his children at some specific point in time, which may be conditional on some other impending circumstance; or he may equally be implying that he hopes (in general) to see them perhaps against all possible odds (i.e., without any particular expectation). Someone having knowledge of the man's circumstance, or who was engaged in a conversation with him would likely know intuitively which particular inference was appropriate. We routinely exercise these kinds of intuitive judgements in almost every aspect of our lives, without necessarily being consciously aware that we are doing so. If we artificially extract the logical content of the statement "I hope to see my children" however, there is no effective means to distinguish which of the two alternatives is inferable. Such abstractions are typical of the way in which information systems treat data objects, and although one could argue it is merely the lack of background information which prevents the distinction, the utterance has no logical interpretation unless it is linked to a particular intention. The tendency in fact is for logical interpretations to treat all such statements as cases in the particular, at the expense of the universal.

It seems to me, therefore, that synthetic judgements of causality entailing absolute certainty of the kind in the first example, and intuitive inferences of universality or particularity in the form of the second example, while indispensible elements in most kinds of human intellectual activity, are beyond the scope of computational devices operating on the basis of machine rationality. Against the attribution of 'artificial intelligence', to accurately define the method of operation of such systems one would need to resort to a more prosaic definition, such as 'parallel logical computation', as in nearly seventy years since the birth of the concept 'thinking machines', no such system has achieved anything which vaguely approximates to the processes of human understanding with which we are all familiar.

We know no higher form of intelligence than human intelligence. Therefore, artificial intelligence, as a theoretical construct, must be an attempt to model or approximate the faculties of the human intellect. It is a sad consequence that, in the eyes of many, it also became imbued with the potential of surpassing human intellect, while at the same time failing to appreciate the fluidity and sophistication that distinguishes the latter. If we propose that a machine is capable of outsmarting people, yet which in practice fails even to approximate the most common practices of human understanding, we risk bracketing out human critical faculties from the sentient business of our daily routines. Artificial Intelligence begins from a premise which underestimates human critical faculties, yet within the terms of its premises cannot even match them. It remains therefore a fallacy, rather than a reality, and one which, for so long as we remain ensnared by its false promises and charms, will inevitably 'dumb us all down'.

  1. Turing, Alan, Computing Machinery and Intelligence (October 1950), Mind LIX (236), p.436: http://somr.info/lib/Mind-1950-TURING-433-60.pdf [back]
  2. Ibid., p.433. [back]
  3. Ibid., p.435. [back]
  4. Spinoza, B., Concerning the Nature and Origin of the Mind, in his Ethics (tr. Boyle, A.; ed. Parkinson, G.H.R.), J.M. Dent & Sons, 1989, p.48. [back]
  5. Russell, Bertrand, The Philosophy of Logical Atomism (1918), Routledge, 2010. [back]

0 comments :

Post a Comment