This website is largely concerned with efforts to build intelligence into machines. Such efforts would seem to presuppose that we have a clear concept of natural intelligence, of the mental and behavioural abilities that can be found in human beings and perhaps also in other animals. In fact, our view of natural intelligence is constantly shifting, medicated not only by new biological knowledge but also by the emerging competence of artificial systems. It is useful to glance at attitudes to intelligence that mostly predate computer progress over the last three or four decades. In this way we can see that the concept of intelligence has many dimensions, not all of which may be relevant to the design and manufacture of clever machines. (And we may also expect a reciprocal effect: progress in designing artificial intelligence will increasingly throw light on what it means to belong to an intelligent biological species.)
In everyday usage 'intelligence' and its derivatives are familiar words: it is rarely thought ambiguous to declare that this or that person is intelligent, or that intelligence is a desirable human characteristic. But when we try to define 'intelligence' we may find that the concept becomes elusive. It is concerned with the manipulation of numbers, words or other symbols? It is related to practical action in the world? Does it bear on creativity and invention? Do we find intelligence in mental activity or in effective behaviour? And how is intelligence related to such phenomena as willing, learning, remembering and emotion? We are finding it possible to build a degree of intelligence into artefacts without understanding fully the nature of natural intelligence. But awareness of views about intelligence in human beings can indicate what may or may not be possible for machines.
WHAT IS INTELLIGENCE?
Definitions - of anything - aspire to completeness. The definition that leaves out some important feature of an entity is rightly regarded as inadequate. And where the topic is complex or nebulous, as intelligence can be, definitions are particularly at risk. Can we define intelligence? Should we even try? Perhaps it is better simply to list appropriate mental and behavioural activities to convey an impression of intelligence.
There are in fact various types of definitions that should not be confused. Miles (1957), for example, identified:
~: real definitions, concerned with the actual entities to which the terms refer;
~: nominal definitions, concerned with the meanings of words rather than the entities that the words are supposed to denote. Here lexical definitions focus on how words are commonly used; stipulative definitions focus on how a person intend to use a term;
~: operational definitions, concerned with meanings in terms of observable, measurable operations.
To these types of definition we may add definition by enumeration, where the nature of an entity or word is described by listing characteristic features. This approach can reveal that there are many different types of the entity in question, and help to avoid the confusion that arises when different people use a word correctly but with different meanings. Confusion of this sort can beset discussion of intelligence.
Types of Intelligence
We commonly imagine intelligence to consist in such things as the ability to solve problems, do sums, learn, cope with new situations, etc. It is assumed that people in responsible positions - managers, doctors, cabinet ministers, etc - are intelligent, an assumption that sometimes leads us into paradox: 'if so and so is so intelligent why does he/she act/talk so foolishly?' This encourages us to identify different kinds of intelligence. The brilliant mathematician may have no political insight, and a first-class degree in Greats says nothing about the individual's capacity to act intelligently in human relationships.
To some extent the different types of intelligence are recognised in language. Thus we may talk of judgement, wisdom, insight, perception and learning. Intelligence may be 'manipulative', 'verbal', 'philosophical', etc. We will see that such distinctions are important to an understanding of artificial intelligence: computers may be seen as highly skilled at mathematics, much less competent at tasks requiring what we would call judgement or wisdom. At the same time we should not be trapped in rigid categories. Biological intelligence has evolved over millions of years, and the infant enlarges its spectrum of intelligence as it grows to maturity: in analogous fashion there is a discernible evolution of computer intelligence.
Sometimes a distinction is made between intelligence and specific intellectual abilities, where intelligence is seen as the capacity to acquire particular skills rather than the skills themselves. Thus someone who takes a month to master simple differential equations may be deemed less intelligent than someone who needs only a few minutes. A skill may be acquired through laborious effort over a lengthy period, and its presence may indicate at least a certain minimal intelligence (some skills cannot be taught to some people).
It has also been suggested that the capacity to acquire intellectual skills is a general capacity, equally relevant to skills of different types. One of the criticisms of artificial intelligence is that individual AI programs tend to focus on specific isolated tasks (game-playing, theorem-proving, story-writing, etc) and so lack the generality that characterises real intelligence. (We will see that this is not a particularly telling criticism: a computer can easily be given access to different AI programs to extend its range of competence, and moreover there are already AI programs that embody the requisite general capacity, eg the General Problem Solver of Ernst and Newell.
Some observers have been tempted to reify intelligence, much in the way that certain mental attributes can be wrongly represented as soul or spirit. In fact intelligence should be regarded as an abstraction from certain kinds of behaviour. It is only through behaviour that we can recognise intelligence, even though the behaviour may then allow us to deduce mental (or even, some would say, metaphysical) capacities. This approach is important to an evaluation of intelligence in computer systems: a computer-based system will be deemed intelligent by virtue of what it can do, ie how it behaves with regard to problem-solving, decision-making, inference drawing, etc.
Pyle (1979) has emphasised that 'intelligence' is a 'situation-specific' word. It can be seen to take on different meanings according to the particular situation. We have noted that it can signify the capacity to acquire a skill, but it can also indicate particular abilities in particular circumstances, eg driving a car, writing a poem, solving a crossword puzzle (Dockrell, 1970). But a capacity to acquire a skill implies a potential for behaving in certain ways. In saying that a person is intelligent we are, at least in part, imagining how the person would behave in circumstances that may not yet have occurred. And we can observe that this idea connects well with what some people see as the general capacity of intelligence: a full description of current performance does not exhaust the possibilities of the intelligent system.
The various definitions and descriptions of intelligence highlight, to some extent, the interests of individual researchers. Pyle lists a few important researchers and their definitions, and notes that most say something about the ability to reason:
Binet: to judge well, to comprehend well, to reason well.
Spearman: general intelligence ... involves mainly the 'education of relations and correlates'.
Terman: the capacity to form concepts and to grasp their significance.
Vernon: 'all-round thinking capacity' of 'mental efficiency'.
Burt: innate, general, cognitive ability.
Heim: intelligent activity consists in grasping the essentials in a situation and responding appropriately to them.
Wechsler: the aggregate or global capacity of the individual to act purposefully, to think rationally and to deal effectively with the environment.
Piaget: adaptation to the physical and social environment.
The mix of abilities (judgement, comprehension, reasoning, concept-formation, appropriate response, adaptation, etc) in these definitions highlights the multifaceted nature of intelligence. No single definition exhausts the possibilities. A system - biological or artificial - with only one skill or capacity has very limited intelligence, if the system is intelligent at all.
It is inevitable that workers in artificial intelligence would need to scrutinise natural intelligence to identify key characteristics, definitive attributes, etc. You cannot begin to frame intelligence in artefacts until you have some concepts about how intelligence is to be recognised in biological systems. Douglas Hofstadter, for example, who carries out AI work at Indiana University, has indicated what he regards as 'essential abilities for intelligence'. These include the capacity to respond to situations flexibly, to exploit fortuitous circumstances, to make sense out of ambiguous or contradictory messages, to find similarities in situations separated by differences, and to generate new concepts and novel ideas. Again the multifaceted character of intelligence is evident.
The Hofstadter 'essential abilities' emphasise such aspects as judgement, insight and creativity, elements which are not normally associated with computers. Indeed the conventional wisdom would have it that computers are inflexible and unimaginative, preoccupied with blind obedience to rules, as incapable of independent initiative as a train on a track. But researchers in artificial intelligence have a different view. Hence Hofstadter (1979) remarks: '... the strange flavour of AI work is that people try to put together long sets of rules in strict formalisms which tell inflexible machines how to be flexible.'
The multidimensional nature of intelligence suggests that some elements of intelligence will more easily be structured into artificial systems than others. It is, for example, much easier to quantify and measure such elements as judgement or creativity. And unless an element can be quantified, in some sense, it is difficult to see how it could be encapsulated in a computer program. Developments in computer science have encouraged a cognitive approach to human psychology, including human intelligence. This psychological approach has in turn stimulated ideas on how human mental processes can be modelled in computer systems.
We have seen that there are many types of intelligence, and that different researchers have individually focused on the areas of interest to them. The emergence of artificial intelligence as a branch of computer science has encouraged a view of intelligence as an information-processing phenomenon. This circumstance has reinforced cognitive interpretations of mental activity and indicated how such activity can be modelled in computer systems.
Artificial intelligence today focuses on particular types of activity: learning, reasoning, language understanding, translation, perception, etc. There is interesting speculation on artificial emotion but little if any successful work. There are many types of human intelligence that machines are not even beginning to emulate, but in other areas computer-based systems are undeniably intelligent. It is worth looking at other aspects of the AI background before exploring some of these areas.