3.1.1 What is intelligence?
This is a course about artificial intelligence. The aim of artificial intelligence is simply this: to build intelligent machines. This goal seems ambitious enough and is certainly easy to state. But before we can even start on such a project, we must have a fairly clear idea of what it really is we are trying to build. We all think we know what a machine is and we all probably feel we can recognise intelligence when we meet it. But maybe this confidence is misplaced? There are two major questions we have to try and settle before we embark. They are:
- What is intelligence anyway?
- If we did manage to build an intelligent machine, how could we tell if it was really intelligent?
Alone, these two seemingly simple questions have spawned a vast literature. However, I only want to deal with them quite briefly, for reasons that I hope will become clear soon, and as a means of making three important observations. Let's start with an exercise that one can find in every course in artificial intelligence.
What do you think intelligence is? Jot down a few notes about this.
The number of possible answers you might have come up with is bewildering. It's possible you offered alternative names for the concept, such as 'ingenuity', 'nous', 'cleverness', and so on. But it's likely that you also backed that up with a fuller description or definition, perhaps something along the lines of 'the capacity to think and reason', 'the ability to apply knowledge' or some such. Most likely of all, though, is that you suggested examples of intelligence, such as logical reasoning, use of language, abstract thought, and so on.
Although we may all think we recognise intelligence when we see it, it seems to be a difficult notion to pin down. In the discussion above, I suggested three overlapping approaches one might take to defining the concept: names, definitions and descriptions. But there are objections to the kinds of answers all these three lead to:
- Names. Suggesting names or synonyms for intelligence gets us no further, really. Saying that 'intelligence' is the same as 'cleverness' hardly tells us anything of interest.
- Definitions. Obviously this is a better idea, but still runs into trouble. For a start, there's little obvious agreement on a definition. A Google search I carried out yielded – after discounting special meanings, such as 'spying' – ten or more competing definitions, among them:
- the ability to comprehend; to understand and profit from experience
- a general mental capability that involves the ability to reason, plan, solve problems, think abstractly, comprehend ideas and language, and learn
- the ability of an individual to understand and cope with the environment
- the capacity to create constructively for the purpose of evolutionary gain.
Several things strike me about these. Leaving aside the obvious disagreements (in some cases they hardly seem to be talking about the same thing at all), a second point is that they all seem very abstract: intelligence is defined in terms of other concepts – 'comprehension', 'understanding', 'reasoning', 'creativity' – which are equally vague. A third related point – slightly less obvious, perhaps – is circularity. Defining 'intelligence' in such terms as 'comprehension', or 'thinking abstractly', words which we inevitably associate with intelligence anyway, is to some extent saying little more than 'intelligence is behaving intelligently'.
- Examples. Most definitions of abstract concepts rely on examples. To define 'intelligence', it is only natural to fall back on instances of what we take to be intelligent behaviour,reasoning, problem solving, use of language, and so on. This is clearly helpful, but has its own problems. Maybe to single out two or three examples of intelligent behaviour as defining properties of 'intelligence' is to be in danger of ignoring others. Suppose we take 'abstract reasoning' as a core property of intelligence: what about the quick-thinking solver of practical problems? Alternatively, suppose we take 'use of language' as key: does this mean the tongue-tied mathematical genius is more stupid than the silver-tongued political rabble-rouser? This may seem like hair-splitting. However, I'll argue later that to define intelligence in terms of just a few key abilities may be to make a serious mistake.
So what about artificial intelligence, the quest to build intelligent machines? You'll find most books on the subject start with a brief attempt to define what it is that practitioners are trying to do. To produce a long list of these would be wearying, so here are just three examples:
[The automation of] activities we associate with human thinking, activities such as decision-making, problem solving, learning ...
The study of computations that make it possible to perceive, reason and act.
The design and study of computer programs that behave intelligently. These programs are constructed to perform as would a human or animal whose behaviour we consider intelligent.
All the problems I outlined above are here: disagreement, abstractness, circularity, a reliance on a few key examples. But my aim is not to belittle these authors – I certainly could have done no better myself. I just want to make three observations about the whole endeavour, which I think relate to the foundations of the whole project of artificial intelligence.
Observation 1: There is an obvious lack of agreement on what intelligence is, and thus of the exact goals of artificial intelligence.
Observation 2: The only really clear and effective definitions of intelligence are in terms of a few examples of intelligent behaviour: perception, reasoning and action, in the case of Winston above; decision making, problem solving and learning in Bellman's definition.
Observation 3: The overwhelming focus is on human intelligence. You may recall that Descartes considered animals to be simply mindless automata. The quotation from Hobbes' Elements of Philosophy above suggests that Hobbes too thought our intelligence made us utterly distinct from the animal world. Bellman and Winston above seem to concentrate on human qualities such as reasoning and problem-solving and most other authors follow them. Only Dean et al. gave explicit consideration to the view that non-human animals are capable of intelligence too. The intelligence of animals, which I am calling natural intelligence, is a major theme of this course.
Actually, one of the most honest comments I've read on the actual practice of artificial intelligence comes from Russell and Norvig:
We have now explained why AI is exciting, but we have not said what it is. We could just say, 'Well it has to do with smart programs, so let's get on and write some.'
As a computer scientist myself, I sympathise with this. But suppose I do write a system of some kind that I'm claiming is intelligent. How could I tell if it was intelligent?
Suggest a few ways it might be possible to tell if a computer system is intelligent. You might find it helpful to consider why you consider your friends and colleagues are intelligent (if indeed you do).
Of course, as before we're hampered by uncertainty about the meaning of the term 'intelligence'. It's instructive, though, to ask why it is we believe other humans, our friends and acquaintances, possess that basic human quality. Well, I can't know for certain that my best friend is an intelligent, reasoning human being, rather than a brilliantly constructed but mindless automaton. But I can assume it from her actions. She holds conversations, responds appropriately and flexibly to the world about her, solves problems, plans ahead, turns up for appointments at the right time, pursues her own goals, etc.
An answer like this looks quite convincing, but may run into some of the same problems we ran into in trying to define intelligence in the first place. I suggested a whole list of actions by which I might judge my friend to be intelligent. But did I leave any out? Are some of the actions I did mention more important than others? Am I promoting some at the expense of others?
This is not a modern problem. Descartes faced up to it in the 17th century, as he pondered the differences between humans and animals. And his answer has had such an immense influence on the founders (and later practitioners) of Symbolic AI, that I think it is worth looking at closely. He wrote:
... if there were machines which had ... the external shape of a monkey or of some other animal without reason, we would have no way of recognizing that they were not exactly the same nature as the animals; whereas, if there was a machine shaped like our bodies which imitated our actions ... we would always have two very certain ways of recognizing that they were not ... true human beings.
The first of these is that they would never be able to use words ... as we do to declare our thoughts to others: for one can easily imagine a machine made in such a way that it expresses words, ... but one cannot imagine a machine that arranges words in various ways to reply to the sense of everything said in its presence, as the most stupid human beings are capable of doing.
The second test is that, although these machines might do several things as well or perhaps better than we do, they are inevitably lacking in some other, through which we discover that they act, not by knowledge, but only by the arrangement of their organs ... As a result of that, it is morally impossible that there is in a machine's organs sufficient variety to act in all the events of our lives in the same way that our reason empowers us to act.
Now, by these two same means, one can also recognize the difference between human beings and animals. For it is really remarkable that there are no men so dull and stupid, including even idiots, who are not capable of putting together different words and of creating out of them a conversation through which they make their thoughts known ...
Despite the elegant 17th-century language, this passage has an extraordinarily modern ring. Descartes saw very clearly some of the problems and challenges of modern artificial intelligence. So, it is worth being very clear about the points he is making.
Sum up the two tests Descartes proposes for detecting true intelligence.
The first test is the ability to use language. Humans alone, who are for Descartes the only creatures capable of intelligence, can put together words so flexibly as to be able to respond to the infinite variety of situations that confront them. A trained animal, or an automaton, he believes, is bound sooner or later to be caught out.
Secondly, humans are versatile. We are always capable of acting flexibly and creatively in novel situations.
In the last paragraph, he suggests, once again, that it is the use of language that can be used as a test for the existence of reason.
As I've argued, Descartes' concentration on language performance had a huge influence on the founding fathers of artificial intelligence. One can see strong echoes of it in the Dartmouth proposal paper you dealt with in Exercise 5, with its stress on language, creativity and problem solving as key features of intelligence. But without doubt the dominant influence on modern thinking about recognising intelligence, human and artificial, and the direct heir of Descartes, was Alan Turing.
In his seminal 1950 paper 'Computing machinery and intelligence', Turing addressed the same question Descartes had faced three hundred years earlier. What are the defining features of intelligence and how can we recognise them? But for Turing, the matter had real urgency, because he believed that in the digital computer we had at last found a machine that could be made intelligent. We will return to the issue of what was Turing's exact idea of the digital computer later in this course. For the moment, let's consider how he tackled the question of how intelligence could be recognised.
Read through the first five sections of Turing’s paper, 'Computing machinery and intelligence', which can be found using a search engine, or directly on one of the following websites:
- https://www.cs.ox.ac.uk/ activities/ ieg/ e-library/ sources/ t_article.pdf
- https://jupyter.brynmawr.edu/ services/ public/ dblank/ CS371%20Cognitive%20Science/ 2016-Fall/ TuringComputing.pdf
What test does Turing propose for the detection of intelligence in a machine?
Turing proposed an investigation that he called the 'Imitation Game', but which is now famously called the Turing Test. In Turing's game, there are two channels of communication, A and B, through to a neutral human observer, C: A comes from a computer and B from an average human being, but C has no knowledge of which is which, as both are hidden and communicate through a standard teletype. The job of the computer at A is, in a dialogue with C, to convince her that it is a human being. C is free to ask any questions, or make any remarks, she likes in the dialogue, and to go on for as long as she wants; but if in the end she is unable to tell which is the human and which the computer, then the computer has passed the Turing Test and can be said to be intelligent.
You should be able to see clearly the influence of Descartes here. For Turing, as for Descartes, the key indicator of intelligence is flexibility of response through language. Turing's proposal has been immensely influential and, although many modern researchers believe it is deeply flawed, as a definitive test for intelligence in machines it has never been seriously challenged. In 1990 Hugh Loebner, in collaboration with the Cambridge Center for Behavioral Studies, set up a yearly competition for The Loebner Prize. He provided the capital for a gold medal and an award of US$ 100,000 to the programmers of the first computer to pass the Turing Test by giving responses indistinguishable from a human's. The prize has not yet been won (2006). However, each year a prize of $2000 and a bronze medal is presented to the designers of the most human computer program, as compared to other entries that year.
Read through the document linked below. These are (lightly edited) transcripts of some of the dialogues between the human judge and the winning computer in 2005, a program called Jabberwacky. I've also included one transcript of a conversation with the human confederate B. Can you tell which of the transcripts is the one of the dialogue with the human confederate? How well does Jabberwacky perform in the Turing Test, in your opinion? What do you think is its chief failing?
It was fairly obvious to me that the only dialogue involving a human was Transcript 2. It's possible that you didn't spot it, but this doesn't necessarily mean any shortcoming on your part. However, I think it does point to a weakness of the Turing Test itself: it is really quite subjective – what seems natural and human to you may seem artificial and machine-like to me. More importantly, one can argue that we (all of us) subconsciously want to be fooled: our tendency is always to read order, pattern and rationality into the situations we encounter, even if they are not present, in just the same way as we see faces and images in the random patterns of the clouds.
It's hard to make a clear judgement of Jabberwacky's performance. Sometimes the replies seem quite normal and human. At other times they seem wildly off the mark, almost random. A general tendency of all systems designed to pass the Turing Test is that they work reasonably well so long as the dialogue follows predictable, standard lines. However, if the observer is prepared to challenge the system, by responding unpredictably, not cooperating, and so on, then the machine soon starts to reveal itself as just that – a machine.
Whatever its shortcomings, the Turing Test remains a gold standard within artificial intelligence for the recognition of intelligence, if only because no one has been able to propose a satisfactory alternative. However, if you look back quickly at the earlier section on the background to Symbolic AI, you might detect one other test of intelligence that the early researchers had in mind.
From your readings so far, can you think of other indications of intelligence that have often been suggested?
You may have thought of several possible answers here. However, one does stand out for me: the pioneers of Symbolic AI were particularly interested in the idea that the ability to play chess is a clear indicator of intelligence at work.
Turing himself took forward the development of this idea. Chess-playing ability quickly became accepted as another clear test of intelligence. In the 1940s, both Turing and Claude Shannon (the founder of the field we now know as information theory) published papers on the mechanics of a theoretical chess-playing computer. Intensive work in this area followed, until 1958, when Allen Newell, Herbert Simon and Cliff Shaw published 'Chess-playing programs and the problem of complexity', in which they stated:
Chess is the intellectual game par excellence ... It pits two intellects against each other in a situation so complex that neither can hope to understand it completely... If one could devise a successful chess machine, one would seem to have penetrated the core of human intellectual endeavour.
Although the authors might not have realised it, these insights were to define much of the programme of Symbolic AI for the next forty years.
Earlier, I claimed that the key concepts of Symbolic AI were representation and search. How do you think that a chess-playing computer might be based on representation and search?
Think about chess for a moment. Even if you are not a player, you know that there is a board of 64 squares, on which are pieces, in certain positions which change from move to move. It seems clear that if a computer is to play chess at all, it must work with some kind of model or representation of the changing state of the board as the game proceeds from move to move, as the patterns of the pieces shift, and as pieces are taken. You can also see that at any point in the game when it is the machine's move, the program will have to choose the best move to make in the circumstances. This means searching for, and selecting, the best move from among all the legal alternatives at each point.
In his early paper, Shannon had seen exactly this. He envisaged that building a chess-playing program was a three-part problem:
- making a representation of the state of the board that could be stored in a computer;
- finding a search strategy that will find the best move;
- translating this search strategy into a series of instructions that the computer can carry out.
The idea of chess-playing as a key indicator of intelligent thought, realised through representation and search, became cemented into place as a core strategy of the Symbolic AI project.
You'll recall that in 'Computing machinery and intelligence' Turing begins with the question 'Can machines think?'. This in turn implies two preliminary questions: 'What is a machine?' and 'What is thinking?'. We've now looked at Turing's and other AI pioneers' answers to the second of these questions – thinking is essentially something that we can recognise externally through behavioural investigations like the Turing Test; internally it relates to problem-solving procedures based on representation and search.
Whether this is altogether a satisfactory account of thinking and intelligence is a question we will address throughout the course. Let's now consider Turing's answer to the first question: 'What is a machine?'.
Try to sum up what Turing meant by a 'machine' in his paper 'Computing machinery and intelligence'. Refer back to the paper if you need to.
Turing leaves us in no doubt that by 'machine' he means the digital computer. For Turing, a digital computer is a device with a store, an executive unit and a control. The store will contain a 'book of rules' telling the computer exactly what to do next at each step. It can also be used as a scratchpad for storing data and intermediate results. Computers are discrete state machines in that the machine moves through a series of states, the next state being determined unambiguously by the current state and the input the control unit is receiving. The states are discrete because there is no ambiguity or middle ground between one state or another – the machine is either in state 1 or state 2: it cannot ever be in state 1/2. Finally, digital computers are universal machines. They can mimic any discrete state machine simply by adding a new book of rules to the store.
If we substitute the more up-to-date terms 'memory' for 'store', 'CPU' for 'executive unit', and 'program' for 'book of rules', we have the modern computer. This is the machine that the founding fathers of Symbolic AI believed could be programmed to think. The digital computer was at the heart of their project from the start.