The Act of Creation

by Arthur Koestler

  In the making of tools we saw similar bisociative processes at work. Sultan's star achievement was the fitting of two hollow bamboo-rods together into one long rod, by pushing the end of the thinner rod into the opening of the other. Let me describe this experiment in some detail:

  Sultan is given the two sticks, and the banana is placed at a distance from the bar greater than the length of each single stick. For quite a while he tries to reach the banana with one stick or the other (M1 obstructed, random trials). He then pulls a wooden box, which has been used in a different type of experiment, to the bar (fumbling for some M2 to provide the 'mediating performance'). He pushes the useless box away again, pushes one stick outside the bar on the ground as far as it will go and prods it with the second stick towards the banana (M3: Sultan learned long ago to push a longer stick about with a shorter one). He succeeds in pushing the first stick into contact with the banana, but obviously cannot pull the banana in. Nevertheless, he repeats the procedure: when he has pushed one stick out of his reach and it is given back to him, he starts once more. 'But although, in trying to steer it cautiously, he puts the stick in his hand exactly to the cut (i.e. the opening) of the stick on the ground, and although one might think that doing so would suggest the possibility of pushing one stick into the other, there is no indication whatever of such a practically valuable solution [no "insight"]. Finally, the observer [i.e. Köhler] gives the animal some help by putting one finger into the opening under the animal's nose (without pointing to the other stick at all). This has no effect [still no insight]; Sultan, as before, pushed one stick with the other towards the objective. . . .

  He finally abandons the attempt altogether. These tries have lasted over an hour. The chimp was then left in the keeper's care, who reported later on:

  Sultan first of all squats indifferently on the box, which has been left standing a little back from the railings; then he gets up, picks up the two sticks, sits down again on the box and plays carelessly with them. While doing this, it happens that he finds himself holding one rod in either hand in such a way that they lie in a straight line; he pushes the thinner one a little way into the opening of the thicker, jumps up and is already on the run towards the railings, to which he has up to now half turned his back, and begins to draw a banana towards him with the double stick. I call the master: meanwhile, one of the animal's rods has fallen out of the other, as he has pushed one of them only a little way into the other; whereupon he connects them again. [1]

  Henceforth Sultan never had any difficulty in connecting two rods; and later on even three.

  At first sight, Sultan's achievement appears not as an integration of existing skills, but as the invention of a totally new one. However, commenting on the keeper's report, Köhler says: (my italics) 'The keeper emphasized the fact that Sultan had first of all connected the sticks in play and without considering the objective [the banana]. The animals are constantly poking about with straws and small sticks in holes and cracks in their play, so that it would be more astonishing if Sultan had never done this [i.e. poking the thinner rod into the hole of the other], while playing about with the two sticks.' [2]

  Thus the discovery again followed the familiar pattern of a playful habit being connected with a blocked matrix, with chance acting as a trigger. Later on, a third matrix was added: Sultan had learned long ago to sharpen sticks by biting off splinters, so that they could be used to poke in keyholes; now this skill was used to make two sticks fit together. Obviously, Sultan would never have invented this sophisticated method of tool-making if each of the three component skills (raking, poking, sharpening) had not been pre-existing items of his habit-repertory. The more familiar and well exercised each of the matrices, the more likely it is that the animal will solve the problem and, other things being equal, the less it will depend on the helping hand of chance. (In the case of the two bamboo rods which 'happened' to fit each other -- a chance which the animal will rarely encounter in the woods -- we have an example of 'guided learning': the experimenter serves as a match-maker in lieu of the favourable chance event; the rest is up to the pupil.)

  Uniform Factors in Learning

  Sultan's original achievements cannot be explained either by stamping-in, or by spontaneous inspiration out of the blue; they are integrations of existing, flexible skills, of previously unconnected codes of behaviour into more complex codes of a higher order. In conditioning and rote-learning, the new code has to be formed more or less from scratch, more or less bit by bit. This 'drilling in' is a gradual, cumulative process; whereas the bisociation of two matrices appears as a sudden fusion. But in between these opposite extremes we find a graded series of learning methods, with certain basic features common to all. Let me enumerate a few of these common features.

  The chimpanzee, straining to reach a banana behind the bus, remembers the stick lying out of sight; he runs to get it and uses it as a rake. This has been mentioned in the early Gestalt literature as one of the criteria of insight learning. But memory enters into all learning processes. Thus retention in delayed reaction tests has been shown to last in cats from three to thirteen hours, in chimpanzees up to forty-eight hours. [3] Another feature which we find in all types of conditioning and learning, from Pavlov upward, is expectancy and anticipatory behaviour. Once the stage of initial bewilderment or frustrated rage is passed, and the animal has embarked on 'learning to learn', random trials are superseded by less random 'fumbling tries'; and these in turn by hypotheses. At the 'fumbling' stage, Sultan's behaviour in the stick-connecting experiment shows no more insight than the cat's: his fetching of the wooden box and attempts to push one stick with the second, were quite inappropriate. On the other hand, however, we have met with plenty of examples of comparable fumblings among human geniuses -- of wild guesses and inappropriate tries preceding the act of discovery. If we remember that Kepler spent seven years of trial and error, pursuing false inspirations and wrong hypotheses before the discovery of his First and Second Laws, we are led to realize the subtleties and wide applications of the try-method -- and how completely wrong it is to equate it with blind random behaviour. The range of learning by trial and error extends from relatively haphazard tries through the whole graded series, to the relative certainty of inductive inference. On the lower reaches the trials are explicit and often temperamental, like the frantic attempts of Thorndike's cats to get out of the box; on the higher reaches, they assume more and more the character of implicit hypothesis -- Adams' cat moving its head up and down as it 'works out' the means of getting at the liver suspended from the string. Lastly, at the top of the series, the inventor toys with an idea in his head before taking the more explicit step of trying out several alternative sketches on paper; after which he may proceed to making a rough model -- an even more explicit, but nevertheless merely symbolic try. The writer, groping for the right adjective, will sample several with his literary taste-buds; even the Lord Almighty, according to Genesis, proceeded by trial and error, as witnessed by the painful episode of the Flood.

  Criteria of Insight Learning

  So far, then, we have a continuous series of learning methods, where the amount of required stamping-in decreases in proportion to the animal's ripeness for the experimental task. But, according to the contentions of the Gestalt school, there is a decisive break in the continuity of the series which puts insightful learning into a category apart from other methods of learning; and this break is said to be reflected in the animal's characteristic behaviour at the moment the true insightful solution occurs. In this view, Sultan's trial-and-error behaviour was merely a preliminary to the true solution, which emerges with dramatic suddenness and all in a piece; whereas in trial-and-error learning it emerges gradually.

  The chief descriptive characteristics of 'insight' which have been proposed by various authors are as follows: (a) dramatic suddenness; (b) 'the appearance of a complete solution with reference to the whole layout of the field' [4]; (c) the smooth, unhesitating manner in which the so
lution is 'suddenly, directly and definitely' [5] carried out; (d) the solution of the problem precedes the actual execution of it [6]; (e) the solution is retained after a single performance; (f) novelty of the solution. [7] It is furthermore generally assumed that 'insight' is closely related to, if not synonymous with, intelligence and understanding and, by implication, that trial-and-error learning is not so or to a lesser degree so. This last point, however, I shall discuss later; first let us turn to the purely descriptive aspects of 'insight learning'.

  When Köhler's experiments are discussed, the authors usually select the star performances of Sultan, and it is often overlooked that these were rather in the nature of rare limit cases. In the experiment I am going to describe, a young chimpanzee, Koko, was faced with the problem how to get a banana hung high from the wall. The only solution was to push a wooden box underneath the banana and to climb on the box. Though Koko is described by Köhler as 'just as gifted as Sultan', it took him no less than nineteen days to learn this -- whereas he had learned to rake in a banana with a stick in a few minutes. The use of sticks is part of the chimpanzee's repertory of habits -- but there are no wooden boxes lying about in the forest. However, before the experiment was started, Koko was given a small wooden box as a toy; 'he pushed it about and sat on it for a moment'. He was then removed to another cage and in his absence the banana was suspended from the wall, three or four yards away from the wooden box (the italics are Köhler's: by 'objective' he means banana):

  Koko . . . first jumped straight upwards several times towards the objective, then took his rope in his hand, and tried to lasso the prize with a loop of it, could not reach so far, and then turned away from the wall, after a variety of such attempts, but without noticing the box. He appeared to have given up his efforts, but always returned to them from time to time. After some time, on turning away from the wall, his eye fell on the box: he approached it, looked straight towards the objective, and gave the box a slight push, which did not, however, move it; his movements had grown much slower; he left the box, took a few paces away from it, but at once returned, and pushed it again and again with his eyes on the objective, but quite gently, and not as though he really intended to alter its position. He turned away again, turned back at once, and gave the box a third tentative shove, after which he again moved slowly about. The box had now been moved 10 centimetres in the direction of the fruit. The objective was rendered more tempting by the addition of a piece of orange (the non plus ultra of delight!), and in a few seconds Koko was once more at the box, seized it, dragged it in one movement up to a point almost directly beneath the objective (that is, he moved it a distance of at least three metres), mounted it and tore down the fruit. A bare quarter of an hour had elapsed since the beginning of the test. [8]

  All's well that ends well. But it does not. A few minutes later the experiment was repeated -- after the banana had been moved about three yards from its former position, while the box was left standing where Koko had dragged it. When Koko was led back onto the stage:

  he sprang at the new banana in the same manner as before, but with somewhat less eagerness; at first he ignored the box. After a time he suddenly approached it, seized and dragged it the greater part of the distance towards the new banana, but at a distance of a quarter of a metre he stopped, gazed at the banana, and stood as if quite puzzled and confused. And now began a tale of woe for both Koko and the box. When he again set himself in motion it was with every sign of rage, as he knocked the box this way and that, but came no nearer to the objective. After waiting a little the experiment was broken off. [9]

  This tale of woe continued for nineteen days during which the experiment was repeated at varying intervals; and even afterwards, when the new skill was firmly established at last, its performance still alternated for a while with random trials.

  Does Koko's behaviour satisfy the descriptive criteria of insight learning?

  (a) Suddenness. Yes, it does -- because at the climactic moment of the first experiment, the solution did appear suddenly and all of a piece. No, it does not -- because prior to it Koko had made several half-hearted attempts at the correct solution and yet abandoned them. (b) 'Complete solution with reference to the whole lyout to the field'. The answer is, No. (c) 'Smooth, unhesitating, direct and definitive' -- on one occasion, Yes, on the others, No. (d) 'Solution precedes execution of solution' -- yes and no. (e) 'Solution retained after a single performance' -- definitely No. (f) Novelty -- yes.

  Köhler's own comments on this experiment are revealing. Although in The Mentality of Apes he stresses that the gulf between Trial-and-Error and Insight is unbridgeable ('the contrast is absolute' [10]), his comment on Koko's initial hesitations and fumblings with the box is: 'there is only one expression that really fits his behaviour at that juncture: it's beginning to dawn on him!' [11] Let us note that for about ten days after that first success, Koko kept manipulating the box, sometimes aimlessly, sometimes angrily, and during this whole period 'no trace of a solution appeared, except an equivalent of the words: "there's something about that box".' [12] In another passage Köhler says (italics Köhler's): 'It may happen that the animal will attempt a solution which, while it may not result in success, yet has some meaning in regard to the situation. Trying around then consists in attempts at solution in the half-understood situation.' [13]

  Preconditions of Insight

  No more need be said to prove that if we apply the descriptive criteria which I have enumerated, we find a graded series from 'trying around', through the 'dawning' of the solution, to the limit case of the sudden solution. But limit cases at the end of a graded series do not require a separate set of postulates to explain them. The break in actual behaviour, the discrete and unitary character of the solution in these cases can be explained in terms which are also applicable to other forms of learning. Thus with regard to criterion (a) we can say that the suddenness of the solution is due to the trigger action of chance in a situation which was ripe for solution -- that is to say, where the animal's repertory comprises all the requisite single skills, and where all that is needed is a link to combine them into a complex skill -- e.g. Sultan accidentally pushing one rod into the opening of the other. In other cases -- Sultan turning round to pick up the remembered stick -- where chance plays no part, memory provides the link; but memory enters into all forms of learning. Regarding (c) and (d) (smooth, unitary execution of the act, indicating that it has been thought out before being acted out), we may say that the animal has formed a hypothesis, or carried out an implicit try, followed by explicit performances of the act. Rats, cats, and dogs also show this brief suspension of activity, this 'attitude of concentrated attention' [14] before they act out a hypothesis -- which may or may not be the correct one. (e) Retention after a single performance can be interpreted as 'induction based on a single case' -- as the chick, from a single experience, draws the correct empirical inference that all cinnabar caterpillars are to be avoided. Lastly, (f), novelty is of course also achieved when the cat learns the open-sesame trick in the puzzle-box. To argue that the cat's novel response was acquired by Trial and Error, the chimp's by Insight, is to argue in a circle, since novelty itself is supposed to serve as a differential criterion.

  Thus Sultan's Eureka processes, once we have got rid of thinking in S.-R. schemata, are interpretable in terms of the same theory which covers all lower forms of learning. They make a spectacular impression, because (since the separate skills which had to be integrated into the new skill were well-exercised items in his repertory of habits), the problem to be solved was just one step beyond the limits of that repertory, and all was set for a single spark to trigger off the fusion.

  At the opposite end of the learning scale, the dog in Pavlov's laboratory is not equipped with pre-existent rules of the game which could be combined with each other; it must construct a new code of rules, starting more or less from scratch. Therein lies the main difference. The dog must start with an agonized reappraisal of which environme
ntal events are relevant and which are not; then extract and codify the recurrent invariant features from the stimuli promoted to significance; then discriminate between finer features within those features. The rat must piece together, bit by bit, his cognitive map of the maze; the cat must gradually extract, by empirical induction, the rules of Thorndike's game from a surrealistic universe.

  The 'missing link' in between the cat and Sultan is provided by Koko. He does not have to start from scratch; he has already played with the box; he has sat on it and pushed it about. Was his first successful solution of the problem a random try? Certainly not. It had all the 'dramatic suddenness, smoothness, directness and definiteness' that one can wish for. It was more than a 'provisional try', rather like a hypothesis which carried implicit conviction; yet on the other hand, it had been preceded by hesitant action along the correct lines which was abandoned; and it was succeeded by forgetting all but a fragment of the successful solution -- the fragment 'there is something about that box'.