A dual process theory of thinking
We started this book with our account of rationality, and we end it with our theory of thinking. We believe that one leads to the other in natural fashion with the dual nature of rationality being mirrored, to an extent, in dual systems of underlying thought. Let us briefly recap and round off the argument to this point.
The distinction between rationality1 and rationality2 was originally introduced to solve an apparent paradox of rationality (Chapter 1). This is that the human species is obviously highly intelligent, being able to achieve many practical and even abstract goals, and yet the psychological literatures on reasoning, judgement and decision making seem to be littered with evidence of error and bias. To resolve the problem, we have distinguished between people’s rationality1, which consists in their ability to achieve many personal goals, and their more limited rationality2, which shows in their violation in many experiments of the rules of logic and decision theory. As argued in Chapter 6, human beings without special training do have some deductive competence. This modest ability, combined with their higher degree of rationality1, makes it no real paradox that they have done so well, and even been able ultimately to create logic and decision theory as formal systems.
In discussing this distinction throughout the book, we have been able to show a number of reasons why people may be irrational2 without being judged to be irrational1. First, the normative systems may themselves be inadequate. We have pointed to the limitations both of formal logic as a standard for good reasoning (Chapter 1) and of formal decision theory as a standard for good decision making (Chapter 2). Secondly, we have shown that behaviour often appears irrational only in the sense that people fail to observe the experimental instructions, whereas the processes applied would normally be effective in everyday life (a point also made by Sperber et al., 1995). For example, a “positivity bias” may be a generally adaptive feature of thinking with a limited capacity system and yet lead to errors described as confirmation or matching biases under particular experimental conditions (see Chapters 3 and 5). We have shown (Chapter 5) that people’s thinking is habitually influenced by prior belief, a tendency that can be highly rational1 but may lead to belief biases in experimental tasks.
We are not saying that people are invariably rational1 because any rationality is clearly bounded by cognitive constraints. Of course, people may fail to achieve their goals because they have not had the opportunity to learn the strategy required or because the task exceeds their cognitive processing capabilities. Particularly difficult to achieve are high-level and long-term goals, conceivable only in language and requiring much explicit processing. Consider the case of probabilistic reasoning. On the one hand, we have evidence that people can detect contingencies and probabilities very accurately, as measured by implicit learning methods (see Reber, 1993). On the other, the literature on statistical reasoning and judgement provides much evidence of bias and fallacious inference, particularly on unfamiliar problems that are merely described to subjects. Some of these biases can certainly have irrational1 consequences. For example, a tendency to generalise from small or biased samples of information in novel cases will clearly lead to some poor decision making. A plausible explanation for this is that it is very difficult to abstract principles of probabilistic reasoning in a purely inductive way, without knowledge of task structure—a problem that is intrinsic to the nature of probability (see Einhorn, 1980). It is clearly important to recognise contexts where formal probability theory has to be applied to achieve effective reasoning and decision making. At the very highest level, it is essential for scientific research, but it is sometimes needed in ordinary affairs as well.
In discussing rationality in this way, we believe that some clear psychological proposals have also arisen. A major advantage of our approach is that there is no ground for making a sharp distinction between reasoning and decision making. All the tasks we have discussed in this book involve the same general stages of problem representation, judgement or inference, and decision/actions. Exactly the same psychological problems arise in accounting for reasoning and decision making. For example, we need to ask (a) how the problem information gets represented, (b) what causes people to focus selectively on some aspects and ignore others, and (c) what mechanism is responsible for the retrieval of relevant prior knowledge? There is no clear distinction either in terms of the processes applied to produce the response to the task: decision making may require explicit reasoning for example, whereas response to reasoning problems may require only intuitive judgements. We hope in particular that our extended discussion of the Wason selection task (Chapter 4) has served to illustrate the false distinction between reasoning and decision tasks. As we have shown, the most investigated single problem in the literature on “reasoning” can best be understood by an analysis in terms of judgemental and decision processes.
The most important psychological distinction, however, is the one we focus on in this final chapter—that between explicit and tacit cognitive processes. In Chapter 3, we introduced our thesis that our thinking is highly focused on what are subjectively relevant features in the task information and from memory. We argued that such focusing can be rational1 and is achieved by preconscious and tacit processes. What we are aware of and what we think about is therefore determined mostly by tacit, implicit processes. We argued that such processes are computationally extremely powerful and operate in a very rapid and almost certainly parallel manner. In fact, we characterise our implicit system in much the same way as do neural network theorists and feel that connectionist models provide a promising approach to understanding how this system works. Such tacit systems can clearly be called, in a sense, “inferential”—for example, in applying general knowledge to specific cases. It is important, however, to distinguish such implicit inferencing from explicit reasoning of the kind discussed in Chapter 6, which involves manipulation of explicit propositional representations.
The notion of implicit processing has also been crucial in our explanation of “biases”. However, we are not adopting a behaviouristic or epiphenomenalist stance in which conscious thinking serves no purpose. On the contrary, we wish to propose a dual process theory of thinking in which tacit and parallel processes of thought combine with explicit and sequential processes in determining our actions. In Chapter 6, for example, we argued both that there is evidence of deductive competence in reasoning, and therefore of a limited capacity to be rational1 and also that such reasoning arises from an explicit verbal process. Support for the latter claim comes from the extent to which rational2 processes, unlike rational1 processes, are subject to the influence of verbal instructions.
Although not previously linked to a theory of rationality, the dual process theory we present has origins in much earlier theorising of the first author in the area of deductive reasoning. However, the emphasis of the theory has shifted over the years in particular with regard to the relationship between the two processes. We discuss this issue first, and then proceed to look more closely at what is known about both tacit and explicit thought processes.
Dual Processes in Reasoning:
Sequential, Parallel, or Interactive?
In order to illustrate the issues involved we give a brief summary of the history of dual factor and dual process accounts of reasoning. Evans (1982) in reviewing the then state-of-the-art in the field of deductive reasoning, presented a descriptive two-factor theory with the claim that almost all reasoning tasks show evidence of a logical and non-logical component of performance. Despite all the large amount of further research that has been conducted since then, we see no reason to revise this analysis, and much of our discussion of competence and bias in Chapter 6 is built upon this dichotomy.
In the two-factor theory, logical and non-logical processes were characterised as competing processes and formalised as an additive probability model by Evans (1977b) who provided statistical evidence of within-subject conflict between the two factors, for example by showing statistical independence between selection rates of different cards on the selection task. The two-factor, within-subject conflict model was also applied by Evans et al. (1983) to explain their belief bias findings. Of particular importance were the results of protocol analyses which showed firstly that, when subjects attended to the conclusion more than the premise, they were more likely to show a belief bias. The crucial evidence for the conflict model was that these two modes of reasoning were not associated with different subjects. (Recently, however, George, 1991, has presented some evidence that subjects differ in the extent to which their reasoning is belief based.)
The descriptive two-factor theory was also cautiously linked to psychological assumptions about conscious and unconscious processes in what was termed the “dual process theory of reasoning” (Evans, 1982, Chapter 12; Evans & Wason, 1976; Wason & Evans, 1975). This theory was motivated by the need to account for discrepancies between introspective reports and experimental evidence of bias on the Wason selection task. Wason and Evans investigated the reasons people gave for choosing cards both with the usual “if p then q” conditional, and also with the form “if p then not q” where the correct choices are greatly facilitated because they are also favoured by matching bias. The striking results were for subjects who received the negative consequent conditional first. Most of course, chose p and q, but also stated that they were turning the cards to seek falsifying combinations—apparent evidence of “full insight”. The same subjects, subsequently given the affirmative conditional also tended to choose p and q, and now gave verification explanations in line with “no insight”.
Wason and Evans thought it implausible that subjects possessed an insight on one version and not on the other and instead proposed their dual process theory. They argued for a distinction between type 1, unconscious, and type 2, conscious processes. They proposed that matching was an unconscious type 1 process that led to the card choices and that the type 2 conscious processes served only to rationalise the choices made. Evans and Wason (1976) went on to show that subjects would happily provide a verbal justification for any common solution to the selection task that was proposed to them as the “correct” solution.
The heuristic-analytic theory of Evans (1984, 1989) has already been discussed at a number of points in this book. This theory developed the earlier dual process theory in several ways. First, the heuristic (type 1) processes were better specified as providing “relevant” problem representation, and the specific if and not-heuristics were proposed to account for abstract selection task performance and other aspects of conditional reasoning. The analytic (type 2) processes were upgraded from a purely rationalising role to form the basis of the logical component of performance in the two-factor theory, albeit by an unspecified mechanism. This development depended on the critical insight that the Wason selection task is a most unrepresentative example of a reasoning task—one in which logical inferences, rather than relevance judgements, are rarely employed.
Though we generally regard the heuristic-analytic theory as an advance, there is a respect in which we now prefer the proposals of the original dual process theory, a name which we would now like to revive. Evans (1989) was particularly concerned to explain the evidence of biases in terms of relevance effects in problem representation. Hence the analyses of that book were mostly concerned with how tacit thought processes lead to selective representation of problem information. The impression created, however, was that the analytic processes take over where the heuristic processes leave off. This sequential model is clearly at odds with the conflict and competition model in the Evans (1982) two-factor theory. Wason and Evans, on the other hand, discussed the possibility of a dynamic interaction between type 1 and type 2 processes.
The dual process theory that we now favour is neither a sequential model nor a conflict model, but rather an interactive model. We would agree with Evans (1989) that the processes responsible for problem representation and focusing are primarily tacit. However, we also believe that inferences and actions often result from implicit processing without intervention from the explicit system. When we look at inferences drawn and decisions made it seems to us that these can reflect either tacit or explicit processes. For example, we can make a decision in a quick “intuitive” way or by conscious elaboration. Subjects on a reasoning task may make an effort at explicit reasoning in accordance with the instructions and they may succeed in deriving the conclusion in this way. Alternatively, they may give up and choose a conclusion that “feels” right. The interactive nature of the two processes lies in the fact that our conscious thinking is always shaped, directed, and limited by tacit, pre-attentive processes. For example, on the selection task not only is people’s conscious thinking restricted in general to the matching cards, but their thought about the consequences of turning the cards is also limited to the matching values on the other side (Evans, 1995b; Wason & Evans, 1975). Clearly such a finding is not consistent with the idea of an implicit stage followed by an explicit stage.
We are pleased to discover that Sloman (1996) has recently argued the case for two systems of reasoning that bears many similarities to the dual process theory we support here. He distinguishes between associative reasoning—which he assumes to be based in connectionist systems—and rule-based reasoning. By rule-based reasoning he refers to symbol manipulation and does not attempt to adjudicate the debate between inference rules and mental models. We agree with him that these two theories share a common agenda that we have described as the deductive competence problem (Chapter 6). Sloman’s discussion also roughly indicates an implicit/explicit distinction similar to ours. First, he argues that awareness of process is a rough if fallible heuristic for distinguishing the two processes. He argues that with associate reasoning we are aware only of the product—say a judgement—and not the process. However, he is cautious about awareness of process with rule-based reasoning, owing to the interpretational problems of introspective report (so are we, see below). The other way in which Sloman’s two reasoning systems are similar to the implicit/explicit distinction is that he relates his distinction to automatic versus controlled processing. Awareness and control are, of course, two defining characteristics of consciousness. Sloman also argues that “the rule-based system may suppress the associative system but not completely inhibit it”. This accords with our own analysis that conscious reasoning can overcome habitual tacit processes, but only to a limited extent (see Chapter 6).
The Nature of Tacit Thought Processes
Discussion in this book has been focused principally on giving rational1 accounts of behaviour on reasoning and decision tasks—for example by reinterpreting evidence of biases—so we have already provided much discussion of psychological processes that we regard as primarily tacit. We obviously cannot make a direct equation between rationality1 as we have defined it and successful type 1 implicit processes, or between rationality2 and correct, by some normative theory, type 2 explicit processes. The notions are, however, very closely correlated for reasons we will explain.
As we have argued, rational1 reasoning and decision making happens when we make practical decisions that help us to achieve our personal goals. Much of the time we do this very effectively and by use of habitual type 1 processes. As indicated earlier in this book, adaptive tacit processes may be either innate or learned. We do not wish to take a strong position on the argument that cognitive processing reflects the possession of innate, specialised modules (e.g. Fodor, 1983; Chomsky, 1986) although we accept that systems such as language and visual perception could not be learned within the lifetime of an individual without a strong innate basis. We are currently unconvinced, however, that higher-level thought involved in reasoning and decision making can plausibly be accounted for in terms of specific evolutionary mechanisms (e.g. that proposed by Cosmides, 1989) and feel that many rational1 actions reflect specific learning experiences of the individual. Nevertheless, the mechanism by which such learning and application comes about is clearly subject to biological constraints such as, perhaps, the first (cognitive) principle of relevance proposed by Sperber et al. (1995—see Chapter 3).
Our assumption is, therefore, that that tacit processes primarily reflect (biologically constrained) learning, and that we have learned how to do most of the things that we require to do to achieve our everyday goals. It is fully necessary that this is the case, because conscious thought is a very scarce and limited resource. The execution of the most ordinary tasks, such as travelling to work, demand extraordinary amounts of information processing, and the interaction of both very large stores of prior knowledge and very complex processing of environmental stimuli. If you do not immediately agree with this statement, then think what would be needed to design and program a robot that could—without external assistance—perform your own journey to work.1 Yet so automated are your own processes, that you will probably not require much application at all of your conscious thinking for travelling to work, and will instead have this resource available during the journey for planning your day’s activities.
As already indicated, there are occasions where to be rational1 you must apply conscious resources, or type 2 processes. It is also possible to conform to a normative theory by entirely type 1 processes. We discussed examples in Chapter 2, where we pointed out that even the foraging behaviour of a bumble-bee can be seen to comply with normative decision theory. We also pointed out that, by our definition, a bumble-bee cannot have rationality2. That implies having good reasons for one’s reasoning or decision making, and this in turn means following rules, sanctioned by a normative theory, in which propositional representations are manipulated. In this case, not only would there have to be propositional representations, but these would have to have numerals attached to them, so that the SEU rule for calculating subjective expected utility could be followed. Surely bumble-bees do not follow any such rules, as opposed to merely conforming to them—i.e. producing behaviour in another way that gives the same result as following the rule would have.
Let us consider a human example to illustrate how closely type 1 and type 2 processes, and rationality1 and rationality2, can be related to each other. Our visual system is highly reliable but sometimes gives us misleading representations of the world. This system has reliable type 1, implicit processes that do not follow normative rules for manipulating propositional representations. But after processing, it presents us with a representation of the world that is sometimes inaccurate. An example would be when we are looking at a fish in water, which would have an illusory location. Now we could learn by type 1, implicit processes to compensate for this inaccuracy, just by throwing a barbed spear at some fish again and again. We would explicitly note when we had failed to spear one and try again. Eventually we could learn how to make a small compensation for our visual illusion, without however being explicitly aware of what this was, and be able to spear fish with great success. In this process we would display mostly rationality1.
We could carry on and explicitly investigate the illusion and how we compensate for it. To do that, we could become fully scientific and follow logical and probabilistic rules for reasoning, in type 2, explicit processing. We would then be rational1 and that would be necessary for acquiring scientific knowledge of what was going on. Even so, we would have to rely on many type 1 processes, and be rational1 again, to gather our scientific data and make relevance judgements for our inferences in our investigation, and of course we would still experience the illusion no matter how extensive our scientific knowledge. We are conscious of and can control to some extent our type 2 processes, but not our type 1 processes. To clarify further the difference between implicit and explicit thought process, some valuable insights can be gained by looking at the literature on implicit learning, to which we now turn.
Implicit learning and tacit processing
Fortunately, the interesting and important field of implicit learning has recently been reviewed and discussed in two books by Reber (1993) and by Berry and Dienes (1993, and see Evans, 1995a, for an extended review and discussion of these books). One way to define implicit learning is that it occurs below a subjective threshold—the subject is unable to report what has been learned. Frequently such learning is implicit in a second sense—subjects may lack awareness that they have learned anything at all. Berry and Dienes take the stronger position that there are separate implicit and explicit cognitive systems that differ in respects other than awareness. For example, they argue that implicit learning leads to only limited and specific transfer to other tasks compared with explicit learning. They also say it gives rise to a phenomenal sense of “intuition”—a point to which we will return. The best evidence that they offer for this view of implicit learning is in control tasks of the sort studied by Berry and Broadbent (e.g. 1984). Here subjects learn by experience to control and predict systems, such as factory production, simulated on computer without acquiring verbal knowledge of how they are doing it.
Berry and Broadbent (1988) proposed that in implicit learning mode subjects store all contingencies in an unselective manner whereas explicit learning is much more focused. They suggested that explicit processing is more similar to what happens in problem solving where subjects evaluate explicit hypotheses, apply explicit rules, or manipulate mental models. They showed that implicitly acquired knowledge would transfer between similar domains but not across problems where the domain was different and only the principle was the same at a higher level of abstraction. They also found that transfer between similar tasks was inhibited if subjects were informed verbally of the connection between them, suggesting that this induced a less effective explicit mode of learning. Reber (1993) also discusses a number of examples in his own experiments where instructions inducing explicit learning strategies impede performance.
Like Berry and Dienes, Reber (1993) argues for separate implicit and explicit systems. One point on which he agrees with Berry and Dienes is that implicit learning is relatively robust—in particular, implicitly acquired knowledge is less affected by neurological damage than is explicitly acquired knowledge. This can be taken as evidence for the connectionist nature of implicit processing. Reber claims also that implicit processes are less variable across individuals and to a large degree independent of age and IQ. He uses these features to build an evolutionary argument for what he calls the “primacy of the cognitive unconscious”. Tacit processes evolved first, are common with other higher mammals, and are relatively robust and invariant.
Because Reber discusses the unconscious, he feels obliged to define what he means by consciousness. He argues that although there is a sense of consciousness that is common with other animals, there is also a sense of the term—he calls it Consciousness II—which is uniquely human:
The kind of consciousness we typically imagine when we think of our own sense of awareness is one that functions not merely to differentiate self from others but one that incorporates a large number of functions that allow us to modulate and refine the actions of self (p. 136).
He goes on to argue that in most of his experiments subjects are aware that they have learned something although they do not know what it is. He adds: “But this kind of consciousness is not really Consciousness II, it is something less well developed … is a function that contains awareness but lacks the self-reflecting, modulating functions.”
Now, if there is a uniquely human consciousness, Consciousness II, that evolved late and which is—unlike implicit systems—related to age, IQ, and so on, as Reber claims, then we should ask (a) what specifically is its function and (b) what have studies of implicit learning informed us about what this function is? We return to this point when we consider the nature of explicit thinking.
One of Berry and Dienes’ characteristics of implicit learning is that it gives rise to a phenomenal sense of intuition. “Intuition” is a term used to describe the experience of processes whose end-product only is posted in consciousness. Most human expertise is intuitive because judgements have to be exercised in complex domains with the interaction of many variables. This is why knowledge elicitation for expert systems has proved such a difficult exercise (see Evans, 1988). For example, an expert chess player thinks about only a small set of the possible moves when analysing a position, and a small set of replies and so on, thus restricting an otherwise unmanageably large problem space. However, the process by which moves are selected for consideration is tacit and unreportable. The problem is that much of the intelligence or expertise lies in these intuitive processes: better players think of better moves to consider in the first place.
Intuitive judgement has been studied extensively in cognitive and social psychology. Statistical or probabilistic judgement has been the main focus of study in the “heuristics and biases” tradition of Kahneman and Tversky. People may be asked, for example, whether a sequence appears to be random, whether two variables seem to be correlated, or which of two samples appears to give better evidence for a hypothesis. Such research is marked by widespread claims of bias: for example, people have the “gambler’s fallacy” expecting sequences of random events to alternate outcomes; they see “illusory correlations” in data biased by their prior expectations and so on (see Baron, 1988; Garnham & Oakhill, 1994, for review of these literatures). Kahneman and Tversky have proposed various “heuristics” as explanations of such biases, most notably representativeness where judgements of probability reflect detection of similarity between features of events and hypotheses (Kahneman & Tversky, 1972b) and availability in which probability of single events is judged by the ease with which examples can be brought to mind (Tversky & Kahneman, 1973).
Of course, such judgements are only correctly described as intuitive if they reflect tacit processing. Kahneman and Tversky did not give a precise account of their heuristics, and in particular did not state whether these operate at a conscious or unconscious level. In the case of the availability heuristic, the processes that determine what “comes to mind” seem to be the same kind of tacit processes that we have discussed as determining selective representations in reasoning, by a combination of salience of the stimuli presented and retrieval from memory of associated and relevant knowledge and limited by processing effort. Availability must be closely related to the idea of relevance and focusing effects discussed in Chapter 3 and elsewhere in this book. The only sense in which this heuristic could be explicit is if the subject was to reason, “I can think of several examples of this event, so it must occur often.” We think it more likely that availability of examples would confer an intuitive feeling of probability.
There is evidence to suggest that at least in the case of subjects lacking training in statistics, probability judgements are indeed tacit. A series of studies by Nisbett and colleagues on people’s understanding of statistical principles in everyday reasoning are relevant here. Nisbett, Krantz, Jepson, and Kunda (1983, Study 4) gave subjects problems in everyday domains that involved the law of large numbers: for example, they might have to explain why players showing good form in early season sporting results do no better than the rest in the longer term. Subjects were given a choice of several explanations, one of which was the correct, statistical reason. The problems were presented in two domains—sports and acting—and the subjects had varied experience in the two domains.
What Nisbett et al. found was that a majority of subjects with experience in the domain chose the statistical explanation, whereas a majority of those inexperienced in the domain in question chose a deterministic account. This suggests that understanding of the law of large numbers can be acquired by experience, but in a domain-specific manner. In a separate study, Jepson, Krantz, and Nisbett (1983) asked subjects to give explanations of events in a large number of domains and classified their open-ended answers as statistical, determinate, or intermediate. Across the nine domains studied, the frequency of statistical explanations ranged from as low as 5% to as high as 93%. Bearing in mind Berry and Dienes’ criterion of specificity of learning and transfer, these findings again suggest that an intuitive understanding of the law of large numbers may be acquired by tacit processing in a domain-specific manner.
In accordance with the dual process theory, just because one has initially learned to comply with a rule implicitly does not mean that one cannot come to learn it and follow it explicitly. Smith et al. (1992) have recently argued that people can follow the law of large numbers as an abstract rule. One piece of evidence they refer to comes from Nisbett et al. (1983), where subjects made appropriate use of sample size and variability when reasoning about the colour of imaginary birds and people on an imaginary island. However, we are not sure that subjects are following a rule to think about this case. After all, they are very familiar with birds and other people, and could have learned only implicitly to take account of the greater colour variability in birds than in people. All ravens may be black, but the males and females of many bird species differ in colour.
A good test, as Smith et al. (1992) also point out, would be whether teaching someone the explicit verbal principle of the law of large numbers would benefit their everyday reasoning. Fong, Krantz, and Nisbett (1986) showed that statistical answers to problems couched in everyday terms increased dramatically as a function of the amount of formal statistical training the subjects had received. Given the normal method of teaching statistics—by verbal instruction—this suggests a benefit of explicit thinking. Stronger evidence is provided by experimental training in studies in the same paper. For example, in Experiment 1 a control group without training gave 42% statistical answers; a group given (explicit) rule-based training 56%, a group given (implicit) examples-based training 54%, and a group trained by both methods 64%. The benefits of rule-based training certainly give evidence that subjects may bring explicit principles to bear in such tasks. Whether the implicit, examples-based training leads to implicit or explicit knowledge of the principle is moot. Studies of concept and rule learning show that people cam learn to conform to a rule in a wholly tacit manner, but that they can be led to explicit knowledge of it, depending on factors such as the complexity of the domain and the mode of learning (see Berry & Dienes, 1993, Chapter 3).
The points discussed here illustrate that people’s conformity with a rule or normative system does not necessarily qualify them as rational2. However, the evidence also suggests that people can learn to follow rules explicitly and thus increase their rationality2. Compare human beings again with bumble-bees. The bees’ behaviour is determined by the quality of the food they get from a small number of flowers they have recently visited. They may have a very limited memory, but they do well with this heuristic in their natural environment, in part because of the variability of their food there (Real, 1991). Human beings can be much more flexible and even eventually learn to follow the law of large numbers explicitly, taking account of both sample size and variability in different circumstances. We now turn to further consideration of the explicit thinking system, which makes us far and away the most adaptable of animals.
The Nature of Explicit Thought Processes
It is clear that the major benefit of the implicit system is its vast computational power, allowing very complex acts of information processing to take place very rapidly. Conscious thought, by contrast, is slow and very limited in capacity. We have shown already that such thinking is highly focused on selected information: we may have difficulty thinking about more than one hypothesis or more than one mental model at a time. As we do not regard consciousness as an epiphenomenon, it is important to consider what functional advantage it gives to us. The answer to this can be found by considering the limitations of the implicit cognitive system, and in particular what aspects of human intelligence cannot be achieved in a tacit manner.
We have already seen that implicit systems lead to learning that is relatively inflexible and domain specific—evidence, perhaps, for the operation of innate modules. Such processes are at least in part shaped by our past history of personal learning: the perceptions, judgements, and actions they prescribe are those that are effective in the environment in which the underlying neural networks have received their training. The advantage of the dual process system is that conscious reflective thought provides the flexibility and foresight that the tacit system cannot, by its very nature, deliver. Most of the time our decision making is automatic and habitual, determined by past learning, but it does not have to be this way. We can make conscious decisions based upon analysis of a novel problem, projection of mental models of future possible worlds, and calculations of risks, costs, and benefits. Granted we are not very good at conscious decision making, just as we are not very good at deductive reasoning, because of severe cognitive constraints. The most striking of these is our very limited span of attention already discussed. Acquisition of effective explicit thinking skills is also very hard work, in contrast with the automatic and apparently effortless acquisition of our tacit and intuitive processes. The point is, however, that consciousness gives us the possibility to deal with novelty and to anticipate the future. We can be rational2 as well as rational1.
Although the concept of explicit thought links with consciousness, it is also connected with language. In making this point, we do not overlook the fact that the mechanism of language operates primarily at a tacit level and may well reflect the possession of innate modules (see, for example, Chomsky, 1980, 1986; Fodor, 1983). Our point is that human beings’ unique possession of language is almost certainly a prerequisite for our special form of reflective consciousness—Reber’s Consciousness II—and hence for our facility for explicit, rational2 thought. We accept the evolutionary argument that tacit processing is shared with animals who can thus be rational1, but assert that rationality2 is uniquely human. We have already indicated ways in which explicit thinking is linked with verbal processes. For one thing, such thinking appears to be frequently verbalisable—i.e. present in verbal working memory and able to be “thought aloud”. Deductive processes are, as we have shown, also highly influenced by verbal instructions. We would not go so far as to say that language is the sole medium of explicit thinking, but it is hard to see how Consciousness II would have been achieved without it. For this reason we are interested in the argument of Polk and Newell (1995) that deductive reasoning processes are verbal and very closely linked with linguistic comprehension processes.
It seems to us that the stream of consciousness is a channel of not just limited but apparently fixed capacity. We cannot increase the capacity when a problem of complexity needs to be dealt with: nor can we reduce it when there is nothing to be done. Never empty, the channel fills with daydreams or occupies itself in gossip or with the vicarious experience of a television show when it has no pressing employment. Goal-directed, explicit thinking—aimed at problem solving or decision making—is also highly demanding of effort and concentration that noticeably tires the individual. Thus we may seek to delay such tasks as filling in a tax form which require a conscious effort, and prefer activities such as idle gossip, which involve far more complex information processing, but primarily at a tacit level with the results delivered effortlessly to our consciousness.
In attempting to study the explicit system of thinking and reasoning, we must be aware of the limitations of self-knowledge and verbal reporting. These issues have been discussed in detail by Evans (1989, Chapter 5) and will be reviewed only briefly here.
Self-knowledge and verbal reports
The notion of explicit, conscious thought is bound up with the problems of introspection and verbal report. Folk psychology includes the belief that we make decisions for conscious reasons that we can report. Vast industries of opinion polling and market research are built upon this belief. People are asked, for example, not only how they will vote but why. They may be given lists of issues and asked to identify which ones are influencing their voting. Politicians and political journalists take very seriously the results of such surveys. From the perspective of the dual process theory this is all very dubious. First, many decisions and actions result directly from tacit, judgemental processes that are by definition inaccessible to consciousness. Secondly, even where explicit thought is employed, there is no guarantee that this will confer insight and accurate introspective reporting.
In the Wason and Evans (1975) study, discussed earlier, subjects were asked to give reasons for their choices. This they did, using type 2 explicit verbal reasoning, relating their actions to the experimental instructions by arguing that the cards turned could prove the rule true or false. However, this reasoning was evidently serving the function of rationalising choices caused by tacit processes such as matching bias, a conclusion supported by the recent study of inspection times and protocols reported by Evans (1995b, and in press—and see Chapter 3 in the present volume). The famous paper of Nisbett and Wilson (1977) demonstrated effects consistent with this conclusion across a range of cognitive and social psychological experiments. They argued that people have no direct or special access to their own thought processes. According to Nisbett and Wilson, when people are asked for introspective reports what they actually do is to theorise about their own behaviour. Although such self-theorising is potentially advantageous, such theories are often erroneous. Thus journalists commissioning opinion polls may be engaging in self-fulfilling prophecies. Their newspaper articles provide people with theories, for example, of why one should vote for particular political parties. When opinion polls then ask people to give their reasons for voting, these theories are delivered back.
Those who criticise introspection may find themselves attacked as anti-rationalist. (For a summary of hostile reactions to Nisbett & Wilson, see White, 1988.) For example, Fellows (1976) once accused the first author (Evans) of regarding human beings as “non-verbal animals” and went on to add:
If Evans dismisses the subject’s reports as rationalisations, then logically he must dismiss his own explanations in the same way.
This line of attack misses the crucial point about the dual process theory. The rationalisations of Wason and Evans and the causal theories of Nisbett and Wilson are clear evidence of the existence of type 2 explicit verbal reasoning. Of course, Fellows is right to say that the kind of thinking used by Evans to write the discussions of his papers is the same as that used by subjects to rationalise their choices. Both are theorising in an explicit verbal manner. However, it is the psychologist, not the subject who has privileged access in this situation. The psychologist knows the previous literature on the topic and also knows the experimental design, what subjects in other groups were doing, and much else besides. Hence the researcher will have a better chance of getting the theory right than the subject. The point is not that subjects do not reason explicitly and verbally—they evidently do—but that they lack insight into the causes of their own behaviour. Even when they give correct accounts, it can be argued that this is because they hold a good theory of their behaviour rather than direct access to their thought processes (see Nisbett & Wilson, 1977).
Now, as we have argued, we do not regard explicit thinking as simply serving to rationalise behaviour, and believe that decisions and actions can result from explicit processes. The findings discussed here, however, militate against introspective reporting as a method of studying such processes. When asked how or why they have performed an action, or to predict future behaviour, people appear to access and apply theories of their own behaviour through their explicit thinking system. Such theories are often inaccurate and cannot be relied upon. So how can we investigate the role of explicit thinking in reasoning and decision making? One method that we have already discussed (e.g. in Chapter 6) is to investigate the effects of verbal instruction on reasoning performance. Tacit thought processes, we assume, are non-verbal and not responsive to such instructions. The other method is that of verbal protocol analysis.
In considering the difference between introspective report and protocol analysis we are happy to adopt the broad position advocated by Ericsson and Simon (1980, 1984). Contrary to much popular mis-citation, Ericsson and Simon did not refute Nisbett and Wilson’s arguments. They also do not expect reports to be accurate if they are (a) retrospective and (b) invite subjects to state reasons for behaviour, or to make strategy reports. What Ericsson and Simon argue for is a method of verbal protocol analysis that differs from introspection in two crucial regards. First subjects report their thoughts concurrently, not retrospectively so that the current contents of short-term memory can be reported before they are forgotten. Next, the subjects only report their locus of attention—what they are thinking about. The process of thinking is to be inferred by the psychologist, not described by the subject.
Ericsson and Simon (1984) say that such verbal reports will be incomplete because there are automatic and “recognition” processes that do not register as sequential steps in verbal short-term memory, or working memory as we might better describe it. The processes they refer to are evidently part of our type 1, tacit system. However, the processes that can be traced by verbal protocol analysis are those that make use of working memory. Protocols can reveal the locus of attention—what information is currently being heeded—and may be informative about the goals that the subject is pursuing. They also reveal intermediate products and stages of problem solving. Even tacit processes may deliver their final products to consciousness. Complex problem solving involves lots of sub-stages—e.g. by pursuit and solution of sub-goals—which register their products in reportable working memory, allowing the processes as a whole to be traced and interpreted.
With these methodological considerations in mind, we turn finally to the mechanisms of explicit reasoning, picking up from where we left the issue of deductive competence in Chapter 6.
The mechanism of explicit reasoning and decision making
In discussing the inference rules versus mental models accounts of deductive competence (Chapter 6) we indicated a general preference for the models approach on the grounds that it was more plausible and psychologically richer. However, in the context of the current chapter we must raise a serious concern about the research programme on reasoning by mental models: the neglect of the problem of consciousness and explicit reasoning. Johnson-Laird and Byrne (1991, p.39) in virtually the only reference to consciousness in their entire book on reasoning, comment:
The tokens of mental models may occur in a visual image, or they may not be directly accessible to consciousness. What matters is not the phenomenal experience, but the structure of the models.
We find this statement very unsatisfactory and note to our surprise that in the massive body of experimental work summarised by Johnson-Laird and Byrne, none appears to include evidence based on verbal protocol analysis. Johnson-Laird and Byrne appear to be adopting an epiphenomenalist stance that just does not stand up to the evidence. First, we have shown that people only make an effort at deductive reasoning when explicitly instructed to do so. Although we welcome Johnson-Laird’s (e.g. 1993) recent extension of mental model theory to the realm of induction, we must ask why the validation stage (search for counter-examples) only arises when deductive reasoning instructions are given, unless the process is in some way explicit? The evidence that belief bias can be inhibited by instructions giving extra emphasis to the principle of logical necessity (Evans et al., 1994) can be interpreted as evidence for the model theory, but only on the assumption that the effort to find counter-examples is under conscious control.
The other difficulty here is the claim that mental models occupy space in working memory—a central proposal for the explanation of differences in difficulty between syllogisms, namely that multiple model problems are more difficult than single model problems (Johnson-Laird & Bara, 1984). Now in the widely accepted Ericsson and Simon model, the contents of working memory are generally regarded as conscious and reportable. We do not argue that the process by which mental models are manipulated can be reported, but we find it most implausible that this manipulation would not manifest itself in any way in a verbal protocol. In particular, it should be possible to tell whether subjects are thinking about single or multiple models. For example, if subjects consider alternative possibilities because they have been instructed to do so, it seems to us that this must show up in a verbal protocol.
Inference rule theorists have produced verbal protocol analysis in support of their arguments for rule-based reasoning—O’Brien, Braine, and Yang (1994) and Rips (1994) provide important evidence of this type. O’Brien et al. gave their subjects relatively long reasoning problems and asked them “… to write down everything that followed logically from the premises in the order that inferences occurred to them …The point is that the steps in the solution of one of these problems should be in a certain order if the subjects are following logical rules. Most subjects did report the predicted order in their protocols. These results do establish that the subjects are following rules in step-by-step fashion under some conscious control, and not merely conforming to logic by coming up with the logically correct overall solution in some other way. But it is more open whether they are following natural deduction rules, as the inference rule theorists would contend, or following rules for constructing mental models.
In replying to O’Brien et al., Johnson-Laird, Byrne, and Schaeken (1994) acknowledge the importance of studying the intermediate steps subjects report themselves as using in solving logical problems. Johnson-Laird et al. admit that their model theory does not yet cover the order in which people will construct models to solve problems like those in O’Brien et al., but claim that this can be done in a way consistent with the data by supposing that “… reasoners will tend to start with a premise that is maximally semantically informative …”. They give simple examples of what they mean, but we would argue that they need much more than this if their theory is to make predictions precise enough to be tested by experiments like those reported by O’Brien et al. In general, Johnson-Laird et al. must introduce some definition of semantic information, or as we would say, epistemic utility, into their theory, and use that to predict the order in which subjects will construct models to tackle any given problem. That prediction could then be tested by the subjects’ protocols.
Despite this current limitation of the model theory, we reiterate our view that the inference rule theory is too limited in proposing a central reasoning mechanism based on abstract natural deduction rules, which maps so poorly on to the kind of reasoning that most of us do most of the time. The theory is also poorly suited to an integrated approach to reasoning and decision making because it can have little to say about the latter, or even about valid inference from uncertain premises (Stevenson & Over, 1995). In this respect we see much greater potential for the mental models approach. Johnson-Laird (1994a, b) has at least started to extend his theory, and in a seemingly natural way, to inductive and probabilistic reasoning. This should allow him to define semantic information or epistemic utility in some way, and beyond that to specify some common mechanism underlying both deductive reasoning and general decision making. In our view, this mechanism is one for constructing and manipulating mental models of alternative possibilities, and of their plausibility and desirability.
Johnson-Laird (1994a, b) has suggested that the probability of a proposition is assessed by the proportion of mental models in which it is true. To us the idea that people often generate, enumerate, or match many possible models is implausible and stands in contradiction to the proposal that deductive reasoning is difficult when only two or three alternative models need to be considered (Johnson-Laird & Bara, 1984). Moreover, Johnson-Laird appears to presuppose that mental models are, in effect, thought of as all equally probable—i.e. any mental model counts as only one item in any proportion. People are good at learning the relative frequencies of events in the world (Chapter 2), but they have to come to think that some of their mental models are more probable than others if they are to learn which hypotheses are more probable than others from experience (Carnap, 1950; Stevenson & Over, 1995). Even when people do judge probability by the proportion of propositions in mental models or the relative frequency of events in the world, that probability would still have to be mentally represented in some other way in order to affect decision making.
We think that some decision making does take place through the explicit consideration of alternative possibilities, but we do not imagine that people make extensive use of numerals to express probabilities and utilities in their mental models. Any explicit reasoning or decision making requires the help of the implicit system, and both probabilities and pay-offs can be embodied in this system. How else could one explain the massive Skinnerian research programme with its vast accumulated evidence for the effect of alternative reinforcement schedules on animal learning? Johnson-Laird and Byrne (1993) themselves allow mental models to be “annotated” with epistemic qualifiers, and this might be enough to reflect the vague thoughts people often have of some states of affairs as more probable than others. One possibility is to treat “is more probable than” as a relation between models in the way that “is to the right of” is a relation between terms within mental models in Johnson-Laird and Byrne (1991). For decision making, one would also need a relation “is of greater value than” between mental models (Manktelow & Over, 1992).
There is still a long way to go to get a full theory of deductive competence that is integrated with one of decision making. We favour the model approach, but we want to emphasise that we do not regard the manipulation of mental models as the sole mechanism of reasoning that people employ Where problems of a similar type are repeatedly attempted and solved, it is highly efficient for people to learn specific heuristics. Neural networks are probably the mechanism that underlies much of this learning, so that the system is behaving in an “as if” manner. In the terms we have used, the system would only comply with or conform to rules, giving the same output as these would if they really were followed by someone. It is nevertheless possible to learn to formulate a rule explicitly, like a version of the law of large numbers, and then to follow it in an explicit process, as the work of Nisbett and colleagues, discussed earlier, suggests. However, discussion of such rules, and the schemas that are sometimes used to try and express them, is beyond the scope of this book.
Conclusions and Final Thoughts
Much can be learned about human thought by studying errors and biases, relative to impersonal normative systems, in the same way as much can be learned about human vision by studying visual illusions. It is generally misleading, however, to impute irrationality to our subjects on the basis of such errors—and their associated cognitive illusions—in laboratory tasks. Our distinction between rationality1 and rationality2 is made to clear up confusions and to assist with the demanding theoretical task of understanding how people think. We have argued that people have a high degree of rationality1 enabling them to achieve many personal goals, but that such rationality is achieved primarily by use of the tacit cognitive system based upon biologically constrained learning. We have also argued that people have a more limited capacity to be rational2 dependent on the explicit thinking system with its access to natural language and reflective consciousness.
Considering future research on the psychology of thinking, we hope that a general effort will be made to achieve theoretical integration, and see promising signs that this is beginning to happen. The emphasis on rationality2 has, in our view, had a divisive influence—sustaining, for example, the unfortunate separation of the study of reasoning and decision processes. In experiments on explicit deductive reasoning tasks, the focus on this notion of rationality has also led to undue emphasis on the problem of deductive competence and the somewhat sterile rules versus models debate. We feel that researchers should first take account of the very powerful tacit processing systems that affect our judgements and inferences. We would like to see a major expansion in the programme of research here on pragmatic processes, relevance, probability, utility, and information. At the same time, attention must be paid to the ways in which the explicit cognitive system can enhance reasoning and decision making and confer the uniquely human form of intelligence. Above all, we need to raise our vision above that of understanding behaviour within particular paradigms and develop a truly general and integrated theory of human thinking.
1. |
We assume here the computational theory of mind within the cognitive science approach. We realise that some philosophers would dispute the validity of our robot analogy, but it is particularly hard to reject with respect to implicit processes. Indeed we would argue that these philosophers concentrate too much on explicit, conscious process, and wrongly assume that these are more important in thinking than implicit processes. |