[2] In the present paper I shall attempt to further discussion by offering accounts of judgment and reason that are better developed and better supported than my accounts as of 1988.{1} I will correct some of the views developed in that book and offer some arguments not found there. Healy and Reiner are, of course, not to be faulted for presenting and criticizing my published views. In the next section I shall discuss the concept of judgment, which plays a central role in my account of rationality; in the following section I will consider those features that must be added to judgment in order to arrive at an account of reason.
[4] While much of the history of epistemology has been concerned with the regress problem just described, there is a related regress that has received less attention. When we examine standard foundational epistemologies we find that the justification of non-foundational propositions requires that these proposition be related to the foundational propositions by appropriate rules. Whether these are rules of deduction, induction, or some other sort, the choice of rules is just as important for justification as is the choice of foundational propositions. A purported deductive justification of a non-foundational proposition that is invalid or a proposed inductive justification that violates norms of inductive logic would not be a successful justification. Now a naturalistic account of cognition has no more room for a magical insight into the appropriate rules than into the foundations of knowledge. Our choice of rules is also subject to challenge and in need of justification. Even in cases in which a rule maps onto a tautology, our ability to recognize that a proposition is a tautology is not infallible. Moreover, reflection on the history of logic and on contemporary disputes over such subjects as intuitionist, relevance, and paraconsistent logics will underline the point that our ability to grasp even the basics of logic is far from infallible. As a result, we face another regress when we attempt to justify the rules we adopt for producing justifications.
[5] Now in spite of both of these regress problems, we do in fact offer justifications and accept rules of deductive logic and other rules as well. This shows that we terminate the regresses, but it does not show that we do so in a rational manner. It is possible that every decision that ends a regress is ultimately unjustified in that it rests on arbitrary personal choices, or groundless social conventions, or the like. But there is another alternative: Perhaps our cognitive repertoire includes fallible but non-arbitrary means of stopping a justificatory regress. It is this ability that I have tried to capture in the concept of judgment. I will develop this concept in some detail here and argue that this concept is instantiated in human cognition in a way that is compatible with naturalism. I want to underline three features of judgment at the outset in the hope that this will forestall a number of misunderstandings as we proceed. First, since this ability is fallible, justifications that rest on judgments can be challenged and reexamined. Second, exercising judgment is not equivalent to having an opinion. Rather, I will argue that individuals acquire the ability to exercise judgment in a specific field by developing expertise in that field. In other words, there are norms governing what counts as the exercise of judgment, but these will be norms that apply to the agent, not to the proposition or rule being assessed.{2} Third, the claim that human beings are able to exercise judgment is an empirical claim. I will have more to say about the way this claim is to be evaluated as we proceed.
[6] There is a further source of potential misunderstanding that I also want to address at the outset. To claim that our knowledge of even the most elementary principles of logic is fallible is not to claim that we should treat these principles as if they are mistaken. Nor does the claim that our most basic cognitive abilities are fallible imply that they never yield correct results. Since truth is a property of propositions and validity a property of arguments, even a totally arbitrary choice may latch onto a true proposition or a valid argument. The issue that I am concerned with is how we assess propositions and arguments. Moreover, to hold that a proposition or an argument is unjustified as long as it is possible for us to be in error about it is to accept the Cartesian assumption that nothing less than indubitability counts as a genuine justification. The question before us here is whether we have the ability to provide genuine, although fallible, justifications.
[7] I also want to comment a bit further on the agent-centered focus of my account of judgment because this may be an issue on which Reiner and I are speaking past each other. Some of Reiner's remarks leave me with the impression that he is concerned with the product of judgment while I am concerned with the process. For example, Reiner maintains that we need only require that judgments conform to rules but that what is important is the choice of rules to which we require conformity. This, however, tells us nothing about how we assess rules and decide which ones we should accept. In paragraph 9 of his paper Reiner asks: "Is the tacit belief underlying the skills that constitute judgement supposed to be true, i.e. is it supposed to be knowledge?" But, again, truth is a property of propositions, not a property of the cognitive processes by which we assess whether a proposition should be accepted as true, or well-confirmed, or sufficiently close to the truth for a given range of actions, and so forth. Judgment, as I am going to discuss this notion, refers primarily to the process by which we make such assessments, and only in a derivative sense to the outcome of that process. This is particularly important because of the central role that judgment will play in my theory of reason. I take it as crucial that an adequate theory of reason must allow for cases in which false propositions are rationally accepted as true and true propositions are rationally accepted as false in a specific situation -- e.g., when ancient astronomers believed that the earth is stationary and explicitly rejected Aristarchus' proposal that the earth moves annually around the sun. In section 5.5 of Rationality I discussed two cases in which contemporaries disagreed on fundamental issues without either side being irrational. Moreover, since naturalism requires fallibilism, a naturalistic account of reason must allow for situations in which people are rationally justified in holding beliefs which are in fact mistaken, and must include an account of the means by which we correct errors. Thus if Reiner's claim that my approach supports dogmatism were correct, I would consider this to be a serious failing. I think it will be clear that no such dogmatism follows from the account of reason that I will develop here.{3}
[8] I turn now to a more detailed account of the nature of judgment and the reasons for believing that humans are capable of exercising judgment. As Healy points out, I conceive of the ability to exercise judgment as a skill. I will work my way towards an account of judgment by first examining some features of physical skills. Paradigm examples of physical skills include the ability to ride a bicycle, catch and hit a baseball, or use a carpenter's tools. There are three aspects of such skills that I want to emphasize.
[9] First, these skills are learned through demonstration and practice, rather than by learning and following a set of rules.{4} Ability will typically improve with practice and deteriorate with lack of practice -- although relearning will tend to be faster than initial learning. That we do not learn a skill just by memorizing a set of rules is particularly clear in cases in which no one is able to formulate a set of rules that will successfully guide this activity. But even in cases in which we can formulate such rules, being able to state the rules will not usually be sufficient for carrying out the activity in a question. For example, a physicist who understands the mechanics of keeping a bicycle balanced, but who has never ridden, will still have to practice in order to learn to ride. Moreover, children typically learn to ride bicycles without being told to follow any set of rules. The same points hold for developing skill at sports, at driving a car, or at using tools. Note also that even though we may be able to formulate rules that a skill conforms to and embody those rules in a robot that is faster and more accurate than we are, the robot has not acquired the skill in the same way that we do. The existence of such robots does not provide evidence against the claim that we learn skills without following explicit rules.
[10] Second, for most physical skills, and especially for the more complex skills, there is a substantial range in which most people can learn and improve. Still, some people learn much more quickly than others; some develop a skill to a higher degree than others; and some are unable to learn a particular skill at all. For example, most people can develop some skill at catching fly balls, but few achieve the level of a major-league ball player, and a few will be unable to develop this skill at all.
[11] Third, the ability to exercise a physical skill is fallible in the sense that even the best practitioner will not always exercise that skill to the highest degree that is within her capability and will sometimes fail altogether. Yet an enormous variety of human accomplishments depends on physical skills. Note especially that the fact that even the most adept sometimes fail does not provide grounds for concluding that there is no such thing as a skill, nor any reason for refusing to invest greater confidence in those who have shown facility at a particular task than in those who have not.
[12] Now let us consider cognitive abilities. Many cognitive tasks are carried out by following rules. Algorithms are the most impressive class of rules since (by definition) algorithms guarantee completion of a task in a finite number of steps. Familiar examples include the standard arithmetic operations, the rules for differentiating functions, and the truth-table method for evaluating propositional arguments. The discovery of an algorithm for a task for which no previous algorithm was known is an important form of cognitive progress. Other tasks are facilitated by rules that do not guarantee a solution. The tree method for evaluating arguments in predicate logic provides an example of this case. But there are also many cognitive tasks that human beings successfully carry out without following any set of rules. In many cases this occurs because there are no known rules or because the application of known rules exceeds our resources, including available time. Paradigm cases include the ability to solve deduction problems in logic, calculate integrals, and invent previously unknown algorithms. Moreover, when using a non-algorithmic method we must often decide whether to continue to apply the method or to try something different. For example, when using the tree method for a problem in predicate logic, we often reach a point where we must decide whether to continue seeking ways of generating instances that may close all branches of the tree, or look instead for an interpretation that will show the argument to be invalid. While there are no precise rules for making this decision, it does not follow that the decision must be arbitrary. When we teach logic we provide progressively more difficult exercises with the aim of bringing the student to a point at which such decisions are not made at random, even though we never provide a rule for making these decisions.
[13] The ability to make the kinds of decision that I have just been illustrating share the characteristics of physical skills noted above: they are taught by demonstration and learned through practice; there is a range of talents in developing and exercising these abilities; and the exercise of these abilities is fallible. The similarities are sufficiently great that I will refer to these as 'cognitive skills'. Indeed, from a naturalistic perspective we should suspect that cognitive skills are not just analogous to physical skills, but continuous with them. This suggestion is supported by cases in which physical and cognitive skills are thoroughly integrated. Clear examples of this situation are provided by many of the arts such as conceiving and producing a piece of sculpture and doing jazz improvisations. However, in this paper I will concentrate on cognitive skills.
[14] I want to focus now on one aspect of the exercise of cognitive skills. In carrying out a task such as developing a deductive proof or writing a computer program we find ourselves "thinking of" possible steps that we might take, and assessing whether these steps are appropriate. Typically we do not generate these possibilities by applying rules, and in many cases our assessment of whether an idea that has occurred to us is worth pursuing is made without benefit of rules. People who have studied and practiced in a particular field will usually be more effective at coming up with viable ideas than those who have not engaged in such study and practice. This is why we bother training people in particular fields and why we rely on trained people when we want to have particular jobs done or problems solved. Moreover, even though in many cases we develop algorithms or design dedicated computers that will carry out specific tasks more rapidly and accurately than unaided humans can, the fact remains that human beings do develop the ability to carry out tasks and solve problems without having to rely on an algorithm. This point holds for a large number of fields of human endeavor. People develop skills and train others in fields that include medical diagnosis, the law, engineering, theoretical and experimental physics, and many others. Note also that the fact that an algorithm for solving a problem is known does not automatically eliminate the need for non-algorithmic decisions. In addition to decisions whether to use a known algorithm, there are also cases in which we must choose between alternative algorithms that do not yield the same results. For example, Galileo and Newton provided us with different algorithms for calculating how long it will take a body to fall. Which one we use in a particular case will depend on such factors as the exact purpose at hand, the degree of precision that we are seeking, and the degree of precision that we are capable of achieving.{5} The degree of precision that we can achieve changes as science and technology develop. We can now show experimentally that Galileo's law is incorrect for falls of just a few feet (e.g., using Atwood's machine), but this might not matter for many practical purposes. Thus the decision to use one of these algorithms rather than the other is conditioned by a host of contextual factors.
[15] I want to emphasize that my account of skills and skill development should not be assimilated to that of Dreyfus and Dreyfus (1986, 1991). While there is much of interest in their discussion, their account is limited to the process by which a learner develops an already existing skill. Nothing in my account precludes an individual from developing levels of a skill that go beyond the accomplishments of those who trained her. Nor does my account deny the existence of geniuses who create entirely new fields or who enter into existing fields without the need of extensive training. It would appear that no one trained Socrates in dialectic and it is clear that no one taught Galileo how to use a telescope or trained Ramanujan in mathematics.
[16] My general point, then, is that there are many fields in which people develop cognitive skills and a great deal of everyday life, as well as intellectual and technical progress, depends on our ability to develop such skills. Now each of the decisions that rely on a cognitive skill provides an example of the exercise of judgment. Note especially that as I am using the term, merely having an opinion is not the same as exercising judgment. One has to develop expertise in a field in order to exercise judgment in that field, and one must know the details of a particular problem in order to exercise judgment with respect to that problem. The ability to exercise judgment in a specific field is an individual accomplishment, but is no more "subjective" than is the ability to drive a car or catch a baseball. This suggests that Healy's use of the term 'judgment' is not quite the same as mine. Given my usage, his talk of "informed judgment" is potentially misleading since it suggests the possibility of an uninformed judgment. As I am using the term, some judgments on a particular matter may be better informed than others, but an uninformed belief is not the result of an exercise of judgment. In other words, as I am using the term, 'judgment' has both a descriptive and a normative dimension. It is not always easy to determine who is competent to exercise judgment in a particular field, but determining how to recognize expertise is a different issue than the one that I am concerned with here. My present concern is just to make the point that people do develop the ability to exercise judgment.{6}
[17] I have been emphasizing situations in which we make decisions without a sufficient set of rules to dictate that decision. Healy argues that the scope of judgment goes well beyond cases in which we run out of rules, that "the exercise of judgment transcends and underpins rule-following at every step of the way," and thus that "Brown vastly underestimates the centrality of judgment to the reasoning process" (paragraph 4). He offers several examples in support of his claim, e.g., deciding which rules are appropriate in a particular case and assessing whether rules have been followed correctly. In fact, I agree with Healy's examples of cases in which we exercise judgment, but I note that all of his examples do have this feature in common: they involve cases in which we must make decisions without having a rule to dictate the outcome. There is, however, another point of disagreement that may be more substantive.
[18] In my view, we do not always exercise judgment in familiar situations and to do so would be wasteful. The exercise of judgment is a cognitive process which requires time and perhaps other resources as well. We are finite beings and it is important for us to apportion our resources. The ability to carry out such everyday tasks as starting my car or washing the dishes without engaging in reflection on how to proceed allows me to think about other matters while also getting these tasks done. It would be extremely wasteful to reassess whether the rules of arithmetic are appropriate every time I set out to balance my checkbook. Moreover, the exercise of judgment requires expertise and none of us have the skills required for determining the proper way to carry out every mundane task. There are two general ways in we which we avoid getting into this kind of impossible situation. In some cases we engage in higher-order reflection and decide to adopt specific procedures in familiar situations. In other cases (this is probably the more common of the two) we make use of procedures that have been adopted in our society and that we have learned as a result of living in that society. Hopefully these procedures were themselves adopted on a rational basis, but in an enormous number of cases we are unable to make this assessment for ourselves. We typically shift into a reflective mode when familiar procedures fail. If my car does not start when I carry out the usual routine I may attempt to assess the reasons for this failure. Alternatively, given an awareness of my ignorance on such matters, perhaps reinforced by previous futile attempts to solve this type of problem, I may have adopted the practice of calling the garage without further reflection. There is nothing irrational about the development of such habits. Thus, while I agree with Healy that "judgment always underpins the reasoning process, even when we are employing fully algorithmic procedures (as when, say, balancing a checkbook or differentiating a function)" (paragraph 4, see also paragraph 16), this need not be my judgment at each time that I am using an algorithm.{7}
[19] Consider, now, the role that judgment plays in blocking justificatory regresses. The regress problem arises when we begin back-tracking as we seek a more certain justification for a proposition. As long as we continue the backwards search, we have not yet achieved a justification. The problem is to stop the regress in some non-arbitrary manner. My thesis is that the exercise of judgment provides a way of stopping a regress that is fallible and open to reconsideration, but not arbitrary. When someone with the relevant expertise proposes, say, a diagnosis or a design approach or a plan of proof, we have a reason for taking the proposal seriously even though we do not have a proof that the proposal is correct or the best of the available options. This does not mean that it is rational to accept this proposal without any further evaluation (remember, the exercise of judgment will be only a part of my full account of rational belief), but it does mean that we have a reason for expending resources on the further evaluation of this proposal that we would not have if the same proposal were picked out of a hat or offered by someone who lacked expertise in this field.
[20] We are now ready to turn to a crucial question: How are we to account for the phenomena I have been describing from a naturalistic perspective? I have described situations in which people arrive at decisions without consciously following rules; we would like to have an account of the underlying cognitive abilities that make such behavior possible. In other words, we are looking for a theory and an adequate theory should exhibit those virtues that we normally expect from a good theory. These include an ability to account for the evidence and coherence with other available relevant knowledge. In addition, where there are competing theories, we will want to adopt the one that meets these demands most successfully. I now want to examine two such theoretical proposals.
[21] The first proposal is that although we do not consciously follow rules when we exercise judgment, we must be engaged in some kind of unconscious rule-following. This view is, I think, currently the dominant view and it is important to keep in mind that this is a theoretical hypothesis, not an a priori truth. It will be helpful in assessing the strengths and weakness of this proposal if we first examine the phenomenon of rule-following a bit more closely.
[22] The paradigm examples of rule-following are provided by familiar situations in which we carry out a task by consciously following a set of rules. For example, suppose a friend has moved to an unfamiliar town and I am making my first visit since that move. I am driving to my friend's new home, following an explicit set of directions that she has supplied. Or, suppose I am calculating the derivative of a complex but straightforward function, explicitly following the standard rules for exponentials, products, and so on. An important variation involves so-called "rules of thumb," that is, rules that can be stated and consciously followed, and that often yield a successful outcome -- but not always. Examples include: when lost while driving in an unfamiliar town, stop at a gas station and ask for directions; when attempting to integrate a function, try a trigonometric substitution; when stuck on a proof in symbolic logic try an indirect proof, or try making obviously correct deductions and see if anything interesting turns up. The important feature of these examples for present purposes is that the practice of following the rules seems to explain why we are able to proceed in a coherent, non-random manner, and why we are able to achieve the successes that we do.{8}
[23] Now let us turn our attention to cases in which individuals are reasonably successful at carrying out a particular kind of task, but do so without consciously following rules. In order to explain this behavior we need to construct a theory, and one approach to theory-construction that has often proved successful is to attempt to model a theory on some familiar situation. In the case at hand, we may attempt to account for skillful behavior by postulating the existence of non-conscious behavior that is closely analogous to the conscious rule-following behavior with which we are familiar. Two features make this an especially attractive theory: first, our familiarity with conscious rule-following gives us a sense that we understand how people exercise cognitive skills and why they are successful; second, there has been a lack of serious competitors. Thus I will respond to this theory in two steps: first I will consider a number of problems with the theory; second, I will propose an alternative theory and argue that is it superior.
[24] The first problem for the theory under examination is that unconscious rule-following requires an entity that carries out these activities. Two candidates seem plausible: an unconscious mind that is modelled on the conscious mind or the brain itself; each of these proposals faces serious difficulties. The unconscious mind approach involves postulating an additional entity whose ontological status is at least as unclear as that of the conscious mind. Moreover, all of the traditional problems about the relation between the mind and the brain will reappear with respect to the unconscious mind and the brain, and we will also have to face another (more recent but now familiar) set of problems about the relation between the unconscious and the conscious mind. Moving directly to the brain is a much more attractive proposal given the history of failures to solve these problems.
[25] But now another difficulty appears. The brain is a physical system and we do not normally attribute the ability to follow rules to physical systems. This may be a mistake. There may be physical systems of a particular degree of complexity or of specific kinds that do follow rules, and the brain may be one of these. But a reasonable defense of this claim requires some non-question-begging evidence that such systems do exist. It may be tempting to offer the computer as an example of a physical system that follows rules, but to do so would be question-begging. Although a computer program is a set of rules, when we enter a program into the computer we change states of a number of circuit elements and thus create an initial condition in the machine. The situation is analogous to one in which I might "program" a rock to hit the floor with a velocity of (roughly) sixteen feet per second by lifting it to a height of four feet and then dropping it. If the computer really is different from the rock we require evidence that this is the case. In the present context, it cannot simply be assumed that the rock does not follow rules but that the computer does.
[26] We are left, then, with our first problem unsolved: If the exercise of judgment occurs through unconscious rule-following, it is not clear what entity follows these rules.
[27] Second, if we postulate unconscious rule following we must face the regress problem once again. Merely locating rules that are being followed unconsciously will not show that these are appropriate rules. For example, we might argue that the regress ends because we reach a set of rules that are hard-wired into us, and that this is as far as we can go. If this is the case, we may be in a situation in which we are permanently doomed to following inappropriate rules. It may, of course, turn out that this is true, but I shall argue shortly that human behavior has shown sufficient flexibility to make this view implausible -- especially in the presence of an alternative account that is more compatible with that flexibility. Before coming to that issue, however, let us consider one way of arguing that we might operate under a set of fixed rules that are in fact appropriate. The approach would be to give an evolutionary account of why this particular set of rules developed. On such an account we would be dealing with rules that developed in response to environmental pressures, not some set of rules that was arbitrarily implanted in us. But while this may be a tempting approach, I suggest that tying evolution to the development of rules will not do the job because the theory of evolution should not lead us to expect anything more than rules that are appropriate for the particular circumstances of our evolutionary history. Yet human behavior shows a wide ranging flexibility and adaptability that includes the ability to pursue knowledge way beyond the environments in which we evolved. Striking examples include sending a space ship to orbit Venus and use radar to map the surface below the planet's cloud cover, and exploring distant parts of the universe by means of radio, X-ray, gamma ray and neutrino telescopes. With the advent of moon landings, the Mir space station, and regular shuttle flights, it is clear that we can learn to function in environments quite far from that in which we evolved. We also have the ability to develop highly abstract systems of mathematics that have no apparent connection with our survival or reproduction. It is not likely that all of these abilities ultimately rest on a hard-wired set of rules that developed in response to the environmental pressures of our early evolutionary history.
[28] Third, consider one aspect of how the process of skill improvement would look on the account under consideration. For example, it is easy to state Newton's laws or to write down the general form of the Schrodinger equation; it is much harder to learn to apply them to specific problems. Learning to carry out such applications is a central part of skill development in physics.{9} Now one way of accounting for skill improvement on the basis of unconscious rule-following is to assume that the more skillful practitioner will have learned a set of rules for applying the general principles to a variety of special circumstances. Yet if the improvement of a skill requires the acquisition of a growing set of ever more specific rules to cover special cases, we would expect performance to deteriorate as the number of rules to be searched and evaluated at each stage increases. This is exactly the opposite of what we find. Yet we do learn to apply highly general rules to a wide variety of specific cases -- including situations having features that we have not encountered before. The suggestion that we do this by following another set of rules that we already have available is most implausible. Nor will it help to say that we just invent new rules for applying rules as the need arises. To take this approach is to concede the main point that concerns me here: We have cognitive abilities that do not operate by following rules.
[29] The overall situation is captured in the following summary of difficulties encountered by artificial intelligence researchers who sought to model human cognition solely on the basis of sets of rules.
While initially the task of writing rule systems capable of accounting for human behavior seemed tractable, intense pursuit of the endeavor raised doubts. Rules systems were hampered by their "brittleness," inflexibility, difficulty learning from experience, inadequate generalization, domain specificity, and inefficiencies due to serial search through large systems. Human cognition, which the rule systems were supposed to be modelling, seemed to be relatively free of such limitations. (Bechtel and Abrahamsen 1991: 17)
[30] Let us turn now to an alternative approach: That the brain is in fact able to carry out coherent evaluations as a result of training and experience, and to do so without following any rules. Now this would be a totally ad hoc proposal were it not for the fact that we do have independent evidence that there are physical systems that are capable of working in just this way. This evidence is provided by the existence of connectionist computers (see, for example, Bechtel and Abrahamsen 1991, Rumelhart and McClelland 1986, Churchland 1989).{10} There are several reasons why these machines are relevant to our present concerns.
[31] First, the physical structure of these machines is considerably closer to that of the brain than the physical design of traditional computers. Instead of having a central processing unit and a distinct memory, connectionist machines consist of a number of distinct processors that are multiply connected to each other and that carry out computations in parallel across the entire machine. The strengths of these connections change as a machine "learns" and information is stored in those connection strengths, rather than in a separate memory. These machines are much better than serial machines at tasks such as pattern recognition -- tasks at which human beings are also very good but which have long resisted efficient implementation on traditional machines. Connectionist machines are capable of learning from experience, and when they suffer localized damage, their behavior deteriorates in ways that are similar to the way human cognitive behavior deteriorates in response to damage. For example, instead of a bit of information simply being lost, as occurs when a memory cell is damaged in a traditional computer, the information becomes somewhat vague, but is not totally gone. (See Hinton, et al. 1993 for a recent example.) To be sure, currently available connectionist machines are not nearly as good as humans at these tasks, but these machines are also much simpler physical systems than a human organism.
[32] Second, these machines learn to carry out tasks without having a program inserted into them. Instead, the machines are trained to do a specific task by giving them an input from a training set and adjusting the connection weights slightly when the machine gives an incorrect output. As the training procedure continues a machine will get progressively better at a task and may sometimes become more efficient at it than the teachers who carried out the training. There is a very real sense in which the way these machines learn to carry out a task is similar to the way that we learn skills. The machine is not given a set of rules to follow and there is no good reason to think that the machine learns by extracting rules from the training set that it later follows. However, even if the machine does learn by extracting rules, that process takes place without the guidance of a set of rules. In other words, the most important feature of these machines for present purposes is that they show that a physical system can learn to carry out tasks and engage in coherent behavior without having to follow some set of rules.
[33] Third, I noted above that one striking cognitive feature of humans is our extreme adaptability. I submit that our capabilities are explained much better by the hypothesis that our brains are extremely flexible systems capable of adapting to a wide range of situations than by assuming that our brains operate in accordance with a set of ground-level rules that were built into us once and for all. Sufficiently complex versions of such machines may well turn out to be capable of adapting to a variety of environments, and even of readapting when a new flow of inputs is encountered.
[34] Finally, there is no reason for thinking that evolution could not have produced such machines. One vital lesson that we have learned from the existence of connectionist machines is that physical systems that operate in this manner are much easier to construct than was imagined a few decades ago. Systems of this sort provide a much clearer naturalistic basis for the existence of skills than does the attempt to reduce behavior that does not appear to involve following rules to behavior that does follow rules at some other level.
[36] Healy describes my view as requiring that "to be termed rational, an individual judgment must be validated by submission for critical appraisal to the relevant expert community" (paragraph 17) and Reiner writes of a quest for social approval (paragraph 11). When I introduced a social element as a necessary condition for rationality I wrote, "for a belief based on judgment to be a rational one, it must be submitted to the community of those who share the relevant expertise for evaluation against their own judgments" (1988: 187). A few pages later, discussing some points on which I disagree with Kuhn, I wrote: "the model of rationality that I am proposing only requires that individuals submit their judgments for evaluation by their peers, and that they take this evaluation seriously. The model does not require that each member of the community agree with the majority, and indeed, agreement with the majority view is neither necessary nor sufficient for rationality" (192). I proceeded to explain why agreement is neither necessary nor sufficient and in Section 5.5 I provided an extended discussion of two cases of fundamental disagreement within a community in which all of the participants were, on my view, rational. Nothing in this discussion yields a social determination of what beliefs are rational. I can best explain the function of social evaluation by explaining why judgment is not sufficient for rationality and considering what else is required.
[37] I take it that our central cognitive aim is the acquisition of true beliefs and that reason is the major instrument by which we pursue this cognitive aim. It is especially important here not to mystify the notion of truth. Depending on context and circumstances, we are interested in arriving at true beliefs on a wide range of topics. Examples include whether it is raining, whether a particular mushroom is poisonous, whether a specific approximate means of carrying out a computation is sufficiently accurate for a given engineering application, whether there are 92 naturally occurring elements in the universe, whether quarks exist, whether all of those who signed the American Declaration of Independence did so on July 4, 1776, and many other more and less dramatic examples. But, as Plato and many of his successors recognized, a true belief is of limited value unless we have reasons for taking the proposition in question to be true. If we have merely stumbled into a true belief, we can just as easily stumble out. Propositions do not wear their truth-values on their sleeves, so some process of evaluation is required to determine whether a proposition is true; reason provides the means by which carry out these evaluations. Yet we are fallible creatures, and our evaluations of proposition will be infected with our own fallibility. One concern of epistemology should be to seek means of controlling for this fallibility and attempting to minimize its effects.
[38] I have argued that judgment plays a key role in all of our cognitive evaluations and that judgment provides reasons for believing a proposition that are more reliable than random beliefs, but that are still highly fallible. Clearly, we should attempt to improve on the reliability of our judgments when this is possible -- not to do so would be less than rational. Moreover, the way in which we should pursue improved judgments becomes clear when we consider that the most important limitations on judgment derive from the fact that the ability to exercise judgment is an individual ability. As a result, judgments will be limited by limitations of the individual who exercises judgment. These include several rather mundane individual limitations: we do not always have a command of all the relevant information and all the relevant techniques, and sometimes we are not even aware that information and techniques we have not considered exist; we do not immediately grasp all the implications of the propositions we consider and even when we proceed to formally derive consequences, we do not always succeed in drawing out the most relevant conclusions; we typically do not recognize all of the alternatives that might be considered; often we are overly fond of our own ideas; and so forth.{11} All of the means of improving our judgments that I am going to discuss are aimed at reducing the impact of these limitations by requiring that we subject judgments to further evaluations using means that take us outside the limits of our own minds and over which we have only limited individual control.
[39] There seem to be three major ways by which we can pursue such improvements: observation, the use of logic and other formal methods, and social evaluation of our judgments. Within the scope of this paper I can only offer some brief illustrations of the role that these play in evaluating judgments. There is, however, one important complication that should be kept in mind throughout this discussion: the use of these three resources overlap in complex ways and each will also involve the exercise of judgment. There is no simple story to be told here and no algorithm for deciding just what should be done at each juncture.
[40] Observation is not relevant to all cognitive tasks but where it is relevant -- as it is in science and in many everyday concerns -- observation provides powerful constraints on our theorizing. This is particularly clear on a realist account of observation, i.e., an account which holds that observation brings us into contact with the items which exist independently of our beliefs about them and, so to speak, allows those items to have their say.{12} I am not maintaining that observational claims are ipso facto more certain than other kinds of claims. In contemporary physical science, for example, the procedures that lead to observation claims often involve complex equipment and data processing techniques, and the acceptance of an observation report requires all of the other factors involved in arriving at rational beliefs.{13} Rather, observation plays a special role in evaluating beliefs because observation brings us into contact with a world that is independent of those beliefs.
[41] Formal methods allow us to draw out consequences of propositions that are often far from obvious. The result of a series of arithmetic operations on a specific set of numbers is already determined before we carry out the arithmetic; doing the arithmetic allows us to discover this result -- which we may find surprising and even unpleasant. In a similar way, whether a set of propositions is consistent is a fact about that set -- a fact about which we may be mistaken, as the classic example of Russell's paradox adequately illustrates. Note, however, that while formal relations are, in an important sense, objective, the role that such results play in evaluating beliefs is far from straightforward. The point comes out in a particularly striking way when we begin with a set of premises that we believe to be true and deduce a conclusion that conflicts with results accepted on the basis of observation. At this point it is clear that something has gone wrong, but many options are available in deciding just what is wrong. We may reconsider one or more of the premises, reconsider the procedure that led to the observational result, recheck the deduction if it is a long and complex one, or any combination of the above. We might even choose to reconsider a previously accepted principle of logic if we are convinced that we are faced with an argument displaying true premises and a false conclusion.{14}
[42] Finally, we come to social evaluation. We should submit our views to evaluation and criticism by others because other people will have perspectives, information and skills that we do not have. This is a major departure from the Cartesian thesis that knowledge is a matter of what individual cognitive agents can accomplish within their own minds, but the history of philosophy and science since Descartes have shown pretty clearly that Descartes was overly optimistic about what an individual mind can accomplish and that he seriously underestimated the complexity of the universe we live in.{15} But perhaps there is a different lesson that we should learn from Descartes's practice -- he did submit the Meditations to seven other philosophers for comments before he published it, and the book was published with these comments and Descartes's replies. Indeed, the value of critical debate is a lesson that most philosophers have long recognized in practice, although they have been reluctant to include it in their official epistemologies.
[43] Testing our judgments by any of these means can lead to a number of possible outcomes: they may provide reasons for rejecting an initial judgment, they may lead to its modification, they may support that judgment, and they may convince us that we should withhold a decision pending further reflection and evaluation. But whatever the outcome of these evaluations, there is no automatic conclusion that determines their impact nor is there any algorithm for deciding when a sufficient body of evaluations have been carried out. Studies in the history and philosophy of science over the past four decades provide myriad illustrations of this point. Theories face anomalies and some of these anomalies are eventually handled by the theory while others lead to minor modifications or major revolutions. Theories that have triumphed for decades are overthrown, and so forth. Such situations arise even in mathematics; Lakatos (1976) is, perhaps, the classic study. Situations of this sort are even more common in everyday life and in the worlds of business, diplomacy and politics.
[44] The upshot, then, is that even after I have subjected my judgments to testing through observation, formal analysis, and interpersonal critical debate, I must rely on my judgment in order to assess what propositions I should believe. The evaluation process results in improved judgments which provide the basis for rational beliefs. Still, my rational beliefs are my beliefs for which I take responsibility. Moreover, our understanding of the fallibility of our judgments, assessments of judgments, and judgments of these assessments, requires that we hold our rational beliefs as fallible and subject to reconsideration when reasons for such reconsideration arise.
[45] Thinking of reason as a tool and truth as the end towards which this tool is directed will help us understand the ways in which reason can also be used in seeking its own improvement.{16} Improvements in our means of observing the world, the development of more powerful formal methods, improved means of storing and accessing information and of engaging in discussion with others, all increase our ability to evaluate and criticize beliefs, and thus increase our ability to pursue truth effectively. Let me mention just one example of each of these kinds of improvements.
[46] First, during the last 50 years astronomical observation has overcome its traditional limitation to the visual realm. Astronomers now regularly tap information throughout the entire electromagnetic spectrum and even off that spectrum (e.g., in neutrino astronomy). These new data sources have significantly changed our image of the universe. Moreover, at the present stage of the development of science, this should come as no surprise. After all, there is no a priori reason why the senses we evolved on the surface of this planet should play a special role in acquiring information about the universe at large.
[47] Second, a contemporary college student who has taken a semester of calculus can easily solve problems that defeated the best minds in the world during the seventeenth century.
[48] Third, as the debate in which this paper is a component will illustrate, new communication techniques have made it possible for people located at great distances from each other who share common interests to exchange criticisms and ideas with a speed that was not even imagined just a few years ago. This increased range and speed of communication can have powerful effects on the ability of an individual with a very limited life span to contribute to a continuing intellectual quest by having errors corrected and new directions for fruitful research pointed out much sooner than was ever possible before.
[49] All of these resources improve our ability to make reliable judgments. Indeed, this occurs even at the level of our initial judgments since improved techniques and information become included in the "training set" that provides the basis for our acquisition of cognitive skills. Moreover, richer and more reliable data provides a superior basis for our initial judgments.
[50] These examples are merely illustrations, but it is important that we not draw the wrong conclusion from them. It does not follow that we are more rational than our predecessors since being rational is a matter of making good use of the resources available to us. Thinking of reason as an instrument aimed at learning truths will help us understand the point: By improving the instrument, we become better able to pursue the goal. The situation is analogous to that in which a craftsperson using relatively primitive tools need not be poorer at her craft than someone who has better tools, even though the latter person will be able to produce a superior product. In a similar way, our ability to pursue truth improves as our ability to make and evaluate judgments improves. At the same time, the availability of a richer body of resources places greater demands on us than on our predecessors since a failure to make reasonable use of these resources will amount to a failure to arrive at our beliefs in a rational manner.
[51] Let us now return to Robinson Crusoe and some related issues that are raised by special situations. A person who is physically isolated and thus unable to engage in intersubjective evaluation is lacking means that might increase the reliability of her judgments. Depending on the degree of isolation, there may be many topics on which she is unable to arrive at rational results because she lacks the necessary expertise and cannot get help. Even in cases where she does have expertise, specific relevant pieces of information that we might normally look up in handbooks may be unavailable. Moreover, if it turns out that this person's knowledge of logic and elementary mathematics are limited, and if her sight and hearing are impaired as well, her ability to arrive at beliefs in a rational manner may be very limited indeed. People who have to make decisions under severe time limitations may also find themselves in a position in which they are unable to make their decisions on a rational basis. And, to takes a somewhat different example, at the present time high-energy physics is in a serious malaise because important theories cannot be rationally evaluated due to a lack of instrumentation for producing the required data. Physicists thus lack a rational basis for evaluating these theories and reason requires that they suspend judgment pending further developments. In other words, while being rational is largely a matter of making good use of available resources, if the resources are sufficiently poor, it may become difficult or impossible to arrive at rational results on some questions. The only rational response becomes one of suspending judgment and if real-world pressures force us to come to a decision, we may have to do so on a less-than-rational basis.
Bechtel, William and Adele Abrahamsen (1991) Connectionism and the Mind: An Introduction to Parallel Processing in Networks. Cambridge, MA: Basil Blackwell.
Brown, Harold I. (1987) Observation and Objectivity. New York: Oxford University Press.
Brown, Harold I. (1988) Rationality. London: Routledge.
Brown, Harold I. (1990) "Prospective Realism." Studies in History and Philosophy of Science 21, 211-42.
Brown, Harold I. and Hooker, C. A. (in preparation) Beyond Formalism: A Naturalistic Theory of Reason.
Cherniak, Christopher (1986) Minimal Rationality. Cambridge, MA: MIT Press.
Churchland, Paul M. (1989) The Neurocomputational Perspective: Philosophical Essays on the Mind, Brain, and Science. Cambridge, MA: MIT Press.
Dreyfus, Hubert L. and Dreyfus, Stuart E. (1986) Mind Over Machine. New York: The Free Press.
Dreyfus, Hubert L. and Dreyfus, Stuart E. (1991) "Towards a Phenomenology of Ethical Expertise." Human Studies 14, 229-250.
Galison, Peter (1987) How Experiments End. Chicago: University of Chicago Press.
Hinton, Geoffrey E., David C. Plaut and Tim Shallice (1993) "Simulating Brain Damage." Scientific American 269, No. 4, 76-82.
Klotz, Irving M. (1980) "The N-Ray Affair." Scientific American 242, No.5, 168-175.
Lakatos, I. (1976) Proofs and Refutations. Cambridge: Cambridge University Press.
Rumelhart, D., J. McClelland, et al. (1986) Parallel Distributed Processing, 2 vols. Cambridge, MA: MIT Press.
Shapere, Dudley (1982) "The Concept of Observation in Science and Philosophy." Philosophy of Science 49, 485-525.
{2} For brevity I will usually write only of accepting propositions with the understanding that accepting a rule is always equivalent to accepting a specific proposition. For example, accepting the validity of modus ponens is equivalent to accepting the proposition that it is always permissible to make the inference embodied in this rule. Return
{3} I am not clear why Reiner insists on the familiar point that any finite body of behavior can be covered by (many) rules. Of course I agree, but this universal point tells us nothing about how we should assess which of these rules are appropriate for specific purposes. Return
{4} At the present stage of the discussion I am only interested in whether we consciously make use of rules. I will much to say about possibility that we follow rules unconsciously beginning at paragraph 21. Return
{5} According to relativity theory, both of these algorithms yield results that are, at best, only approximately correct. Return
{6} Deciding whether someone does have an expertise that she claims to have, or whether a purported expertise exists at all, are difficult practical problems for which there is no easy solution. Deceptions, including self-deceptions, are not uncommon, but it is not impossible to ferret these out. Wood's demonstration that Blondlot and his colleagues were deceiving themselves when they thought that they could detect effects of N-rays provides a nice example (see Klotz 1980). One way in which we attempt to deal with this problem is through the creation of social structures such as the licensing of airline pilots and physicians. These procedures are extremely helpful to the layperson and ameliorate the problem to a significant degree, even though they are far from infallible. Further discussion of this issue here would take us too far afield. Return
{7} It might be suggested that judgment is still required to recognize that a set routine is relevant in a given situation, but I think this is a mistake for two reasons. First, I want to resist any temptation to label every cognitive decision a "judgment." To do so would seriously reduce the content of concept of judgment. The claim that human beings can learn to exercise judgment is intended as a substantive claim, not a tautology. Second, I suggest (although I will not argue the point here) that many concepts and descriptions have a body of core instances which are recognized without any need to engage in the particular cognitive processes involved in judgment. If we may assume for the sake of illustration that Fodor's modularity thesis is correct, then I suggest that the judgment module will not be involved in such cases. Return
{8} In spite of the familiarity of these situations there are eliminativists who maintain that consciousness and related notions are theoretical concepts from an outmoded theory that should be abandoned along with such concepts as phlogiston and N-rays. If these philosophers are correct, our descriptions of these situations will have to be revised. Such a result would be a serious blow to those who attempt to account for cognition in terms of following rules because it would undermine their paradigm cases. Since the account that I will defend holds that intellectual skills are not dependent on rule-following, nothing in my own theory depends on whether this description of rule-following picks out a genuine feature of cognitive life. Return
{9} This is just one of the many forms in which the problem of applying rules arises. Return
{10} These machines are also referred to as "neural nets" and as "parallel distributed processing machines. Return
{11} Many of these points are discussed in Cherniak (1986). Return
{12} I have defended this view of observation in Brown (1987 and 1990). Return
{13} This thesis has been developed and defended in considerable detail in Galison (1987). Return
{14} Some advocates of quantum logics make this move; see Brown (1988: 76-77) for discussion and references. See Brown (1988: 92-93) for an example of a case in which the deduction of an inconsistency from a pair of premises resulted in the rejection of an implicit premises that had not previously been considered to be a revisable hypothesis. See also Bahcall (1989), Brown (1987: 66-70) and Shapere (1982) for discussions of the solar neutrino experiment -- an empirical result that indicates that something is wrong somewhere in the body of contemporary physics, but that has resisted a definitive resolution for more than 25 years. Return
{15} Recall that Descartes thought that he could eliminate the demon and lay the foundations of metaphysics and science once and for all in a rather short book. Note also that Descartes's radical critique of perception and Locke's distinction between primary and secondary qualities were aimed at showing that some of the items which appear in our perceptual experience have no archetype in the physical world. The recognition that the physical world is full of items that we cannot detect with our unaided senses but that we still can study is a later development. Return
{16} The following discussion of the improvement of reason was developed jointly with C. A. Hooker. I cannot disentangle his contributions from my own. Return