1. Introduction

Few products of intellectual life are more exhilarating, more pleasing to give and receive, than a good explanation. While theories of explanation can be quite technical, and the content of particular scientific explanations quite arcane, the cue for acceptable explanation remains cheerfully informal. A good explanation “feels right.” In all cases, the cue seems the same: A familiar sense of understanding that the explanatory story, causal or otherwise, delivers to us.Footnote ¹ It is not news, of course, that this sense of understanding alone is not necessarily a reliable guide to truth, nor is it a necessary condition for good explanation. Still less is it sufficient for good explanation. As one philosopher of explanation puts this observation: “It is no explanation to provide a distorted representation of the world, and the ‘understanding’ induced by such incorrect models is illusory at best” (Reference HumphreysHumphreys 1989, 103).

What causes us to accept an explanation? Is it the sense of a coherent story it conveys? Undeniably, there is a special kind of intellectual satisfaction—an affective component—that occasions the acceptance of an explanation, a sense that we have achieved understanding of the phenomena. Peirce identifies the distinctive cognitive experience of explanatory understanding by isolating the moment of final acceptance; the good explanation “is turned back and forth like a key in a lock” (Reference Peirce1908, 100). This description alone should supply little solace to those holding that good explanations are epistemically reliable. After all, alchemists surely felt the key turn, but once inside we find only false descriptions of causal mechanisms. And when Galen arrived at a diagnosis of melancholy due to black bile, his sense of understanding was so gratifying it must have balanced his humors. Ptolemy claimed that “if the earth did not lie in the middle [of the universe], the whole order of things which we observe in the increase and decrease in the length of daylight would be fundamentally upset” (Reference Ptolemy and Toomer1984, 42/H20).

Some of our favorite stories in the history of science, such as Kekule's famous “Eureka” episode, feature dramatic journeys to compelling explanations.Footnote ² In these scenarios, the explanation “felt right.” This sense of understanding is causally potent. It can be so comforting that explanation stops when this feeling is experienced. I will argue that the psychological sense of understanding is just a kind of confidence, abetted by hindsight, of intellectual satisfaction that a question has been adequately answered.Footnote ³ Thus this sense of satisfaction is confidence that one enjoys an accurate description of the underlying causal factors sufficient (under the circumstances) to bring about the phenomenon we are examining. But confidence is, notoriously, not an indicator of truth.

The track record of this sense of understanding is not entirely treacherous, however. This sense sometimes seems to be associated with very secure explanations as well. Darwin claimed that the explanation of the distribution and origin of species in terms of natural section was “so satisfactory” (Reference DarwinDarwin 1859, 476). Avogadro drew a similar inference to the best explanation for the existence of molecules. Diverse domains and measurement procedures led to the unification of molecular and atomic phenomena. Perrin commented on the confidence with which the molecular hypothesis could be held, and did so in distinctly psychological terms:

This subjective sense of understanding may be conveyed by a psychological impression that the explanatory mechanisms are transparent and coherent, or that the explanation seems plausible, and so should be confidently accepted. At the moment, there is neither a satisfying formal account of explanation, nor agreement about the important informal criteria for good explanation, producing what one review casts as “an embarrassment for the philosophy of science” (Reference Newton-Smith and Newton-SmithNewton-Smith, 2000, 132).

Current theories of explanation may leave us embarrassed, but we should be at least as embarrassed by our “feels-right” diagnostic standard for the acceptance of an explanation. Sadly, it has been the fate of humans to have more counterfeit than genuine experiences of “understanding”. Upon reflection, we are aware of this unreliability. What, then, beckons us to make such judgments? There is, in fact, a very substantial literature in cognitive psychology establishing an overconfidence effect for scientists and laypeople alike (see Reference Fischhoff Baruch and LichtensteinFischhoff, Slovic, and Lichtenstein 1977), an effect that conveys the mere “feeling of knowing” (see Reference KoriatKoriat 1993, Reference Koriat1995). This overconfidence is not the mere result of individual differences in personality or of clinical delusions of grandeur; it is the normal consequence of routine cognitive activity.

What conception of understanding is important to theories of explanation? The traditional account casts understanding as the result of a process that is fundamentally reconstructive: understanding requires that the individual be able to piece together bits of information in their cognitive possession. Reconstructive accounts of explanation agree on one point: Understanding is centrally involved in explanation, whether as an intellectual goal or as a means of unifying practice. Peter Achinstein asserts a “fundamental relationship between explanation and understanding” (Reference Achinstein1983, 16). David Lewis requires of an explanation that “the recipient understands and believes what he is told” (Reference Lewis and Ruben1993, 185). Wesley Salmon proposes that scientific understanding is achieved in two ways: by “fitting phenomena into a comprehensive scientific world-picture” (Reference Salmon and Salmon1998, 77), and by detailing and thereby exposing the “inner mechanisms” of a process (Reference Salmon and Salmon1998, 77). Michael Friedman claims that the relation of phenomena that “gives understanding of the explained phenomenon” is “the central problem of scientific explanation” (Reference Friedman and Pitt1988, 189). Peter Railton associates the scientific understanding of quantum mechanics with “the ability to fill out arbitrarily bits of ideal explanatory text” (Reference Railton and RubenRailton 1993, 170–171), which produces a mass of theoretical detail about the object of explanation. Philip Kitcher relates understanding and explanation so closely that elucidation of this connection in a theory of explanation “should show us how scientific explanation advances our understanding” (Reference Kitcher and Pitt1988, 168). James Woodward claims that a theory of explanation should “identify the structural features of such explanation which function so as to produce understanding in the ordinary user” (Reference Woodward and Ruben1993, 249).

While these declarations associate explanation with understanding, none of these accounts have much to say about the precise nature of understanding, and certainly none of them entail that a sense of understanding is criterial for good explanation. But most of them do draw upon our ordinary, ungarnished notion of understanding, in terms of “detailing” and “identifying” the inner mechanisms and structural features of interest. And this contingent fact is a significant feature in contemporary accounts of explanation. Terms like “detailing”, “identifying” and “fitting” imply that explaining a phenomenon is a metacognitive process, not an implicit or otherwise inarticulate process or skill. Accordingly, explanation is portrayed as detective work, in which the investigator consciously pieces together otherwise disparate facts into a coherent, global picture. But scientists are not little epistemologists and, apart from the lay appetite for scientists’ grand Nobellizing on the “big” issues, working scientists’ explanatory concerns are local rather than global. Given the philosophical tendency to focus on the explicit detailing of nomic subsumption, establishing fit, and describing mechanisms, Ronald Giere's characterization of philosophical work on explanation nicely captures the current situation: “[M]ost philosophical writing on ‘scientific explanation’ is not really about explanation within science, but about the use of scientific knowledge in the explanation of events in everyday life” (Reference Giere1988, 104–105).Footnote ⁴

2. The Traditional Epistemic Analysis Of Explanatory Understanding

If the research findings on judgment and decision-making examined below are accepted, then the sense of understanding introduces a systematic, but regrettably inaccurate, index of intellectual achievement. In particular, this sense of understanding results from hindsight and overconfidence biases, as well as mistaken attachment to the idea that transparency is routinely achievable. These biases diminish the truth-tracking role of explanation. It is not just philosophical theories of explanation that have accorded to the sense of understanding an essential role in explanation. Psychological theories of explanation, too, appeal to the important role of a sense of understanding in both everyday and scientific explanation. Like some global, unifying accounts of explanation in the philosophy of science, a prominent psychological account focuses on the unified conceptual framework it provides: “. . . [I]n everyday use an explanation is an account that provides a conceptual framework for a phenomenon (e.g., fact, law, theory) that leads to a feeling of understanding in the reader-hearer” (Reference Brewer William and SamarapungavanBrewer et al. 1998, 120). And scientific explanations are no different in this respect; they should “provide a feeling of understanding” (Reference Brewer William and Samarapungavan1998, 121), in addition to whatever objective virtues the explanation might possess. The operative assumption may be that this sense of understanding is an epistemic virtue; the fact that an explanation conveys this sense of understanding offers a reason for thinking it is also a true, or at least a good, explanation. However, in order to be an epistemic virtue, it should be at least positively correlated with accurate causal descriptions, or the goodness of an explanation. But it isn't; or so I will argue. Instead, this sense is a consequence of largely nonepistemic forces, such as the demonstrated psychological tendency of overconfidence, or the desire to reduce, as William James put it, “that peculiar feeling of inward unrest known as indecision” (Reference James[1890] 1981, Chapter 26 (Will), 1136).Footnote ⁵ Formulating a unified, consistent story is one way to eradicate that peculiar feeling of inward unrest. An explanation pleasantly discharges that feeling of intellectual unease.Footnote ⁶ But unity and consistency with background knowledge is a poor substitute for accuracy and truth, as the litany of false but unifying and consistent theories in the history of science should warn. A good story, and so a good explanation, is coherent. Coherence appears to consist of three features: completeness, plausibility and consistency. One might suppose that a plausibility requirement addresses the worry about false theories. However, judgments of plausibility are themselves theory-dependent, and so if your theory is badly false you cannot expect your explanations to fare much better.Footnote ⁷

Most of the widely discussed accounts of explanation have been objectivist: What makes an explanation good concerns a property that it has independent of the psychology of the explainers; it concerns features of external objects, independent of particular minds. These properties might be formal or they might be causal. As unassailable as these objectivist credentials might be, even the most trenchantly objectivist philosophers of science are tempted by the allure of internal access. Hempel treated explanations as formal arguments, weakening the formal criteria in response to counterexamples as seemed fit. Hempel's Deductive-Nomological (D-N) model of explanation attempted to capture this subjective component in the psychological experience of expectability: “[G]iven particular circumstances and the laws in question, the occurrence of the phenomenon was to be expected; and it is in this sense that the explanation enables us to understand why the phenomenon occurred” (Reference Hempel1965, 337). Of course, the crucial phrase “understand why” must be glossed in a nomic way here.Footnote ⁸

Wesley Salmon, too, expressed the centrality of understanding to explanation, even in the context of his treatment of explanation as “causal-mechanical” and “ontic”: “To understand the world and what goes on in it, we must expose its inner workings. To the extent that causal mechanisms operate, they explain how the world works” (Reference SalmonSalmon 1984, 133). Presumably ‘understand’ here is cast in terms of internal access; ‘exposing’ the world's inner workings normally requires the ability to represent implementations of (in this case) mechanisms. That is, what makes the knowledge appropriately expository is that significant parts of its complexity can be represented, and what affords understanding is that we can represent it. This latter feature is made possible by subjective access.

In more recent work, Salmon became more explicitly permissive about the requirements of explanatory understanding, and located the intellectual value of scientific explanations in their power to achieve a number of different goals, all of which

In his wonderfully accessible “The Importance of Scientific Understanding”, Wesley Reference Salmon and SalmonSalmon (1998) distinguished between what he called “scientific” and “psychological” senses of understanding. The scientific sense of understanding

The objectivity of explanation is undermined, however, if the “good reason to suppose” condition is interpreted in internalist terms.

Fitting these mechanisms into a scientific “world-picture” is an admirable aim, and it animated the philosophes of the Enlightenment. But nascent specialization later dashed those hopes, making the theoretical arcaneness of scientific knowledge at once a powerful source of theoretical progress and an equally potent source of suspicion among those wanting scientific matters to be decidable by any person who was intelligent and generally educated. The admirable aim was not achieved, and the science that Enlightenment figures so loved left them behind.

Salmon's discussion tempts many questions, mostly urgent requests for further elaboration of the character of this global understanding. Because there is significant and honest disagreement among scientists about what the correct scientific world-picture is, it is not clear what constitutes fitting phenomena into the scientific world-picture. What Salmon might intend here is the attempted integration of the best-tested and influential views of the recent history of science: quantum mechanics, atomic theory, the germ theory of disease, Darwinian evolution, and so on. Taken together, these theories form a truly impressive monument to the scope, detail, and accuracy of modern science; perhaps they constitute, as Salmon put it, a scientific world picture. But if explanatory scientific understanding requires seeing “how we can fit them [phenomena] into the general scheme of things, that is, into the scientific world-picture” (Reference Salmon and Salmon1998, 87), then most people are incapable of explanatory scientific understanding, including most scientists. Indeed, when scientists piece together phenomena, they do so by focussing on the detailed findings of their (usually) narrow specialization. In contemporary science, global unification arises spontaneously from coordinated piecemeal efforts, not from a meta-level at which the philosopher or reflective scientist assembles remote domains (Reference MillerMiller, 1987). In fact, in light of the arcaneness of contemporary theoretical knowledge, no single individual is so situated.Footnote ⁹ Does Salmon's talk of global explanation entail this internalist conception of understanding? Of course not. Natural definitions seldom have such entailments. But these accounts of explanation are the work of philosophers, not psychologists. We should expect philosophical accounts of understanding to begin, and perhaps end, with the primitive internalist account of subjective sense.

But this convenient account of justification should not be mistaken for an accurate description of scientific practice. When a scientist (or anyone else, for that matter) ultimately accepts an explanation, it is more likely that the scientist, without any such detailed internal representation of a global (and so diverse) subject matter, simply defers to outside experts or appeals to textbook knowledge.Footnote ¹⁰ But when philosophers piece together a global explanation, they look to the subjective sense of fit as a cue for an acceptable explanation.

Michael Friedman's account of explanation, too, is tied to the goal of understanding, and it stipulates special criteria for increased understanding:

The connection that Friedman draws between understanding and number of independent phenomena concerns cognitive efficiency and tractability. If phenomenon P is reduced to Q, then Q is more basic, and so more likely itself to be irreducible. If what makes a phenomenon irreducible is that it can't be explained in terms of anything else more basic, then it is inexplicable. A theory (or explanation) that has fewer inexplicable commitments would seem to be rationally preferable to one that has more. As a psychological hypothesis, it may be easier to comprehend a theory that has just one, rather than two or more, irreducible mysteries.

Friedman proposes that the kind of understanding provided by scientific knowledge is global rather than local. Accordingly, accounts of explanation that identify the objects of understanding as individual phenomena can be expected to fail: “Scientific explanations do not confer intelligibility on individual phenomena by showing them to be somehow natural, necessary, familiar, or inevitable” (Reference Friedman and PittFriedman 1988, 197). It is for this reason that scientific understanding reduces the number of ultimate (and thus inexplicable) commitments.Footnote ¹¹

Philip Kitcher advances a novel and detailed version of the unification approach, adding crucial categories of learning to this account. Kitcher's approach is crafted to honor the work-a-day details of scientific practice. On Kitcher's account, comprehensibility does not emerge spontaneously from the mere reduction of independent theoretical commitments; comprehensibility depends upon the possession of background knowledge and the training required for the assimilation of superficially different phenomena to common categories:

Kitcher captures this objective, nonfoundationalist and pedagogical feature of explanation by asserting that knowledge of a theory

The identification and internalization of argument patterns is an effect of learning, of scientific pedagogy. Of course, familiarity with the esoteric argument patterns that are characteristic of a theory can be acquired only as an arcane skill. I will have more to say about the implicit learning of esoteric detail in section 3.

Unlike other descriptions of unification, for Kitcher the role of explanation issues from dirtying our hands with local details of scientific practice, rather than from philosophical rumination or the supposed scientific aesthetic preference for grand coherence. Explanation plays an epistemic role in science not chiefly because global scientific worldviews set an explicit, top-down research agenda, but because working scientists must piece together local findings within the very small handful of models and theories they use and evaluate.

Once reconstituted as an openly causal account, this treatment of explanation has sufficient detail and descriptive accuracy to accommodate the history of science and current practice. Any such argument patterns—if they are to represent the esoteric practice of science—must be pedagogically valuable and highly enthymematic. Also, there is no reason that the internalization of argument patterns cannot be understood in a more straightforward metaphysical way, in terms of the causal features of internal representations of the explainer and the relevant parts of the world being represented.

3. Understanding, Learning, And Transparent Capture Of Information

‘Understanding’ has many senses. Consider a sense of understanding associated with an explanation of why jets fly. According to one sense of understanding, I understand why planes fly if I know Bernoulli's principle. In such a case, I must have some knowledge of relevant background conditions, of course—how much the plane weighs, the area of the wings, etc. But in a stronger sense of understanding, I can know Bernoulli's principle and still not understand why planes fly, because my knowledge that Bernoulli's principle applies in this case is not associated with a specific kind of subjective state or feeling; it does not have the phenomenology of understanding.

Understanding is usually the consequence of a learning process, generally construed. But there are many modes of learning. We can learn by feedback, with explicit instruction or not. We can learn by analogy, representing to oneself or “internalizing” a model of a process, and then generating various scenarios about the performance of the system under a range of circumstances. Models, of the sort described above, provide a kind of metaphor by which one can transfer a well-understood structure to one less well-understood. Indeed, learning is typically characterized as the “transfer of structure”, where this means either the transfer of informational structure from the environment to an individual (as when one learns first-hand that bees sting), from one individual to others (as when a teacher explains the Pythagorean Theorem to students), or from one psychological subsystem to another (as when one uses their visual information about facial expression to correctly interpret an auditory sequence of spoken language).

We often, perhaps routinely, learn without awareness. In fact, we acquire enormous amounts of information quite incidentally about unattended dimensions of objects. This point can be pressed even further: Surprisingly little learning occurs via the learner transparently representing to itself the as-yet unintellected object. In the last 40 years or so, cognitive psychology has produced a spate of research establishing the importance of implicit learning in a variety of cognitive domains. Consider the research on implicit learning of particular voices. You begin by exposing individuals over several days to a series of words uttered by a number of people. You then ask them to distinguish between words, half of which were, and half of which were not, among the original test set. In this case, reaction times are faster, and error rates lower, for words that were uttered by the same people as those in the original test set. Thus, the same words, uttered by voices different from those in the original test set, required more perceptual and cognitive effort to identify. The redundancy gains achieved by memory for concrete vocal detail were the product of implicit learning. Throughout the training sessions, the participants were not aware that they were learning vocal detail of the particular word-tokens; they thought they were simply learning word-types, as the instructions asked them to do (see Reference Nygaard, Sommers and PisoniNygaard et. al. 1994, Reference GoldingerGoldinger 1996, and Reference Church and SchacterChurch and Schacter 1994). In effect, subjects developed a tacit model of each talker's vocal characteristics, and their cognitive mechanisms engaged that model in matching the input to a prior lexical entry. Because their model was accurate, they enjoyed increased speed and accuracy. Implicit learning has been demonstrated on a variety of perceptual and cognitive tasks, such as lexical decision, picture naming, object decision, word association, category instance generation, and answering general knowledge questions (Reference BaddeleyBaddeley 1996). The lesson here is the same as that for the implicit learning of particular voices: Awareness of the dimensions attended to is not required for learning. This sensitivity to relevant dimensions—a recognitional capacity—is certainly one form of understanding.

If we are trying to sort a collection of instances into natural classes or “kinds” of objects—a taxonomic task required by any science—we must rely on feedback about negative instances in order to improve our discriminative sensitivity. In other words, we must see some non-gazelles if we are to sort objects into classes of gazelles and ibis. There is one major exception to this generalization, and this exception provides the basis of the pedagogical account of explanation I favor. We can learn without exposure to negative instances only when, to quote a famous study on medical diagnostic classification, we have a theoretical model that gives us “access to the major causal influences, possesses accurate measuring instruments to assess them, and uses a well-corroborated theory to make the transition from the theory to fact (that is, when the expert has access to a specific model)” (Reference Dawes Robyn and MeehlDawes, Faust, and Meehl 1989, 1670). In short, for a combination of accuracy and efficiency, there is no explanatory substitute for an accurate model, or a good theory. Let us now turn to the cognitive limits that make the contingent adoption of an accurate model so important.

4. The Sense Of Understanding As The Product Of Hindsight And Overconfidence Biases

If evidence from the history of science shows that sense of understanding is not produced by a reliable relation between belief and truth, where does this subjective sense of understanding derive from? I will argue that in a significant number of cases it comes from two well-documented psychological biases—hindsight and overconfidence—biases that are difficult to correct, and which survive different experimental methods, test items, and classes of people.Footnote ¹² The history of science is a rich source of examples. In a classic moment of hindsight and overconfidence, Ptolemy claimed that “it is idle to seek for causes for the motion of objects towards the centre, once it has been so clearly established from the actual phenomena that the earth occupies the middle place in the universe, and that all heavy objects are carried towards the earth” (Reference Ptolemy and Toomer1984, 43/H22). Ptolemy's conservative appeal to coherence with (in this case, false) background beliefs, together with his dismissive treatment of alternatives, displays his influence by hindsight. But what is the evidence that there is a hindsight bias?

4.1. Hindsight Bias

Explanation accounts for events after the fact. People are notably unaware of the influence that outcome information has on them. This is precisely the retrodictive epistemic position of the explainer. The traditional manner of establishing the hindsight bias begins by asking subjects to estimate the likelihood of various outcomes of an upcoming event, and then retesting them after the event, asking them to recall how likely they had found each of the possible outcomes the first time around. Reference Fischhoff and BeythFischhoff and Beyth (1975) did just that in an early study of the hindsight bias. Prior to President Nixon's trip to China and the Soviet Union in 1972, subjects were asked how likely they found a variety of possible outcomes (for example, whether Nixon would meet Mao, that the Soviet Union and US would establish a joint space program, etc.). Two weeks to six months after the trip the subjects were asked to fill out the same questionnaire. They were asked to recall the probabilities they assigned initially to the same events and, if they couldn't recall, to assign the probability they would have assigned immediately before Nixon's trip. They were also asked if each of the listed outcomes had, in fact, occurred.

The results were a striking demonstration of the distorting influence of hindsight. For those outcomes that subjects thought had occurred, they remembered their estimates as more accurate than they in fact were. For those outcomes thought not to have occurred, subjects recalled their estimates as having been lower than they in fact were. The effect seems to strengthen with the passage of time. After three to six months, 84 percent of the subjects displayed hindsight biases. Therefore, after learning the results of Nixon's trip, subjects believed the outcomes were more predictable than they actually were.

Learning the outcomes of scientific theory testing—the preamble to any explanation—places scientists in a position similar to the hindsight subjects. Explanation is retrospective. When constructing an explanation, we draw one line through many events, a line that could have been drawn in different ways, even if we know the point of termination. As we draw this line, we are not very accurate judges of how much we are affected by information about an outcome, be it a mass extinction, an explosion, or infection. But once the line is drawn, we conceptualize the event as inevitable, and thus people tend to say that the event was fairly predictable all along. Thus, the hindsight bias is also known as the “I-knew-it-all-along effect.” In particular, people tend to overestimate how probable they thought the event was before it occurred.

The hindsight bias is all-too familiar. Many of us have had the experience of telling a colleague over lunch about an experimental finding about human behavior. The reported finding is immediately met with a knowing chuckle: “You didn't need to run an experiment to know that.” Your interlocutor, insensitive to the effect that the reported outcome has just had on him or her, claims to have “known it all along.” Of course, your interlocutor had no such knowledge. Most psychological processes and social behavior are complex, and you don't gain knowledge of them by mere reflection or casual observation; that is why experiments are needed.Footnote ¹³

It is the hindsight bias that lies behind our tendency to confuse predictability with statistical contingency. Prediction is often thought to lift the epistemic burden from explanation. And while predictive accuracy can be an important index of scientific integrity, prediction is both epistemically over-rated and often difficult to secure. For instance, knowing the factors contributing to an effect in a complex system does not necessarily allow you to predict the outcome. Plane crashes are especially good examples of the limited fare supplied by prediction. Upon analysis of cockpit transcripts, a handful of important factors are identified that contributed to the crash. But as Robyn Dawes puts the point: “[T]hese factors would not allow us to predict future crashes very well at all. Airplane crews are often fatigued; bad weather occurs frequently; miscommunication is not that unusual, nor are temporary breakdowns of radio communication or panic at the last minute” (Reference Dawes1999, 37). These effects are too high-frequency to satisfy a low-probability condition for prediction. But more importantly, the hindsight bias causes us to acquiesce in the belief that we have an understanding of an effect. After that, we regard our search as complete. This unjustified conceit is what makes the hindsight bias so damaging to the search for accurate explanations. And the hindsight bias is not alone in the aid and comfort of indolent arrogance.Footnote ¹⁴ The overconfidence bias feeds the lazy beast as well.

4.2 Overconfidence Bias

If error is the constant companion of inquiry, so is overconfidence. Lay adults are systematically prone to believing that they are right when they are not. The literature demonstrating overconfidence is very large, and counterexamples to the effect have been difficult to produce. To cite one representative example of lay overconfidence, Reference Fischhoff Baruch and LichtensteinFischhoff, Slovic and Lichtenstein (1977) asked subjects to indicate the most frequent cause of death in the U.S., and to estimate their confidence that their choice was correct (in terms of “odds”). It turns out that, when subjects set the odds of their answer's correctness at 100:1, they were correct only 73% of the time. Remarkably, even when they were so certain as to set the odds between 10,000:1 and 1,000,000:1, they were correct only between 85% and 90% of the time. It is important to note that the overconfidence effect is systematic and directional. It is highly replicable, and survives changes in task and setting. And, the effect is in the direction of over rather than underconfidence. Given this, it is not surprising that expert training is not the key to deliverance from the overconfidence bias. Physicists, economists, and demographers have all been observed to suffer from this bias, even when reasoning about the content of their special discipline (Reference Henrion and FischoffHenrion and Fischoff 1986). It is of little consolation, then, that good explanation has been associated with the “feeling” or “sense” of understanding. The same sense of understanding, of intellectual conviction, accompanied the above subjects’ incorrect answers. And there is no reason to think that answers to explanatory why-questions are different in a way that allows them to evade that fact.

In fact, the “settled” subjective feeling of understanding that is associated with overconfidence (and so confidence) may be the subjective state that prompts a stopping rule, or a more informal decision that we can stop explaining or considering alternative explanations of an event, on the grounds that we now understand the relevant causes. Overconfidence, then, may be a truly disastrous component in explanatory reasoning. Indeed, our judgments of accuracy are systematically correct only when we have a good theory or model of the process we are making judgments about. This tendency toward overconfidence would not be so damaging if our judgments were “calibrated”, a term used to designate the extent to which confidence matches accuracy. But in science we seldom have such accuracy information available to us.Footnote ¹⁵

Often, we hold a primitive view, one that is independent of or prior to any existing, empirically supported theory, especially when we are innocent of the arcane causes of our behavior. In this situation, the most important determinant of the plausibility of the explanation is the accuracy of the causal theory used to generate it. I take a similar approach to explanation. Accurate explanation appears to be a truth-conducive factor in science; we seem to have scientific knowledge that is explanatory, and this knowledge helps us to generate further knowledge. Unless we rest content with an account of scientific explanation resulting from pure philosophical analysis, we need a scientifically respectable account of scientific explanation itself.

People who believe p, and others who believe ∼p, can both believe overconfidently. True, the priority assigned to this sense of understanding is responsible for much error in the history of science. But there is a systematic and directional effect of this error: Overconfidence extends the life of a false belief, and propagates a true belief. After all, Copernicus claims, of two alternatives to heliocentrism, that “the mind shudders at either of these suppositions” (Reference Copernicus and Adler1952, 514). This is merely an interesting personal report, even though heliocentrism is an accurate theory. The important question is whether heliocentrism is true, not whether envisioning an alternative is too intellectually painful to bear.Footnote ¹⁶

Why do people find it so difficult to learn from the lessons of overconfidence? The answer has two parts. First, it is difficult to learn from individual cases when we typically don't get (or don't retain) systematic feedback about the quality of our judgments. To mention just one restriction on feedback, we can't compare the long-term outcomes of our actual decisions against alternative decisions we didn't implement. Second, the overconfidence is general and persistent, and this attachment to subjective evaluation of complex incoming evidence has specific consequences for learning. Perhaps this is best seen in the “interview effect”, a very prominent example of this overconfidence in our powers of subjective evaluation. When “experts”, such as hiring and admissions officers, are able to review applicants in unstructured interviews, they are still outperformed by statistical prediction rules that take no account of the interviews. In fact, unstructured interviews actually degrade the reliability of human prediction (Reference Bloom, Brundage and StuitBloom and Brundage 1947; Reference DeVaul, Jervey, Chappell, Carver, Short and O’KeefeDeVaul et al. 1957; Reference OskampOskamp 1965; Reference Milstein, Wildkinson, Burrow and KessenMilstein et al. 1981). That is, people degrade the reliability of their predictions by availing themselves of unstructured interviews; the interview information used is irrelevant to (and thus dilutive of) accurate prediction about future performance. Although the interview effect is one of the most robust findings in psychology, highly educated people ignore its obvious practical implications. This occurs in part due to our overconfidence in our subjective ability to “read” people, but is not limited to it. We suppose that our insight into human nature or into the subtleties of nature is so powerful that we can plumb the depths of not only an individual, but a domain, or control the processes used to evaluate it.

Our little conceits are abetted by a classic and systematic frailty in interpreting probabilistic information. This conceit accounts for the general finding that, in the face of a half century of experiments showing the inferiority of human judgment to well-tested statistical prediction rules in such contexts, many disciplinary experts and others still base judgments on subjective impressions and unmonitored evaluation of the evidence. Resistance to these findings runs very deep, and typically comes in the form of a self-serving bias we might call Peirce's Problem.Footnote ¹⁷ Reference Peirce and PeircePeirce ([1878] Reference Peirce and Peirce1982, 281–2) raised what is now the classic worry about frequentist interpretations of probability: How can a probability claim (say, the claim that 99 out of 100 cards are red) be relevant to a judgment about a particular case (whether the next card will be red)? After all, the next card will be red or not, and the other 99 cards can't change that fact. Those who resist the findings of inferior performance are typically quite willing to admit that in the long run, a simple statistical prediction rule will be right more often than human experts. In short, their (over)confidence in subjective powers of reflection leads them to deny that we should believe we are unreliable in some particular case. Whatever the long-run performance of humans, so the sentiment goes, in this case I am the right person to make the subjective judgment. And the effect of this lack of discipline is made worse because it is difficult to secure all of the information necessary for clear feedback: We can't compare the long-term outcomes of our actual decisions against the decisions we would have made if we hadn't interviewed the candidates but simply used a rule.

Our reluctance to envision our fallibility, and then admit error, goes a long way toward accounting for our unwillingness to acknowledge the embarrassments of overconfidence and hindsight biases. On the basis of this evidence, humans appear naturally disposed to exaggerate the powers of our subjective faculties. No matter what this disposition's origin, it won't do to reply to my claims of unreliability that this result seems implausible, for that is exactly the point. For example, the theory predicts that people will find the sad lesson of the interview effect implausible. Despite a moral that should be humbling, we continue to look for, and when it is present we use, that special, subjective sense of illumination as a cue of correctness. By now it should be clear that this cue, by itself, is unreliable, and often isn't present when we have explanatory knowledge. So clearly if explanation imparts understanding, understanding must consist in something other than a subjective sense of metacognitive control.

5. Conclusion

Despite these intellectual warts on the phenomenology of explanatory understanding, many philosophers of science, such as scientific realists, hold that explanation plays a robust, epistemic role in theory construction. But if the traditional, phenomenological sense of understanding is not a valid cue of a good explanation, what accounts for the epistemic function of explanation in theory construction? I have not attempted to answer this question here. While the conception of good explanation I would defend is objectivist, it is virtually unique in its utter disregard for the role of the subjective sense of understanding as a determinant of explanatory goodness.

On a genuinely realist account, understanding is a cluster of epistemic virtues. How could there be intellectual virtue at work without transparent appreciation of the content of one's theoretical beliefs, content that characterizes the causes of the phenomenon to be explained? The parallel with perceptual knowing is instructive here. A nonskeptic would regard the faculty of vision as a justification-conferring faculty even though the successful perceiver may know nothing about visual transduction, or about the vast array of other physiological causes of visual knowledge. It would seem that knowledge does not require conscious access to information that mediates input and behavior. Why, then, should explanation be essentially tied to a subjective sense of understanding? Nonformal approaches to explanation may employ causal, or model-theoretic notions. But they also use sense of understanding as a cue of those specific notions’ plausibility. This is a problem faced by all philosophers of science who suppose that the feeling of understanding attending the intellectual assembling of mechanisms constitutes a reason for thinking our account is correct.

The fact is, our history is littered with inaccurate explanations we confidently thought were obviously true: the explanation for mental illness in terms of demonic possession, the humoral theory of illness, and so on. The sense of understanding would be epistemically idle phenomenology were it not so poisonous a combination of seduction and unreliability. It actually does harm, sometimes making us squeamish about accepting true claims that we don't personally understand, and more often operating in the opposite direction, causing us to overconfidently accept false claims because they have a kind of anecdotal or theoretical charm.

Moreover, the understanding conveyed by a good explanation may be a community achievement. Except for the simplest of events, explanatory understanding is not essentially an achievement of an individual. And any alternative account of explanation that requires the transmission of a sense of understanding must address this criticism. My positive account of scientific explanation asserts that, as a contingent matter of fact, the only feature of an explanation that can render explanation epistemic is its systematic tendency to produce increasingly accurate theories. In effect, only explanations capable of sustaining theoretical progress are good explanations. Such explanations are usually approximately true, rather than merely better than available alternatives, as the latter may simply be the best of a rotten lot. In order to accord explanation the epistemic role it seems to play in successful theory selection in contemporary science, we must abandon our sentimental attachment to the comforting sense of understanding, or, at least, abandon the idea that this sense is a valid cue of truth. This will not be easy. Explanation is a backward-looking affair and, as we have seen, 30 years of research on judgment documents that people are not good at tracking how they are affected by knowledge of outcomes. As Robyn Dawes puts it, “The problem is that there is a many-many relationship between antecedents and consequences in the course of human life. As we retrospect, in contrast, we can create many-one relationships.” (Reference Dawes1999, 37) Explanation creates many-one relationships.

The preceding assault on the role of a subjective sense of understanding in explanation leaves untouched many admirable features of more orthodox accounts of explanation. Nomic subsumption, unification, and logical derivation all may advance scientific aims. In the same sense that casting problems in terms of an independently well supported theory (e.g., materialism, germ theory of disease, etc.) can help to solve them correctly, an explanation that is accurate—no matter how nontransparent its components—is better than an inaccurate one that merely makes us feel that we understand. Only the former is related to the world in a way that conduces to the success of science. If our subjective sense of understanding seems to be reliably related to truth only when we have a good theory, one must wonder whether that subjective sense of understanding is doing any epistemic work at all. It appears that, when it comes to explanation, there is no substitute for simply being (approximately) right. A consistent scientific realist is driven to this historically contingent, radically objectivist account of explanation. What would this scientific realist account of explanation look like? Given our proclivity to confidently concoct a story when we don't have a well-confirmed account ready-to-hand,Footnote ¹⁸ saying less is truly more. So, for the time being, I will take the cue.