Rationality and Science
Paul Thagard
Philosophy Department
University of Waterloo

Thagard, P. (2004). Rationality and science. In A.Mele & P. Rawlings (Eds.), Handbook of rationality. Oxford: Oxford University Press, (pp. 363-379).


Are scientists rational? What would constitute scientific rationality? In the philosophy of science, these questions are usually discussed in the context of theory choice: What are the appropriate standards for evaluating scientific theories, and do scientists follow them? But there are many kinds of scientific reasoning besides theory choice, such as analyzing experimental data. Moreover, reasoning in science is sometimes practical, for example when scientists decide what research programs to pursue and what experiments to perform. Scientific rationality involves groups as well as individuals, for we can ask whether scientific communities are rational in their collective pursuit of the aims of science.

This chapter provides a review and assessment of central aspects of rationality in science. It deals first with the traditional question: What is the nature of the reasoning by which individual scientists accept and reject conflicting hypotheses? I will also discuss the nature of practical reason in science, and then turn to the question of the nature of group rationality in science. The remainder of the chapter considers whether scientists are in fact rational, that is, whether they conform to normative standards of individual and group rationality. I consider various psychological and sociological factors that have been taken to undermine the rationality of science.

What is Science For?

First, however, it is necessary to deal with a prior issue: What are the goals of science? In general, rationality requires reasoning strategies that are effective for accomplishing goals, so discussion of the rationality of science must consider what science is supposed to accomplish. To begin, we can distinguish between the epistemic and the practical goals of science. Possible epistemic goals include truth, explanation and empirical adequacy. Possible practical goals include increasing human welfare through technological advances. My view is that science has all of these goals, but let us consider some more extreme views.

Some philosophers have advocated the view that the primary epistemic aim of science is the achievement of truth and the avoidance of error (Goldman, 1999). On this view, science is rational to the extent that the beliefs that it accumulates are true, and scientific reasoning is rational to the extent that it tends to produce true beliefs. The philosophical position of scientific realism maintains that science aims for true theories and to some extent accomplishes this aim, producing some theories that are at least approximately true. In contrast, the position of anti-realism is that truth is not a concern of science. One of the most prominent anti-realists is Bas van Fraassen (1980), who argues that science aims only for empirical adequacy: scientific theories should make predictions about observable phenomena, but should not be construed as true or false. The anti-realist view, however, is at odds with the practice and success of science (see Psillos, 1999, for a systematic defense). Most scientists talk and act as if they are trying to figure out how the world actually works, not just attempting to make accurate predictions. Moreover, the impressive technological successes of science are utterly mysterious unless the scientific theories that made them possible are at least approximately true. For example, my computer would not be processing this chapter unless there really are electrons moving through its silicon chips.

But truth is not the only goal of science. The most impressive accomplishments of science are not individual facts or even general laws, but broad theories that explain a great variety of phenomena. For example, in physics the theory of relativity and quantum theory each provide understanding of many phenomena, and in biology the theory of evolution and genetic theory have very broad application. Thus a major part of what scientists strive to do is to generate explanations that tie together many facts that would individually not be very interesting. A scientist who aimed only to accumulate truths and avoid errors would be awash in trivialities. Hence science aims for explanation as well as truth. These two goals subsume the goal of empirical adequacy, because for most scientists the point of describing and predicting observed phenomena is to find out what is true about them and to explain them.

But there are also practical goals that science accomplishes. Nineteenth-century physicists such as Faraday and Maxwell were primarily driven by epistemic goals of understanding electrical and magnetic phenomena, but their work made possible the electronic technologies that now pervade human life. Research on topics such as superconductivity and lasers has operated with both scientific and technological aims. Molecular biology is also a field that began with primarily epistemic aims but that has increasingly been motivated by potential practical applications in medicine and agriculture. Similarly, the major focus of the cognitive sciences such as psychology and neuroscience has been understanding the basic mechanisms of thinking, but there have also been practical motivations such as improving education and the treatment of mental illnesses. It is clear, therefore, that one aim of the scientific enterprise is the improvement of human welfare through technological applications. This is not say that each scientist must have that aim, since many scientists work far from areas of immediate application, but science as a whole has made and should continue to make technological contributions.

More critical views on the practical aims of science are extant. It has been claimed that science functions largely to help maintain the hegemony of dominant political and economic forces by providing ideologies and technologies that forestall the uprising of oppressed peoples. This claim is a gross exaggeration, but there is no question that the products of scientific research can have adverse effects, for example the use of dubious theories of racial superiority to justify social policies, and the use of advanced technology to produce devastating weapons. But saying that the aims of science are truth, explanation, and human welfare does not imply that these aims are always accomplished, only that these are the aims that science generally does and should have. We can now address the question of what strategies of rational thinking best serve the accomplishment of these aims.

Models of Individual Rationality

Consider a recent example of scientific reasoning, the collision theory of dinosaur extinction. Since the discovery of dinosaur fossils in the nineteenth century, scientists have pondered why the dinosaurs became extinct. Dozens of different explanations have been proposed, but in the past two decades one hypothesis has come to be widely accepted: dinosaurs became extinct around 65 million years ago because a large asteroid collided with the earth. Evidence for the collision hypothesis includes the discovery of a layer of iridium (a substance more common in asteroids than on earth) in geological formations laid down around the same time that the dinosaurs became extinct. What is the nature of the reasoning that led most paleontologists and geologists to accept the collision hypothesis and reject its competitors? I shall consider three main answers to this question, derived from confirmation theory, Bayesian probability theory, and the theory of explanatory coherence. In each case, I will describe a kind of ideal epistemic agent, and consider whether scientists are in fact agents of the specified kind.

Confirmation and Falsification

Much work in the philosophy of science has presumed that scientists are confirmation agents that operate roughly as follows (see, for example, Hempel, 1965). Scientists starts with hypotheses that they use to make predictions about observable phenomena. If experiments or other observations show that the predictions are true, then the hypotheses are said to be confirmed. A hypothesis that has received substantial empirical confirmation can be accepted as true, or at least as empirically adequate. For example, the hypothesis that dinosaurs became extinct because of an asteroid collision should be accepted if it as been confirmed by successful predictions.
Popper (1959) argued that scientists should not aim for confirmation, but should operate as the following sort of falsification agents. Scientists use hypotheses to make predictions, but their primary aim should be to find evidence that contradicts the predicted results, leading to the rejection of hypotheses rather than their acceptance. Hypotheses that have survived severe attempts to falsify them are said to be corroborated. On this view, the proponents of the collision theory of dinosaur extinction should attempt to falsify their theory by stringent tests, and only then consider them as corroborated, but not as accepted as true.

Although hypotheses are often used to make predictions, the process of science is much too complex for scientists to function generally as either confirmation agents or falsification agents. In particular, it is exceedingly rare for scientists to set out to refute their own hypotheses, and, given the difficulty of performing complex experiments, it is fortunate that they aim for confirmations rather than refutations. There are many reasons why an experimental prediction might fail, ranging from problems with instruments or personnel to failure to control for key variables. A falsification agent would frequently end up throwing away good hypotheses.
But scientists are not just confirmation agents either, since hypotheses often get support, not just from new predictions, but from explaining data already maintained. Moreover, it often happens in science that there are conflicting hypotheses that are to some extent confirmed by empirical data. As Lakatos (1970) argued, the task then is not just to determine what hypotheses are confirmed, but rather what hypotheses are better confirmed than their competitors. Hypothesis assessment is rarely a matter of evaluating a hypothesis with respect to its predictions, but rather requires evaluating competing hypotheses, with the best to be accepted and the others to be rejected. There are both probabilistic and explanatory approaches to such comparative assessment.


Carnap and numerous other philosophers of science have attempted to use the resources of probability theory to illuminate scientific reasoning (Carnap, 1950; Howson and Urbach, 1989; Maher, 1993). Probabilistic agents operate as follows. They assess hypotheses by considering the probability of a hypothesis given the evidence, expressed as the conditional probability P(H/E). The standard tool for calculating such probabilities is Bayes's Theorem, one form of which is:
P(H/E) = P(H) * P(E/H) / P(E).
This says that the posterior probability of the hypothesis H given the evidence E is calculated by multiplying the prior probability of the hypothesis by the probability of the evidence given the hypothesis, all divided by the probability of the evidence. Intuitively, the theorem is very appealing, with a hypothesis becoming more probable to the extent that it makes improbable evidence more probable. Probabilistic agents look at all the relevant evidence, calculate values for P(E) and P(E/H), take into account some prior value of P(H), and then calculate P(H/E). Of two incompatible hypotheses, probabilistic agents prefer the one with the highest posterior probability. A probabilistic agent would accept the collision theory of dinosaur extinction if its probability given the evidence is higher than the probability of competing theories.

Unfortunately, it is not so easy as it sounds for a scientist to be a probabilistic agent. Various philosophers, e.g. Glymour (1980) and Earman (1992), have discussed technical problems with applying probability theory to scientific reasoning, but I will mention only what I consider to be the three biggest roadblocks. First, what is the interpretation of probability in P(H/E)? Probability has its clearest interpretation as frequencies in populations of observable events; for example, the probability that a die will turn up a 3 is 1/6, meaning that in a large number of trials there will tend to be 1 event in 6 that turns up a 3. But what meaning can we attach to the probability of dinosaur extinction being caused by an asteroid collision? There is no obvious way to interpret the probability of such causal hypotheses in terms of objective frequencies in specifiable populations.
The alternative interpretation is that such probabilities are degrees of belief, but there is substantial evidence that people's thinking does not conform to probability theory (e.g. Kahneman, Slovic, and Tversky, 1982; Tversky, 1994). One might say that the probability of a hypothesis is an idealized degree of belief, but it is not clear what this means. Degree of belief is sometimes cashed out in terms of betting behavior, but what would it mean to bet on the truth of various theories of dinosaur extinction?

The second difficulty in viewing scientists as probabilistic agents is that it there are computational problems in calculating probabilities in accord with Bayes's theorem. In general, the problem of calculating probabilities is computationally intractable in the sense that the number of conditional probabilities required increases exponentially with the number of propositions. However, powerful and efficient algorithms have been developed for calculating probabilities in Bayesian networks that make simplifying assumptions about the mutual independence of different propositions (Pearl, 1988). No one, however, has yet used Bayesian networks to simulate a complex case of scientific reasoning such as debates about dinosaur extinction. In contrast, the next section discusses a computationally feasible account of scientific inference based on explanatory coherence.
The third difficulty with probabilistic agents is that they may ignore qualitative factors affecting theory choice. Scientists' arguments suggest that they care not only how much evidence there is for a theory, but also about the variety of the evidence, the simplicity of the theory that accounts for it, and analogies between proposed explanations and other established ones. Perhaps simplicity and analogy could be accounted for in terms of prior probabilities: a simpler theory or one offering analogous explanations would get a higher value for P(H) to be fed into the calculation via Bayes's theorem of the posterior probability P(E/H). But the view of probability as subjective degree of belief leaves it mysterious how people do or should arrive at prior probabilities.

Explanatory Coherence

If scientists are not confirmation, falsification, or probabilistic agents, what are they? One answer, which goes back to two nineteenth-century philosophers of science, William Whewell and Charles Peirce, is that they are explanation agents. On this view, what scientists do in theoretical inference is to generate explanations of observed phenomena, and a theory is to be preferred to its competitors if it provides a better explanation of the evidence. Theories are accepted on the basis of an inference to the best explanation. Such inferences are not merely a matter of counting which of competing theories explains more pieces of evidence evidence, but requires assessment in terms of the overall explanatory coherence of each hypothesis with respect to a scientist's whole belief system. Factors that go into this assessment for a particular hypothesis include the evidence that it explains, its explanation by higher-level hypotheses, its consistency with background information, its simplicity, and analogies between the explanations offered by the hypothesis and explanations offered by established explanations (Harman, 1986; Lipton, 1991; Thagard, 1988).

The major difficulty with the conception of scientists as explanatory agents is the vagueness of concepts such as explanation, inference to the best explanation, and explanatory coherence. Historically, explanation has been conceptualized as a deductive relation, a probabilistic relation, and a causal relation. The deductive conceptualization of explanation fits well with the confirmation and falsification view of agents: a hypothesis explains a piece of evidence if the a description of the evidence follows deductively from the hypothesis. Similarly, the probabilistic conceptualization of explanation fits well with the probabilistic view of agents: a hypothesis explains a piece of evidence if the probability of the evidence given the hypothesis is higher than the probability of the evidence without the hypothesis. Like Salmon (1984) and others, I prefer a conceptualization of explanation as the provision of causes: a hypothesis explains a piece of evidence if it provides a cause of what the evidence describes. The causal conceptualization must face the problem of saying what causes are and how causal relations are distinct from deductive and probabilistic ones (see Thagard 1999, ch. 7)

Assuming we know what an explanation is, how can we characterize inference to the best explanation? I have shown how a precise and easily computable notion of explanatory coherence can be applied to many central cases in the history of science (Thagard, 1992). For example, we can understand why the collision theory of dinosaur extinction has been accepted by many scientists but rejected by others by assessing its explanatory coherence with respect to the evidence available to different scientists (see Thagard, 1991, for computer simulations of the dinosaur debate using the program ECHO).

I prefer to view scientists as explanation agents rather than as confirmation, falsification, or probabilistic agents because this view fits better with the historical practice of scientists as evident in their writings, as well as with psychological theories that are skeptical about the applicability of deductive and probabilistic reasoning in human thinking. But I acknowledge that the probabilistic agent view is probably the most popular one in contemporary philosophy of science; it has largely absorbed the confirmation agent view by the plausible principle that evidence confirms a hypothesis if and only if the evidence makes the hypothesis more probable, i.e. P(H/E) > P (H). It is also possible that that scientists are not rational agents of any of these types, but rather are reasoners of a very different sort. For example, Mayo (1996) develops a view of scientists as modeling patterns of experimental results that are useful for distinguishing errors. Solomon (2001) describes scientists as reaching conclusions based on a wide variety of "decision vectors," ranging from empirical factors such as salience of data to non-empirical factors such as ideology.

Practical Reason

As mentioned in this chapter's introduction, there is much more to scientific rationality than accepting and rejecting hypotheses. Here are some of the important decisions that scientists make in the course of their careers:
1. What general field of study should I enter, e.g. should I become a paleontologist or a geologist?
2. Where and with whom should I study?
3. What research topics should I pursue?
4. What experiments should I do?
5. With whom should I collaborate?
When scientists make these decisions, they are obviously acting for more than epistemic reasons, entering a field for more reasons than that it would maximize their stock of truths and explanations. Scientists have personal aims as well as epistemic ones, such as having fun, being successful, living well, becoming famous, and so on. Let us now consider two models of scientists as practical decision makers: scientists as utility agents and scientists as emotional agents.
The utility agent view is the familiar one from economics, with an agent performing an action because of a calculation that the action has more expected utility than alternative actions, where expected utility is a function of the utilities and probabilities of different outcomes. This view is consonant with the epistemic view of scientists as probabilistic agents, and has many of the same difficulties. When scientists are considering between different research topics, do they have any idea of the relevant probabilities and utilities? Suppose I am a molecular biologist doing genome research, and have to decide whether to work with yeast or with worms? I may have hunches about which research program may yield the more interesting results, but it is hard to see how these hunches could be translated into anything as precise as probabilities and utilities.

A more realistic view of the decision making of scientists and people in general is that we choose the actions that receive the most positive emotional evaluation based on their coherence with our goals (Thagard, 2000, ch. 6; Thagard, 2001). On this view, decision making is based on intuition rather than on numerical calculation: unconsciously we balance different actions and different goals, arriving at a somewhat coherent set of accepted ones. The importance of goals is affected by how they fit with other goals as well as with the different actions that are available to us. We may have little conscious awareness of this balancing process, but the results of the process comes to consciousness via emotions. For example, scientists may feel excited by a particular research program and bored or even disgusted by an alternative program. Psychologists use the term valence to refer to positive or negative emotional evaluations. For discussions of the role of emotions in scientific thinking, see Thagard (forthcoming-a, forthcoming-b). Like Nussbaum (2001), I view emotions as intelligent reactions to perceptions of value, including epistemic value.

Just as there is a concordance between the probabilistic view of epistemic agents and the utility view of practical agents, there is a concordance between the explanatory coherence view of epistemic agents and the emotional coherence view of practical agents. In fact, emotions play a significant role in inference to hypotheses as well as in inference to actions, because the inputs to and outputs from both kinds of inference are emotional as well as cognitive. The similarity of outputs is evident when scientists appreciate the great explanatory power of a theory and characterize it as elegant, exciting, or even beautiful. As with practical judgments of emotional coherence in practical decision making, we have no direct conscious access to the cognitive processes by which we judge some hypotheses to be more coherent than others. What emerges to consciousness from a judgment of explanatory coherence is often emotional, in the form of liking or even joy with respect to one hypothesis, and dislike or even contempt for rejected competing hypotheses. For example, when Walter and Luis Alvarez came up with the hypothesis that dinosaurs had become extinct because of an asteroid collision, they found the hypothesis not only plausible but exciting (Alvarez, 1998). In contrast, some skeptical paleontologists thought the hypothesis was not only dubious but ridiculous. Emotional inputs to hypothesis evaluation include the varying attitudes that scientists toward different experimental results and even to different experiments ­ any good scientist knows some experiments are better than others. Another kind of emotional input is analogical: a theory analogous to a positively-viewed theory such as evolution will have greater positive valence than one that is analogous to a scorned theory such as cold fusion.

Thus my view of scientists as explanatory-emotional agents is very different from the view of them as probabilistic-utility agents. My emphasis on emotions will probably have readers wondering whether scientists are rational at all. Perhaps they are just swayed by their various intellectual prejudices and personal desires to plan research programs and accept hypotheses in ways that disregard the epistemic aims of truth and explanation. There are, unfortunately, cases where scientists are deviant in these ways, with disastrous results such as fraud and other kinds of unethical behavior. But the temperaments and training of most scientists is such that they have an emotional attachment to the crucial epistemic aims. Many scientists become scientists because they enjoy finding out how things work, so that the aims of truth and explanation are with them from the beginnings of their scientific training. These attachments can be fostered by working with advisors who not only value these aims but transmit their emotional evaluations of them to the students and postdoctoral fellows with whom they work. So, for most scientists, a commitment to fostering explanation and truth is an emotional input into their practical decision making.

Models of Group Rationality

As Kuhn (1970) and many other historians, philosophers, and sociologists of science have noted, science is not merely a matter of individual rationality. Scientists do their work in the context of groups of various sizes, from the research teams in their own laboratories to community of scientists working on similar projects to the overall scientific community. As I have documented elsewhere (Thagard 1999, ch. 11), most scientific articles have multiple authors, and the trend is toward increasing collaboration. In addition, all scientists operate within the context of a wider community with shared societies, journals, and conferences. Therefore the question of the rationality of science can be raised for groups as well as individuals: What is it for a group of scientists to be collectively rational, and are such groups generally rational? I will assume that groups of scientists have the same primary aims that I attributed to science in general: truth, explanation, and human welfare via technological applications.

It might seem that the rationality of scientific groups is just the sum of the rationality of the individuals that comprise them. Then a group is rational if and only if the individual scientists in it are rational. But it is possible to have individual rationality without group rationality, if the pursuit of scientific aims by each scientist does not add up to optimal group performance. For example, suppose that each scientist rationally chooses to pursue exactly the same research strategy as the others, with the result that there is little diversity in the resulting investigations and paths that would be more fertile with respect to truth and explanation are not taken. Philosophers such as Kitcher (1993) have emphasized the need for cognitive diversity in science.
On the other hand, it might be possible to have group rationality despite lack of individual rationality. Hull (1989) has suggested that individual scientists who seek fame and power rather than truth and explanation may in fact contribute to the overall aims of science, because their individualistic pursuit of non-epistemic motives in fact leads the scientific group as a whole to prosper. This is analogous to Adam Smith's economic model in which individual greed leads to overall economic growth and efficiency.

It is important to recognize also that group rationality in science is both epistemic and practical. Of a particular scientific community, we can ask two kinds of question:
(1) Epistemic: Given the evidence, what should be the distribution of beliefs in the community?
(2) Practical: What should be the distribution of research initiatives in the community?
For the epistemic question, it might be argued that if all scientists have access to the same evidence and hypotheses, then they should all acquire the same beliefs. Such unanimity would, however, be detrimental to the long-term success of science, since it would reduce cognitive diversity. For example, if Copernicus had been enmeshed within the Ptolemaic theory of the universe, he might never have generated his alternative heliocentric theory that turned out to be superior with respect to both truth and explanation. Similarly, in the dinosaur case Walter Alvarez would never have formulated his theory of why dinosaurs became extinct if he had been a conventional paleontologist.

Moreover, epistemic uniformity would contribute to practical uniformity, which would clearly be disastrous. It would be folly to have all scientists within a scientific community following just a few promising leads, since this would reduce the total accomplishment of explanations as well as retard the development of novel explanations. Garrett Hardin (1968) coined the term "tragedy of the commons" to describe a situation in which individual rationality could promote group irrationality. Consider sheep herders who share a common grazing area. Each herder separately may reason that adding one more sheep to his or her herd would not have any serious effect on the common area. But such individual decisions might collectively produce over-grazing, so that there is not enough food for any of the sheep, with the result that all sheep herders are worse off. Analogously, we can imagine in science and other organizations a kind of "tragedy of consensus", in which the individuals all reach similar conclusions about what to believe, stifling creative growth.
So, what should be our model of group rationality in science? Kitcher (1993) and Goldman (1999) develop models of group rationality that assume that individual scientists are probabilistic agents. Although these analyses are interesting with respect to cognitive diversity and truth attainment, I do not find them plausible because of the problems with the probabilistic view discussed in the last section. As an alternative, I have developed a model of scientific consensus based on explanatory coherence.

This model is called CCC, for consensus = coherence + communication (Thagard 2000, ch. 10). It assumes that each scientist is an explanation agent, accepting and rejecting hypotheses on the basis of their explanatory coherence with evidence and alternative hypotheses. Communication takes place as the result of meetings between scientists in which they exchange information about available evidence and hypotheses. If all scientists acquire exactly the same information, then they will agree about what hypotheses to accept and reject. However, in any scientific community, exchange of information is not perfect, so that some scientists may not hear about some of the evidence and hypotheses. Moreover, different scientists have different antecedent belief systems, so the overall coherence of a new hypothesis may be different for different scientists. Ideally, however, if communication continues there will eventually be community consensus as scientists accumulate the same sets of evidence and hypotheses and therefore reach the same coherence judgments. The CCC model has been implemented as a computational extension of the explanatory coherence program ECHO in which individual scientists evaluate hypotheses on the basis of their explanatory coherence but also exchange hypotheses and evidence with other scientists. These simulated meetings can either be pairwise exchanges between randomly selected pairs of scientists, or "lectures" of the sort that take place at scientific conferences in which one scientist can broadcast sets of hypotheses and evidence to a group of scientists. Of course, communication is never perfect, so it can take many meetings before all scientists acquire approximately the same hypotheses and evidence. I have performed computational experiments in which different numbers of simulated scientists with varying communication rates achieve consensus in two interesting historical cases: theories of the causes of ulcers, and theories of the origins of the moon.

The CCC model shows how epistemic group rationality can arise in explanation agents who communicate with each other, but it tells us nothing about practical group rationality in science. One possibility would be to attempt to extend the probabilistic-utility model of individual practical reason. On this model, each scientist makes practical decisions about research strategy based on calculations concerning the expected utility of different courses of action. Research diversity arises because different scientists attach different utilities to various experimental and theoretical projects. For reasons already given, I would prefer to extend the explanatory-emotional model described in the previous section.

The extension arises naturally from the CCC model just described, except that in large, diverse communities we should not expect the same degree of practical consensus as there is of epistemic consensus, for reasons given below. For the moment, let us focus on particular research groups rather than on whole scientific communities. At this level, we can find a kind of local consensus that arises because of emotional coherence and communication. The characteristics of the group include the following:
1. Each scientist is an explanation agent with evidence, hypotheses, and the ability to accept and reject them on the basis of explanatory coherence.
2. In addition, each scientist is an emotional agent with actions, goals, valences and the ability to make decisions on the basis of emotional coherence.
3. Each scientist can communicate evidence and hypotheses with other scientists.
4. Each scientists can, at least sometimes, communicate actions, goals, and valences to other scientists.
5. As the result of cognitive and emotional communication, consensus is sometimes reached about what to believe and also about what to do.

The hard part to implement is the component of (4) that involves valences. It is easy to extend the CCC model of consensus to include emotional coherence simply by allowing actions, goals, and valences to be exchanged just like evidence, hypotheses, and explanations.
In real life, valences are not so easily exchanged as verbal information about actions, goals, and what actions accomplish which goals. Just hearing someone say that they really care about something does not suffice to make you care about it too, nor should it, because your goals and valences may be orthogonal or even antagonistic to mine. So in a computational model of emotional consensus the likelihood of exchange of goals and valences in any meeting should be much lower than the likelihood of exchange of hypotheses, evidence, and actions.
Still, in real-life decision making involving scientists and other groups such as corporate executives, emotional consensus is sometimes reached. What are the mechanisms of valence exchange, that is, how do people pass their emotional values on to other people? Two relevant social mechanisms are emotional contagion and attachment-based learning. Emotional contagion occurs when person A expresses an emotion, person B unconsciously mimics A's facial and bodily expressions, and then begins to acquire the same emotion (Hatfield, Cacioppo, and Rapson, 1994). For example, if a group member enthusiastically presents a research strategy, then the enthusiasm may be conveyed through both cognitive and emotional means to other members of the group. The cognitive part is that the other group members become aware of possible actions and their potential good consequences, and the emotional part is conveyed by the facial expressions and gestures of the enthusiast, so that the positive valence felt by one person spreads to the whole group. Negative valence can also spread, not just from a critic pointing out drawbacks to a proposed action as well as more promising alternatives, but also by contagion of the negative facial and bodily expressions.
Another social mechanism for valence exchange is what Minsky (2001) calls attachment-based learning. Minsky points out that cognitive science has developed good theories of how people use goals to generate sub-goals, but has had little to say about how people acquire their basic goals. Similarly, economists employing the expected utility model of decision making take preferences as given, just as many philosophers who hold a belief-desire model of rationality take desires as given.

Minsky suggests that basic goals arise in children as the result of praise from people to whom the children are emotionally attached. For example, when young children share their toys with their playmates, they often receive praise from their parents or other caregivers. The parents have positive valence for the act of sharing, and the children may also acquire a positive emotional attitude toward sharing as the result of seeing that it is something cared about by people whom they care about and who care about them. It is not just that sharing becomes a sub-goal to accomplish the goal of getting praised by parents; rather, being kind to playmates becomes an internalized goal that has intrinsic emotional value to the children.

I conjecture that attachment-based learning also occurs in science and other contexts of group decision making. If your supervisor is not just a boss but a mentor, then you may form an emotional attachment that makes you particularly responsive to what the supervisor praises and criticizes. This makes possible the attachment-based transmission of positive values such as zeal for truth and understanding, or, more locally, for integrity in dealing with data and explanations.
Notice that both emotional contagion and attachment-based learning require quite intense interpersonal contacts that will not be achieved in a large lecture hall or video conference room, let alone through reading a published article. The distinguished social psychologist, Richard Nisbett, told me that he learned how to do good experiments through discussions with his supervisor, Stanley Schacter. Nisbett said (personal communication, Feb. 23, 2001) "He let me know how good my idea was by grunts: non-committal (hmmm...), clearly disapproving (ahnn...) or (very rarely) approving (ah!)." These grunts and their attendant facial expressions conveyed emotional information that shaped the valences of the budding researcher.

Accordingly, when I extend my CCC model of consensus as coherence plus communication to include group decisions, I will include two new variables to determine the degree of valence transmission between agents: degree of personal contact, and degree of attachment. If personal contact and attachment are high, then the likelihood of valence transmission will be much greater than in the ordinary case of scientific communication, in which the success of verbal transmission of information of hypotheses, evidence, and actions is much higher than the transmission of valences.

There may, however, be quasi-verbal mechanisms for valence transfer. Thagard and Shelley (2001) discuss emotional analogies whose purpose is to transfer valences as well as verbal information. For example, if a scientist presents a research project as analogous to a scientific triumph such as the asteroid theory of dinosaur extinction, then listeners may transfer the positive value they feel for the asteroid theory to the proposed research project. Alternatively, if a project is analogous to the cold fusion debacle, then the negative valence attached to that case may be projected onto the proposed project. Thus emotional analogies are a third mechanism, in addition to emotional contagion and attachment-based learning, for transfer of valences. All three mechanisms may interact with each other, for example when a mentor uses an emotional analogy and facial expressions to convey values to a protégée. Alternatively, the mentor may function as a role model, providing a different kind of emotional analogy: students who see themselves as analogous to their role models may tend to transfer to themselves some of the motivational and emotional characteristics of their models.

I hope it is obvious from my discussion of practical group rationality in science why science need not succumb to the tragedy of consensus, especially with respect to practical rationality. Communication between scientists is imperfect, both with respect to cognitive information such as hypotheses and evidence and especially with respect to emotional valences for particular approaches. Scientists may get together for consensus conferences such as the ones sponsored by the National Institutes of Health that regularly deal with controversial issues in medical treatment (see Thagard, 1999, ch. 12 for a discussion). But not all scientists in a community attend such conferences or read the publications that emanate from them. Moreover, the kinds of close interpersonal contact needed for communication of values by emotional contagion and attachment-based learning occur only in small subsets of the whole scientific community. Hence accomplishment of the general scientific aims of truth, explanation, and technological applications need not be hindered in a scientific community by a dearth of practical diversity. Solomon (2001) provides a rich discussion of consensus and dissent in science.

Is Science Rational?

A person or group is rational to the extent that its practices enable it to accomplish its legitimate goals. At the beginning of this paper, I argued that the legitimate goals of science are truth, explanation, and technologies that promote human welfare. Do scientific individuals and groups function in ways that further these goals, or do they actually pursue other personal and social aims that are orthogonal or even antagonistic to the legitimate goals? I will now consider several psychological and sociological challenges to the rationality of science.

Psychological challenges can be based on either cold cognition, which involves processes such as problem solving and reasoning, or hot cognition, which includes emotional factors such as motivation. The cold-cognition challenge to scientific rationality would be that people's cognitive processes are such that it is difficult or impossible for them to reason in ways that promote the aims of science. If scientific rationality required people to be falsification agents or probabilistic agents, then the cold-cognition challenge would be a serious threat: I cited earlier some of the experimental and historical data that suggest that probabilistic reasoning and falsification are not natural aspects of human thinking. In contrast, there is evidence that people can use explanatory coherence successfully in social judgments (Read and Marcus-Newhall, 1993).

One might argue that there is evidence that people are confirmation agents, and not very good ones in that they tend towards confirmation bias in looking excessively to confirm their hypotheses rather than falsify them (Klayman and Ha, 1987). However, the psychological experiments that find confirmation biases involve reasoning tasks that are much simpler than those performed by actual scientists. Typically, non-scientific subjects are asked to form generalizations from observable data, for example in seeing patterns in numerical sequences. The generalization tasks of real scientists are more complex, in that data interpretation requires determining whether apparent patterns in the data are real or just artifacts of the experimental design. If scientists did not try hard to get their experiments to confirm their hypotheses, the experiments would rarely turn out to be interesting. Notably, trying hard to confirm is not always sufficient to produce confirming results, so scientists sometimes have falsification thrust upon them. But their bias toward finding confirmations is not inherently destructive to scientific rationality.

A more serious challenge to the rationality of science comes from hot cognition. Like all people, scientists are emotional beings, and their emotions may lead to distortions in their scientific works if they are attached to values that are inimical to the legitimate aims of science. Here are some kinds of cases where emotions have distorted scientific practice:
1. Scientists sometimes advance their own careers by fabricating or distorting data in order to support their own hypotheses. In such cases, they have greater motivation to enhance their own careers than to pursue truth, explanation, or welfare.
2. Scientists sometimes block the publication of theories that challenge their own by fabricating problems with submitted articles or grant proposals that they have been asked to review.
3. Without being fraudulent or intentionally evil, scientists sometimes unintentionally deceive themselves into thinking that their hypotheses and data are better than those of their rivals.
4. Scientists sometimes further their careers by going along with politically mandated views, for example the Nazi rejection of Einsteinian physics and the Soviet advocacy of Lysenko's genetic theories.
Cases like these show indubitably that science is not always rational. Some sociologists such as Latour (1987) have depicted scientists as largely concerned with gaining power through the mobilization of allies and resources.

It is important to recognize, however, that the natural emotionality of scientists is not in itself a cause of irrationality. As I documented elsewhere, scientists are often motivated by emotions that further the goals of science, such as curiosity, the joy of discovery, and appreciation of the beauty of highly coherent theories (Thagard, forthcoming-b). Given the modest incentive structure of science, a passion for finding things out is a much more powerful motivator of the intense work required for scientific success than are extrinsic rewards such as money and fame. Thus hot cognition can promote scientific rationality, not just deviations from it. The mobilization of resources and allies can be in the direct or indirect service of the aims of science, not just the personal aims of individual scientists.

A useful response to the question "Is science rational?" is: "Compared to what?" Are scientists as individuals more adept than non-scientists at fostering truth, explanation, and human welfare? The history of science and technology over the past two hundred years strongly suggests that the answer is yes. We have acquired very broadly explanatory theories such as electromagnetism, relativity, quantum theory, evolution, germ theory, and genetics. Thousands of scientific journals constitute an astonishing accumulation of truths that ordinary life would never have allowed. Moreover, technologies such as electronics and pharmaceuticals have enriched and lengthened human lives. So the occasional irrationality of individual scientists and groups is compatible with an overall judgment that science is in general a highly rational enterprise.

In recent decades, the most aggressive challenge to the ideal of scientists as rational agents has come from sociologists and historians who claim that scientific knowledge is "socially constructed." Obviously, the development of scientific knowledge is a social as well as an individual process, but the social construction thesis is usually intended to make the much stronger claim that truth and rationality have nothing to do with the development of science. My own view is that an integrated psychological/sociological view of the development of scientific knowledge is perfectly compatible with scientific rationality involving the frequently successful pursuit of truth, explanation, and human welfare (Thagard, 1999).

Crucially, however, the assessment of scientific rationality needs to employ models of individual reasoning and group practices that reflect the thought processes and methodologies of real scientists. Models based on formal logic and probability theory have tended to be so remote from scientific practice that they encourage the inference that scientists are irrational. In contrast, psychologically realistic models based on explanatory and emotional coherence, along with socially realistic models of consensus, can help to illuminate the often impressive rationality of the enterprise of science.

Alvarez, W. (1998). T. rex and the crater of doom. New York: Vintage.
Carnap, R. (1950). Logical foundations of probability. Chicago: University of Chicago Press.
Earman, J. (1992). Bayes or bust? Cambridge, MA: MIT Press.
Glymour, C. (1980). Theory and evidence. Princeton: Princeton University Press.
Goldman, A. (1999). Knowledge in a social world. Oxford: Oxford University Press.
Hardin, G. (1968). The tragedy of the commons. Science, 162, 1243-1248.
Harman, G. (1986). Change in view: Principles of reasoning. Cambridge, MA: MIT Press/Bradford Books.
Hatfield, E., Cacioppo, J. T., & Rapson, R. L. (1994). Emotional contagion. Cambridge: Cambridge University Press.
Hempel, C. G. (1965). Aspects of scientific explanation. New York: The Free Press.
Howson, C., & Urbach, P. (1989). Scientific reasoning: The Bayesian tradition. Lasalle, IL: Open Court.
Hull, D. (1989). Science as a process. Chicago: University of Chicago Press.
Kahneman, D., Slovic, P., & Tversky, A. (1982). Judgment under uncertainty: Heuristics and biases. New York: Cambridge University Press.
Kitcher, P. (1993). The advancement of science. Oxford: Oxford University Press.
Klayman, J., & Ha, Y. (1987). Confirmation, disconfirmation, and information in hypothesis testing. Psychological Review, 94, 211-228.
Kuhn, T. (1970). Structure of scientific revolutions ( 2 ed.). Chicago: University of Chicago Press.
Lakatos, I. (1970). Falsification and the methodology of scientific research programs. In I. Lakatos & A. Musgrave (Eds.), Criticism and the growth of knowledge (pp. 91-195). Cambridge: Cambridge University Press.
Latour, B. (1987). Science in action: How to follow scientists and engineers through society. Cambridge, MA: Harvard University Press.
Lipton, P. (1991). Inference to the best explanation. London: Routledge.
Maher, P. (1993). Betting on theories. Cambridge: Cambridge University Press.
Mayo, D. (1996). Error and the growth of experimental knowledge. Chicago: University of Chicago Press.
Minsky, M. (2001). The emotion machine.: http://www.media.mit.edu/people/minsky/.
Nussbaum, M. (2001). Upheavals of thought. Cambridge: Cambridge University Press.
Pearl, J. (1988). Probabilistic reasoning in intelligent systems. San Mateo: Morgan Kaufman.
Popper, K. (1959). The logic of scientific discovery. London: Hutchinson.
Psillos, S. (1999). Scientific realism: How science tracks the truth. London: Routledge.
Read, S., & Marcus-Newhall, A. (1993). The role of explanatory coherence in the construction of social explanations. Journal of Personality and Social Psychology, 65, 429-447.
Salmon, W. (1984). Scientific explanation and the causal structure of the world. Princeton: Princeton University Press.
Solomon, M. (2001). Social empiricism. Cambridge, MA: MIT Press.
Thagard, P. (1988). Computational philosophy of science. Cambridge, MA: MIT Press/Bradford Books.
Thagard, P. (1991). The dinosaur debate: Explanatory coherence and the problem of competing hypotheses. In J. Pollock & R. Cummins (Eds.), Philosophy and AI: Essays at the Interface. (pp. 279-300). Cambridge, Mass.: MIT Press/Bradford Books.
Thagard, P. (1992). Conceptual revolutions. Princeton: Princeton University Press.
Thagard, P. (1999). How scientists explain disease. Princeton: Princeton University Press.
Thagard, P. (2000). Coherence in thought and action. Cambridge, MA: MIT Press.
Thagard, P. (2001). How to make decisions: Coherence, emotion, and practical inference. In E. Millgram (Ed.), Varieties of practical inference (pp. 355-371). Cambridge, MA: MIT Press.
Thagard, P. (forthcoming-a). Curing cancer? Patrick Lee's path to the reovirus treatment. International Studies in the Philosophy of Science.
Thagard, P. (forthcoming-b). The passionate scientist: Emotion in scientific cognition. In P. Carruthers & S. Stich & M. Siegal (Eds.), The cognitive basis of science. Cambridge: Cambridge University Press.
Thagard, P., & Shelley, C. P. (2001). Emotional analogies and analogical inference. In D. Gentner & K. H. Holyoak & B. K. Kokinov (Eds.), The analogical mind: Perspectives from cognitive science (pp. 335-362). Cambridge, MA: MIT Press.
Tversky, A., & Koehler, D. J. (1994). Support theory: A nonextensional representation of subjective probability. Psychological Review, 101, 547-567.
van Fraassen, B. (1980). The scientific image. Oxford: Clarendon Press.

Back to science and medicine articles.

Back to emotion articles table of contents.

Back to Computational Epistemology Laboratory.