THE LIMITS OF RATIONALITY*

Patrick SUPPES
Stanford University
In this lecture I survey a number of approaches to rationality, but
emphasize throughout the expected-utility or Bayesian model as the
most important and developed theory of rationality proposed in
recent years. Because the problems of group rationality as reflected
in our laws and customs have an importance equal to problems of
individual rationality, I move back and forth from the individual to
society, and I begin with one of my favorite historical examples of
bureaucracy out of control.
The passion for bureaucratic regulation of human activity is
nowhere better exemplified, than in the sumptuary laws repeatedly
laid down and partially enforced in ancient Rome and later in
Europe from medieval to early modern times. The Didean law of 143
B.C. of the Roman Republic limited the number of guests and the
amount of food that could be served in a private residence. In 1294,
Philippe le Bel of France restricted family meals to two dishes plus
soup for dinner. In the 14th century, Edward III of England imposed
sumptuary legislation on the food that was to be given servants. In
- 15 th-century Florence, three courses was the maximum permitted at
private dinners; two at supper. Police had the right of inspection of
banqueting halls at any time to be certain that the sumptuary laws
were being obeyed.
I n 16thcentury Venice, pheasant, peacock, guinea fowl, black-
cock, and woodcock were all banned, along with trout and other
freshwater fish. Repeated edicts in the 16th century all imposed re-
straint on Parisian eating habits, so that even at festive occasions only
three courses were supposed to be served. A host who violated the
*An earlier draft of this lecture was first given as the Howison Lecture in Phi-
losophy, University of California, Berkeley, February 22, 1979.
Grazer Philosophische Studien, 12/13, 1981, pp. 85-101.
84
edicts was liable to fine; his guests and the cook as well could be
87
selves are for the purpose of achieving an end in view. This Aristotelian
view might be characterized as a kind of qualitative dynamics of ra-
tionality. The good reasons play the role of causes or of forces in
mechanics. There is an ambiguity in this view as to whether good
reasons can be given for ultimate ends, but this rather delicate point
need not concern us now. Indeed, my objective is not to refine or to
bring together these two approaches to rationality but rather to
argue that we have not recognized sufficiently the extent to which
either is subject to severe limitations. Except in the most limited and
usually trivial situations is it possible to say that an action is perfect-
ly rational or that a man has acted in a perfectly rational way. We
can, with Thomas Aquinas, speak of the perfection of God but not
of man.
Other philosophical senses of being rational are set aside as not
relevant to the present focus. For example, rationalism as the view
that reason rather than the senses provides the proper foundation of
knowledge is irrelevant. Thus, historically what has been known as
the rational school of physicians - being rationalists in the sense
just defined - would in the sense of interest here be regarded as irra-
tional because of their refusal to use empirical evidence. In a similar
way I am not here concerned with the cntrast between rational and
divine theology. Nor am I concerned with rational as a form of ab-
straction as in that glorious subject of 18th-century natural philos-
ophy, rational mechanics.
I. The Market Model
- Returning to the forms of rationality of interest here, one important
response to the sumptuary laws of the past is that the rational ap-
proach would have been to give free play to market forces to deter-
- mine the distribution of food. For many economists, the perfection
of God is nearly matched by the perfection of markets, which often
seem to be thought of as an instrument of God. In the broader philo-
sophical tradition of analyzing the concept of rationality, it may
seem strange to speak of markets as a model of rationality, but,
especially now, strong claims are made, either implicit or explicit, for
the use of market mechanisms as a rational approach to problems of
distribution and therefore as the appropriate rational approach to
distributive justice.
Although the analysis is, I think, obvious, it may be useful to
88
ed as a mechanism for dealing with limited rationality on the part of
buyers and sellers. The complete exchange of information necessary
for explicit rational calculation, it may be argued, is a fantasy of
rationality that in its worst extremes is widely recognized. Even in
the most rigid Communist countries, for example, some kind of con-
sumer market exists in the sense that centralized decisions are not
made on precisely what is to be issued in terms of food, clothing, and
shelter to each consumer. The necessity for some pocket money
has been recognized from time immemorial, even by the most regi-
mented and egalitarian of armies.
The important feature of the market model is that it provides a
practical alternative to any attempts at complete rationality in the
distribution of goods and services by a state or corporate bureau-
cracy.
II. Expected Utility Model
The theory of markets as developed by Adam Smith, David Ricardo,
and others was the greatest early intellectual accomplishment of the
social sciences. An explicit concept of utility on which to build a
quantitative demand theory of the market came somewhat later in
the 19th century in the work of Gossen, J evons, Walras, and others,
but they made the overly strong assumption that utility was additive,
in the Same way that height or weight is. At the next stage this as-
sumption was weakened by Edgeworth, Antonelli, and I. Fisher to
the concept of an indifference surface for commodity bundles. Early
in this century, Pareto took the step of proposing a theory of
demand in terms of the behavioristic concept of purely ordinal pref-
erences for commodity bundles.
Contemporary expected utility theory goes beyond the ideas of
Pareto in two essential respects. First, preferences are not restricted
to commodity bundles, but to any field of choice about which deci-
sions are to be made. And, second, the theory is set up explicitly to
deal with uncertainty. Ordinal preferences are sufficient in general
only in situations of essentially complete knowledge of the relevant
possible states of affairs. (These extensions beyond Pareto actually
have an independent history that goes back to the 18th century.)
The intuition back of the expected utility mode is one that is
widely accepted. When, as individuals, we are forced to make deci-
sions, especially consequential ones, we must deal with two main fac-
90
tors: first, our beliefs about what is going to happen and, second, t
e the ideas in so
a decision abou
91
tant to us in almost all circumstances than what happens during the
coming year. Thus, whatever the patients decision, we can in a parti-
cular case apply the expected utility model to give an analysis that
shows that the patients choices were in the particular instance ra-
tional under some plausible set of assumptions.
As the last remark hints at, there are various ways of applying the
expected utility model. I want to say something about each of these
ways, which I describe as the biological, the psychological, and the
normative model.
Biological Model
The biological approach is to regard the model as part of the compu-
tational machinery of the organism. Instantaneously, the model is
always satisfied in the actions taken by the individual, and without
difficulty the model could be extended to the behavior of other
mammals and even lower in the phylogenetic scale if desired. This
biological model has consequences that can be investigated, but its
use in the theory of rationality is limited, because if we only require
instantaneous satisfiability, then enough parameters can be adjusted
to fit the data of any momentary choice.
Psychological Model
A more interesting approach, and one that has a fairly substantial
empirical history by now, is the application of the expected utility
model to the actual behavior of individuals in a variety of situations.
I n this case, the model is tested by assuming some constancy over
time and thus reducing the parameters that can be adjusted. An indi-
vidual who is too inconstant in his behavior will, under this standard
set of assumptions, not satisfy the model and will be said not to be-
have in accordance with this rational canon. There is a problem that
has been repeatedly faced in the application of the model to actual
behavior. I t is that some theory of error is required, just as a theory
of error is required even for the best physical theories of atomic or
astronomical bodies. Introduction of this theory of error is a rather
complicated affair but seems essential to any serious application to
real behavior. It should be obvious, of course, that the introduction
of a theory of error itself represents a limit on the degree to which a
purely rational account can be given of an individuals behavior. In
fact, in ,our conceptual thinking about these matters there is not a
92
itly by the individual but are part
cious evaluation sf his beliefs and
93
consistency of computations required by the expected utility model
does not guarantee the exercise of judgment and wisdom in the tradi-
tional sense. The Aristotelian view that the rational man acts in ac-
cordance with good reasons is not a necessary part of the normative
model of expected utility, and it is far from clear how these two con-
cepts can be put together in satisfactory form.
There is still another fundamental difficulty with the normative
model as a complete theory of rationality. When weattempt to ex-
tend it from individual to social decision processes, we do not know
how to aggregate the individual utility models to make a rational
decision among alternatives for the group. The literature on these
matters is large, and the results are mainly negative. But a central fact
about most of the studies makes the situation more negative than the
negative results themselves seem to. This is that almost all studies,
from Arrows classical impossibility theorem onward, ignore con-
flicts of interest and strategic moves to reflect those conflicts. The
analysis of social decision procedures is ordinarily based on the as-
sumption that the problem is to assess in straightforward fashion
each plain mans desires and to choose a social policy in a manner
that best reflects differing individual preferences. But the politics of
the real world cast skepticism on this assumption. It seems to me
that there is good ground for thinking that when serious conflicts of
interest exist, the political negotiations that follow are both neces-
sary and, at the same time, follow no obvious model of rationality.
Negotiations between industry and labor, government and farmers -
choose whatever conflicting groups you want - end up with deci-
sions that are based on compromise, but the details of the compro-
mise are shaped by idiosyncratic features of negotiators, yesterdays
weather, and other random happenings. There is not, and in my view
never will be, any complete model of rationality that is at all satisfac-
tory for accounting for the results.
III. The Legal Model
Once problems of conflict are brought to the surface, it might be sug-
gested that the proper place to look for a model of rationality is in
the practice of the law. The judge, weighing justice in the balance, is
often thought of as the very model of the rational deliberator
reaching an impartial ,and objective judgment of the rights and in-
terests of the parties in conflict. It might even be claimed that the
94
95
Before arriving where I can observe the freeway and make a decision,
I estimate the probability as about a half that the freeway is crowded.
If it is crowded I will take the parallel regular street; if it is not, I will
take the freeway. I arrive on the scene, make an observation that the
freeway is crowded, and thus take the parallel regular street. In the
Bayesian view, it is natural to say that my observation that the free-
way is crowded is certain, and therefore that I assign a probability of
one to this observational evidence.
Although there are clear differences, this desirability of evidence
being certain is in the grand tradition of the theory of demonstration
in Aristotles Posterior Analytics, Descartes Discourse on Method,
and Kants Critique of Pure Reason - all elaborate fantasies, in my
judgment. No harm seems to be done by sticking with a simple ac-
count of certain events in the example of deciding which road to tra-
vel, and I would agree that there is much ordinary talk that uses in a
natural and easy way the notion of being certain about not only pres-
ent observed events but also future ones. For example, it does not
seem at all strange or strained to say that it is certain I will not be
going to New York this week and therefore cannot meet someone
who calls about a conference to be held there.
Nonetheless, for either the theory or practice of complex deci-
sions, the search for certainty is one of the disastrous romantic fea-
tures of both ancient and modern thought. Whether we are deciding
what measures to take to control inflation, what the effects of low-
level radiation are, or what political alliances to make for tomorrows
security, we operate in an uncertain world, which cannot be brought
- to a point of certainty, and moreover, much of the evidence that we
deal in is equally far from certain and is dubious in character. To
some, it sounds like an alliance with the devil to speak of using du-
c bious evidence, but dubiety has an honored place in the repertoire of
any realistic decision maker. The technical ways of conditioning on
uncertain evidence are important if for neo-Bayesians, but I will not
pursue the matter further here.
Criticism of Completeness
Implicit claims of completeness are very much a part of the standard
Bayesian or expected utility model of rationality. The conceptual
machinery is there to deal with any problem, and it satisfies a canon
of completeness in the sense that it gives an answer in principle,
96
97
must be made. There is no serious evidence that this set of decisions
can be reduced to an explicit theory of rationality. Many practical
and conceptually inexplicit decisions must be made in every area of
experimentation, or indeed of empirical inquiry, ranging from high-
energy physics to the psychology of auditory judgments.
The second point is that it is relatively straightforward to think
about applying a criterion of rationality to decisions formulated in a
theoretical framework. This reflects a general scientific fact of life. In
every branch of science, the theoretical language is relatively simple
and easy to learn compared with the complex, arcane, and esoteric
language used to express the lore of experimental procedures, from
the calibration of instruments to the collection of data. I will not
here attempt to document this thesis. I , in fact, consider it obvious
to anyone acquainted with actual scientific practice in some devel-
oped branch of investigation. Decisions taken in the framework of
such highly empirical language and thought routinely seem too
highly context-dependent to be assessed in terms of any general
criterion of rationality.
The third point is that the fantasy of explicit formality, with
elimination of all need for intuitive judgment, is recognized as an
outmoded concept, even in mathematics. Hilberts program of for-
malism was upset years ago by Gdels incompleteness theorems, and
the gap between formal theories of proof, for example, and actual
mathematical practice, requiring judgments of all kinds, is now
widely acknowledged. I emphasize the essential place of judgment in
science and mathematics, because this is the area of thought and ex-
perience that seems most amenable to regimentation by models of ra-
tionality.
The essential place of judgment in decision making has been rec-
ognized by various people, including de Finetti, for some time. But
there has been a countertendency, generated by the thicket of tech-
nical results about decision making in modern mathematical statis-
tics, that tends to suggest that judgment can be eliminated in favor of
the application, in a routine and objective and possibly even algorith-
mic way, of technical results. There has also been a pernicious trend
of long standing in classical philosophy of both ancient and modern
times that reenforces this view - perhaps the best example is the
foolish attempt to construct our knowledge of the world from indu-
bitable sense data. What I have said about the essential role of intui-
98
tive judgment in assessing evidence in scientific experiments extends
99
practice - and in most cases explicit teaching by an experienced
coach. We are, for a variety of reasons, much more romantic and
unsystematic about learning mental skills.
To a considerable extent, however, the area of agreement appears
to be greater than that of difference. A more important point is one
that cuts across the division between mental and physical skills. This
is whether or not there is a simple and agreed-upon objective crite-
rion of performance. If we are training a marathon runner or a com-
petitor for the 1000-meter race, we can, by measuring his running
time each day, have an excellent simple and objective criterion of
performance and improvement in performance. In the case of com-
petitive games, the criterion is more complicated but still in the end
rather simple because of our ability to tabulate who won and who
lost when real play starts. The same can be said for mental skills that
involve objective tasks like those of arithmetical computation. In the
case of 'physical skills like that of mountain climbing, or mental skills
such as those of judging the qualities of complex objects, for ex-
ample, paintings, horses, or beautiful people, there is absent agreed-
upon, simple, and easily applicable criteria of performance. All the
same, the absence of such simplicity does not prohibit highly success-
ful training and teaching; this general observation applies whether we
are talking about physical skills like mountain climbing or mental
skills like judging works of art of a given category.
I emphasize also my optimism about developing a proper psychol-
ogy of such skills, in particular, a proper educational psychology for
teaching such skills. The important reservation to this optimism is
that the teaching will not be verbally explicit; it will not be algorith-
mic in the sense of providing the student an algorithm for reaching
the intended result. Rather, it will depend upon the kinds of compli-
cated indirect and subtle methods already familiar in the training of
many different physical and mental skills. The long tradition of ap-
prenticeship in training craftsmen is one of the finest examples. This
tradition, it should be noted, has flourished in modern scientific
laboratories.
It is my own view that, because of the overly rationalistic and lin-
guistically oriented approach to intuitive judgment in the past, we
have not developed in any adequate way the kinds of effective
teaching regimes that are possible. The first step is to understand the
nature of intuitive judgment - that it is a skill more like running
101
there is not even a theoretical exact value to be measured and thus
there is no hope even theoretically of refining the approximations to
an exact value, which was one of the mistaken concepts of classical
physics.
The utility or loss functions of the statisticians, which are <at the
heart of the expected utility model, are designed for the factory and
the marketplace. As we move to tea and crumpets and philosophical
delicacies in the drawing room, conceptual refinements will be neces-
sary. But with the Aristotelian tradition to draw on, we should be
able to follow the proper golden way of calculation and judgment in
undogmatic splendor. What we teach our students or ourselves about
1 practical decision making cannot be reduced to algorithms or even to
explicit axioms, but we can, on the other hand, improve on Aristotle
just because we can apply axioms and algorithms of rationality as ap-
propriate. The use of modem quantitative methods of decision mak-
ing is necessarily limited but powerful, all the same, when properly
applied. The role of judgment and practical wisdom in applying these
methods will continue to be of central importance, and so, con-
sequently, will be continual philosophical analysis of them.
9