Evolutionary Psychology and the

Generation of Culture, Part II

Case Study: A Computational Theory of Social
Exchange

Leda Cosmides and John Tooby

Department of Psychology, Stanford University, Stanford, California

Models of the various adaptive specializations that have evolved in the human psyche
could become the building blocks of a scientific theory of culture. The first step in
creating such models is the derivation of a so-called “computational theory” of the
adaptive problem each psychological specialization has evolved to solve. In Part II, as
a case study, a sketch of a computational theory of social exchange (cooperation for
mutual benefit) is developed. The dynamics of natural selection in Pleistocene ecological
conditions define adaptive information processing problems that humans must be able
to solve in order to participate in social exchange: individual recognition, memory for
one’s history of interaction, value communication, value modeling, and a shared gram-
mar of social contracts that specifies representational structure and inferential pro-
cedures. The nature of these adaptive information processing problems places con-
straints on the class of cognitive programs capable of solving them; this allows one to
make empirical predictions about how the cognitive processes involved in attention,
communication, memory, learning, and reasoning are mobilized in situations of social
exchange. Once the cognitive programs specialized for regulating social exchange are
mapped, the variation and invariances in social exchange within and between cultures
can be meaningfully discussed.

KEY WORDS: Reciprocal Altruism; Cooperation; Tit for tat; Cognition; Reasoning;
Evolution; Learning; Culture.

INTRODUCTION

H
uman beings live in groups, and their behavior is affected by
information derived from the other individuals with whom they
live. The study of culture is the study of how different kinds of
information from each individual’s environment, especially from

Received April 1, 1987; revised October 17, 1987.

Address reprint requests to: Leda Cosmides, Department of Psychology, Bldg. 420, Stanford
University, Stanford, CA 94305.

Ethology and Sociobiology 10: 51-97 (1989)

0 Elsevier Science Publishing Co., Inc., 1989 0162-3095/89/$3.50
655 Avenue of the Americas, New York, NY 10010
52 L. Cosmides and J. Tooby

his or her social environment, can be expected to affect that individual’s

behavior. The behavior elicited by this information reverberates throughout
the individual’s social group, as information that other individuals may act
on in turn. The ongoing cycle that results is the generation of culture. By
directly regulating individual learning and behavior, those psychological
mechanisms that select and process information from the individual’s social
environment govern the resulting cultural dynamics. The key to understand-
ing cultural processes therefore lies in the discovery, and subsequent map-
ping, of the properties of these innate information processing mechanisms,
an enterprise now feasible using an approach that is both evolutionary and
cognitive (Tooby and Cosmides this issue).
Evolutionary biology provides a powerful heuristic for guiding psycho-
logical research. This heuristic rests on the recognition that psychological
mechanisms evolved as responses to selection pressures. The more impor-
tant the adaptive problem, the more intensely selection will have specialized
and improved the performance of these mechanisms. Some of these mech-
anisms evolved to meet the adaptive problem of social exchange. We will
use the case of social exchange to illustrate the method of evolutionary
psychology discussed in Part I of “Evolutionary Psychology and the Gen-
eration of Culture” (Tooby and Cosmides this issue).
Social exchange-cooperation between two or more individuals for mu-
tual benefit-is a pervasive aspect of all human cultures. It manifests itself
in many different forms within and across cultures: ritualized gift giving, the
exchange of favors between friends, trade in a market economy, and barter
are all forms of social exchange. Successfully conducted social exchange
was a critically important feature of hominid evolution (Tooby and DeVore
1987). The ability to successfully engage in social exchange depends on the
structured processing of specific kinds of information. This is because nat-
ural selection permits the evolution of only certain strategies for engaging
in social exchange: Some strategies are evolutionarily stable (Maynard Smith
1982), others are not. By studying the nature of these strategies, one can
deduce many properties that the human information processing algorithms
regulating social exchange must have, as well as much about the associated
capabilities such algorithms require to function properly. Using this frame-
work, one can then make empirical predictions about human performance
in areas that are the traditional concern of cognitive psychologists: attention,
communication, the organization of memory, learning, and reasoning.
By examining of the nature of the selective pressures on social exchange
in human evolution, some things can be inferred about the psychological
basis for social exchange in humans. These inferences can be used to con-
struct a computational theory of social exchange: a theory defining the in-
formation processing problems that a human engaging in social exchange
must be able to solve. A psychological mechanism is a solution to an in-
formation processing problem. Thus, a computational theory provides a pre-
 A Computational Theory of Social Exchange 53

dictive framework that facilitates the design of experiments that can map
the structure of the cognitive programs that guide social exchange in humans.
This article is divided into four sections:

Part 1. Only certain strategies for engaging in social exchange can evolve:
Natural selection’s game theoretic structure defines what properties these
strategies must have.
Part 2. The ecological conditions necessary for the evolution of social ex-
change were manifest during hominid evolution; hominid behavioral ecol-
ogy further constrains a computational theory of social exchange.
Part 3. These strategic and ecological constraints define a set of information
processing problems that must be solved by any human engaging in social
exchange. Computational theories of these problems are developed.
Part 4. Aspects of the computational theory of social exchange have been
tested. Experimental evidence from tests of logical reasoning verifies the
existence of algorithms for detecting cheaters; as predicted, these algo-
rithms operate on item-independent, cost-benefit representations of ex-
change interactions.
Parts 1 and 2 review the constraints from which a computational theory
should be built. Part 3 presents a first attempt to build a computational theory
of social exchange; Part 4 briefly reviews the results of experiments designed
to test aspects of the computational theory developed.

PART 1. NATURAL SELECTION AND SOCIAL

EXCHANGE

The critical act in formulating computational theories turns out to be the

discovery of valid constraints on the way the world is structured. . . . (Mar-r
and Nishihara 1978)
There are laws inherent in the dynamics of natural selection that should
shape every species. Many of these laws govern the evolution of social
behavior; they constrain the kinds of social behavior that can evolve under
a given set of circumstances. Although this is common knowledge among
evolutionary biologists, it is something psychologists need to incorporate
into their methodology: knowledge of natural selection constitutes knowl-
edge of a set of valid constraints on what kinds of mental mechanisms could
have evolved and on the properties they have.
Traits can be thought of as the embodiment of strategies for the prop-
agation of the genes that code for them. By analyzing the dynamics of gene
flow through populations, one can determine what kinds of traits will be
quickly selected out and what kinds of traits are likely to become universal
and species-typical. Formally, this analysis can be cast in terms of game
theory: One strategy is pitted against others to see which ones come to
54 L. Cosmides and J. Tooby

dominate the gene pool (Maynard Smith 1982). During the last 20 years,
game-theoretic models of the dynamics of natural selection have proliferated
in evolutionary biology. The elaboration of these methods now allows a more
precise characterization of the differences between those strategies that can
be selected for and those that will be selected against (e.g., Hamilton 1964;
Williams 1966; Maynard Smith 1982; Dawkins 1982). In the case of social
exchange (“cooperation” or “reciprocation”), Axelrod and Hamilton (1981)
and Axelrod (1984) have shown that only certain families of strategies with
certain distinctive properties can evolve.
Using the iterated Prisoner’s Dilemma as their paradigm of coopera-
tion,’ Axelrod and Hamilton (1981), and Axelrod (1984), following Williams
(1966) and Trivet-s (1971), explored the envelope of conditions limiting the
evolution of social exchange. It is the possibility of cheating (or defecting
once one side of a mutually beneficial exchange has been carried out) that
makes the evolution of cooperation difficult. For this reason, indiscriminate
cooperation under conditions that allow cheating is an unstable strategy that
would be quickly selected out under all models of biologically plausible
conditions. Virtually any nonsimultaneous exchange creates a possibility of
defection, and most “natural” opportunities for exchange are not simul-
taneous. For example:
1. A common “item” of exchange between primates is protection from con-
specifics or predators. Two or more individuals develop coalitional re-
lationships for mutual defense, aggression, or protection (e.g., baboons:
Hall and DeVore 1965; chimpanzees: Wrangham 1986; de Waal 1982). If
one individual is attacked, and another comes to its defense, there is
nothing the aided individual can do at that time to repay his rescuer.
Reciprocation is possible only at another time, when the rescuer is himself
attacked.
2. Interactants are foraging for patchy resources. One individual finds, for
example, a patch containing more than can be easily eaten by itself, and
gives a call to guide others to the patch (or returns and shares the resource
with others). Repayment in kind, to be valued, would have to take place
subsequently (e.g., chimpanzees: Goodall 1968, 1971; bats: McCracken
and Bradbury 1981; Wilkinson 1984).
3. In hunter-gatherer meat sharing, kills may be larger than can be easily
consumed by those who were directly in the hunt, and only irregularly
obtained. The value of consuming the whole kill is less than the value of
sharing the unneeded or less-needed portions with others, provided that
the act is reciprocated at some future time (Lee and DeVore 1968). Again,
repayment on the spot is both unlikely and not valuable.

t Other models of social exchange are possible, but they will not change the basic conclusion
of this section: that reciprocation is necessary for the evolution of social exchange. For example,
the Prisoner’s Dilemma assumes that enforceable threats and enforceable contracts are im-
possibilities (Axelrod 1984), assumptions that are frequently violated in nature. The introduction
of these factors would not obviate reciprocation-in fact, they would enforce it.
 A Computational Theory of Social Exchange 55

4. There is mounting evidence that a baboon male forms “special relation-

ships” with a few lactating (and therefore infertile) females and their
infants: He protects them from conspecifics and predators, and may sub-
sequently obtain sexual access when the females wean their infants and
become fertile again (e.g., Smuts 1982; Strum 1985). If this does constitute
an exchange, the male’s repayment, by necessity, comes at a much later
time.

The opportunity for on-the-spot repayment or withdrawal of benefits in

the face of cheating is rare in nature for several reasons:

The “items” of exchange are frequently acts that, once done, cannot be
undone (e.g., protection from attack, alerting others to the presence of
a food source).
Opportunities for simultaneous mutual aid are rare because the needs and
abilities of organisms are rarely exactly and simultaneously complemen-
tary: The female baboon is not fertile when her infant needs protection,
yet this is when the male’s ability to protect is of most value to her.
On those occasions when repayment could be made simultaneously and
in the same currency, declining marginal utilities makes the exchange
senseless: If meat sharers both make a kill on the same day, neither
benefits from the other’s windfall.

Thus, in the absence of a widely accepted medium of exchange,2 most ex-

changes do involve opportunities for defection.
A system of mutual cooperation cannot emerge in a one move Prisoner’s
Dilemma, because it is always in the interest of each player to defect (Lute
and Raiffa 1957; see Fig. 1 for explanation). In fact, the argument is general
to any known, fixed number of games (Lute and Raiffa 1957). However,
selection pressures change radically when individuals play a series of Pris-
oner’s Dilemma games. Mutual cooperation-and therefore social ex-
change-can emerge between two players when 1) there is a high probability
that they will meet again, 2) neither knows for sure exactly how many times
they will meet,3 and 3) they do not value later payoffs by too much less than
earlier payoffs (Axelrod and Hamilton 1981; Axelrod 1984). If the parties
are making a series of moves rather than just one, one party’s behavior on
one move can influence the other’s behavior on future moves. (For ease of
explication, in the discussions of social exchange throughout this paper, the

’ Indeed, such factors are exactly why it is so useful to have a medium of exchange. One party
doesn’t have to be able to provide the particular goods or services the other party wants because
money can be converted into anything others are willing to exchange for it. Furthermore, money
permits a simultaneous exchange, in which either party can, in fact, withhold the money if he
or she suspects that the other is attempting to cheat.
s The game “unravels” if they do. If we both know we are playing three games, then we both
know we will mutually defect on the last game. In practice, then, our second game is our last
game. But we know that we will, therefore, mutually defect on that game, so, in practice, we
are playing only one game. The argument is general to any known, fixed number of games
(Lute and Raiffa 1957).
56 L. Cosmides and J. Tooby

cooperate defect

me: R = B(me) - C(me) me: S = C(me)

cooperate
you: R = B(you) - C(you) you: T = B(you)

me

me: T = B(me) me: P = O(me)

defect
you: s = C(you) you: P = O(you)

FIGURE 1. Social exchange sets up a Prisoner’s Dilemma: Payoff Schedule.

B(X) = Benefit to X; C(X) = Cost to X; O(x) = X’s inclusive fitness is un-
changed. R = Reward for mutual cooperation; T = Temptation to defect; S =
Sucker’s payoff; P = Punishment for mutual defection. T > R > P > S; R > (T +
S)/2 (for an iterated game; this prevents players from “cooperating” to maximize
their utility by alternately defecting on one another. C(me) need not equal C(you),
and B(me) need not equal B(you); an exchange will have the structure of a Prisoner’s
Dilemma as long as mutual cooperation would produce a net benefit for both players.
Let us assume that you and I are playing a one move Prisoner’s Dilemma game. I
would reason thus: “You will either cooperate or defect. If you cooperate, then I
get a higher payoff by defecting, because T, the Temptation to defect, is greater than
R, the reward I would get for mutual cooperation. If you defect, then I get a higher
payoff by also defecting, because P, the payoff I receive as a Punishment for mutual
defection, is greater than S, the Sucker’s payoff I would get if I cooperate and you
defect. Therefore, no matter what you do, I am better off defecting.” Your reasoning
process would be identical, so we would both defect, and we would both get P, the
Punishment for mutual defection. If you cooperate, I get B(me) for defecting instead
of &me) - C(me) for cooperating. If you defect, I lose nothing by defecting instead
of losing C(me) by cooperating.

two interactants in the social exchange will be designated “you” and “I,”
with appropriate possessive If “I” defect when “you” coop-
pronouns).
erated, then you can retaliate by defecting on the next move;4 if I cooperate,
you can reward me by cooperating on the next move. In an iterated Pris-
oner’s Dilemma, a system can emerge that has incentives for cooperation
and disincentives for defection.
The work of Trivers (1971), Axelrod and Hamilton (1981), and Axelrod
(1984), has shown that indiscriminate cooperation cannot be selected for
when the opportunity for cheating exists. A cooperative strategy can invade

4 In nature, I can also retaliate by inflicting a cost on you through the use of violence. However,
if I can, reliably, do this, the game is no longer a Prisoner’s Dilemma. Violent retaliation is a
“tax” on defection that wipes out the incentive to defect (i.e., T minus R). If T s R, then the
situation no longer presents a dilemma-we both have an incentive to cooperate and no incentive
to cheat. The key word in the above scenario is reliably. From a “veil of ignorance” as to the
relative strength of two individuals, on average, half the time I (the cheated on) will be able to
inflict a cost on you, and half the time you (the cheater) will be able to inflict a cost on me. Of
course, most animals are not acting from a veil of ignorance, and one would expect them to
assess their relative strength and adjust their strategies accordingly.
 A Computational Theory of Social Exchange 57

a population of noncooperators if, and only if, it cooperates with other co-
operators and excludes (or retaliates against) cheaters. If a cognitive decision
rule regulating when one should cooperate and when one should cheat does
not instantiate this constraint, then it will be selected against. However,
Axelrod (1984) has shown that there are many decision rules that do in-
stantiate this constraint. Any of these could (other things being equal), there-
fore, have been selected for in humans; which decision rule, out of this
constrained family, actually evolved in the human lineage is an empirical
question. The most general statement about such decision rules that natural
selection theory permits is this: Humans have the ability to cooperate for
mutual benefit; this capacity could not have evolved unless it included al-
gorithms for detecting, and being provoked by, cheating.

PART 2. SOCIAL EXCHANGE AND THE PLEISTOCENE

ENVIRONMENT

The ecological conditions necessary for the evolution of social exchange

were manifest during hominid evolution; hominid behavioral ecology fur-
ther constrains a computational theory of social exchange.
Cooperation can evolve only when 1) there are many situations in which
individuals can benefit each other at relatively low cost to themselves (i.e.,
an iterated Prisoner’s Dilemma game is possible), and 2) the probability of
two individuals meeting again is sufficiently high.5 The probability that two
individuals will meet again is increased if the individuals are long-lived and
have low dispersal rates. These life-history factors also increase the number
of situations for mutual help that two individuals are likely to encounter.
The ecological and life-history factors characteristic of the human environ-
ment of evolutionary adaptedness fulfill the conditions necessary for the
evolution of cooperation. Pleistocene hunter-gatherers were not only long-
lived, but they lived in small, relatively stable bands. Thus, the probability
was high that an individual you had helped would be around when you needed
help. Moreover, in all probability these individuals, like modern hunter-
gatherers, were closely related; kin selection can promote the evolution of
cooperation (Trivers 1971; Axelrod and Hamilton 1981).
The intellectual capacities of hominids allowed them to generate many
situations for which cooperation paid off. The most important of these was
the capacity to make and use tools, and the capacity to generate novel be-
havioral procedures to achieve a goal. The expanded exploitation of the
Savannah and woodland niche-made possible by tool use-allowed indi-
viduals to acquire food items too large to be consumed by a single individual
(Isaac 1978; Tooby and DeVore 1987). This created effective opportunities

5 For example, TIT FOR TAT is an ESS if, and only if, the probability that two individuals
will meet again is greater than the larger of these two numbers: (T-R)I(T-P) and (T-R)/(R-S)
(Axelrod 1984).
58 L. Cosmides and J. Tooby

to provide large benefits to other individuals at a low cost to oneself. There

is virtually no cost to sharing food that you cannot consume anyway, and
tomorrow you may be the one who has found no food. Fossil evidence
suggests that Pleistocene hunter-gatherers, like their modern counterparts,
engaged in extensive food-sharing (e.g., Isaac 1978). Similarly, the cost of
sharing tools is low compared to the benefits one can garner through using
them-and the cost of sharing information about tool making may be even
lower.
When combined with their capacity to opportunistically manipulate the
environment through tool use, our ancestors’ ability to generate novel be-
havioral procedure@ created situations in which coordinated, cooperative
behavior could produce vast payoffs. Perhaps one of the best examples is
the “profits” to be made through cooperative hunting. Acting together, sev-
eral armed men can bring down a large game animal; acting alone, a single
armed man is more likely to fail.
These conditions set the stage for the coevolution of a tightly interwoven
complex of adaptations whose mutual reinforcement made cooperation more
and more advantageous (Tooby and DeVore 1987, and references therein).
Cooperative hunting provided a compact and nutritious food source that
provided an efficient means for males to invest in offspring; the increased
payoff of this pattern of male parental investment favored the restructuring
of mating relationships associated with paternity certainty, potentially mod-
erating intra-band male-male competition; such increased parental invest-
ment allows larger brains and longer periods for maturation and learning;
these, in turn, allow more efficient cooperation and tool use, which leads
to even more nutritious food sources from both hunting and gathering; these
developments allow the allocation of an increasing proportion of metabolic
resources to brain tissues. Such developments may be mutually reinforcing
because the evolution of sophisticated cooperation in a large array of ac-
tivities depends upon a complex cognitive base.
Reconstruction of the causal sequences that fed the evolution of co-
operation (and associated developments) is still a matter of debate (cf. Kin-
zey 1987). The most important point is that the Pleistocene hunter-gatherer
environment in which we evolved provided many opportunities for individ-
uals to benefit from mutual cooperation, and cooperation for mutual benefit
is a pervasive and inextricable aspect of all past and modern human cultures.
The peculiarities of hominid behavioral ecology place some species-
specific constraints on a computational theory of social exchange in humans.
Exchange in most primates is restricted to relatively few “items”: food,
sexual access, defense, grooming. The fewer the items for exchange, the

6 An ability that some other primates also possess, to a lesser extent. For example, de Waal
(1982) shows pictures of chimpanzees who have discovered that they can get past an electrified
fence surrounding a tree with edible leaves. One chimpansee holds a large branch against the
tree as a ladder, while another climbs it into the tree. The chimpanzee in the tree then throws
juicy leaves down to his compatriots on the ground.
 A Computational Theory of Social Exchange 59

more “item-specific” the algorithms regulating exchange can (and should)

be: What counts as “error” -cheating or underreciprocating-can be more
closely defined, increasing the accuracy of one’s mental accounting system
and the accuracy of reference (see Part 3). In contrast, because of the human
penetration of the “cognitive niche” (Tooby and DeVore 1987), human al-
gorithms for regulating social exchange should be able to handle a wide and
ever-changing array of “items” for exchange: tools, participation in coali-.
tional aggression, information about tool-making, participation in opportun-
istically created, coordinated behavioral routines. This suggests that our
algorithms for regulating social exchange, and the associated cognitive ca-
pacities that they require to function properly, will have some human-spe-
cific properties. These will be discussed in Part 3.

PART 3. A COMPUTATIONAL THEORY OF SOCIAL

EXCHANGE

David Mat-r has argued that the first and most important step in understand-
ing an information-processing problem is developing a “theory of the com-
putation” (Mar-r 1982; Marr and Nishihara 1978). This theory defines the
nature of the problem to be solved; in so doing, it allows one to predict
properties that any algorithm capable of solving the problem must have.
Computational theories incorporate “valid constraints on the way the world
is structured-constraints that provide sufficient information to allow the
processing to succeed” (Marr and Nishihara 1978, p. 41).
For humans, an evolved species, natural selection in a particular eco-
logical situation defines and constitutes “valid constraints on the way the
world is structured” for a particular adaptive information processing prob-
lem (Cosmides and Tooby 1987). In the case of social exchange, the eco-
logical and game-theoretic aspects of hominid social exchange discussed
above provide the ingredients for the construction of just such a computa-
tional theory. The ability to engage in a possible strategy of social exchange
presupposes the ability to solve a number of information processing prob-
lems that are highly specialized. The elucidation of these information pro-
cessing problems constitutes a computational theory of social exchange. Any
psychological theory purporting to account for the fact that humans are able
to engage in social exchange must be powerful enough to realize this com-
putational theory-that is, its information processing mechanisms must pro-
duce behavior that respects the constraints imposed by the selective process.
Thus, it must be powerful enough to 1) permit the realization of a “possible”
social exchange strategy, i.e., a strategy that can be selected for, and 2)
exclude “impossible” strategies, i.e., strategies would be selected against.
The problems most specific to social exchange will be incorporated into
a “grammar of social contracts” in the second half of Part 3. A grammar
of social contracts is the set of assumptions about the rules governing a par-
60 L. Cosmides and J. Tooby

titular social exchange that must somehow be incarnated in the psycholog-

ical mechanisms of both participants. However, the grammar of social con-
tracts does not exhaust the set of information processing problems posed
by social exchange. The ability to successfully participate in social exchange
also requires a number of other, associated cognitive capacities, some of
which are necessary in a wide range of other evolutionarily crucial social
interactions, such as mating, pair-bonding, parenting, and aggression. Before
we progress to the grammar of social contracts, five associated cognitive
capacities entailed by social exchange will be examined:
1. The ability to recognize many different individuals.
2. The ability to remember aspects of one’s history of interaction with dif-
ferent individuals.
3. The ability to communicate one’s values to others.
4. The ability to model the values of other individuals.
5. The ability to view as costs and benefits items that one perceives as
causally connected to biologically significant variables; human algorithms
regulating social exchange should not be too closely tied to particular
items of exchange.
Undoubtedly, a clever programmer could design many different algo-
rithms capable of solving these problems. It is even possible that one or two
of them could be solved, albeit less efficiently, by domain general mecha-
nisms such as associative nets. But to demonstrate that such mechanisms
could, in theory, solve these problems would be to miss the point. The point
of using natural selection theory in creating computational theories is that
it allows you to specify a set of problems that humans ought to be able to
solve quickly, reliably, efficiently, and without explicit instruction. These
are problems for which natural selection should have produced specialized,
domain specific Darwinian algorithms: modules in Marr’s (1982) or Fodor’s
(1983) terminology, mental organs or cognitive competences in Chomsky’s
(1975) terminology, adaptations in the terminology of evolutionary biology.
It is the presumption that natural selection has designed psychological mech-
anisms that are particularly good at solving these problems that carries im-
plications for the study of attention, communication, the organization of
memory, implicit inference, and learning. We shall briefly sketch a few of
these implications, together with some of the relevant data.

HUMAN SOCIAL EXCHANGE REQUIRES SOME

FUNDAMENTAL COGNITIVE CAPACITIES

Proposition 1. One Must be Able to Recognize Many Different

Individual Humans
The basic idea is that an individual must not be able to get away with
defecting without the other individuals being able to retaliate effectively.
 A Computational Theory of Social Exchange 61

The response requires that the defecting individual not be lost in a sea of
anonymous others. (Axelrod and Hamilton 1981)

Individual recognition is important even if one has an exchange relationship

with only one individual. It is that much more important if one has such
relationships with a number of individuals; the ability to cooperate with more
than one individual is particularly useful to a hunter-gatherer. In order to
cooperate only with individuals who are likely to reciprocate, and avoid (or
cheat on) individuals who are likely to cheat, one must be able to discriminate
different individuals.’ One need not rely on “preliminary hunches” (Carey
and Diamond 1980, p. 60) in singling out individual recognition as a domain
for which humans ought to have specialized mechanisms; it is a direct pre-
diction of an evolutionary perspective.
Indeed, humans do seem to have a highly developed ability to recognize
large numbers of different individuals. Recognition rates are over 90% for
familiar faces that have not been seen for up to 34 years (Bahrick et al. 1975).
Patients with a lesion in a specific part of the right hemisphere develop a
selective deficit in their ability to recognize faces, called prosopagnosia
(Gardner 1974). Carey and Diamond (1980) present and review an impressive
array of evidence from a wide variety of sources suggesting that humans
have innately specified face-encoding schemas. We are also good at iden-
tifying individuals solely through recognizing their idiosyncratic gaits (Cut-
ting et al. 1978; Kozlowski and Cutting 1977).

Proposition 2. One Must be Able to Remember Some Aspects

of the Histories of One’s Interactions with Different Individuals
First, one must be able to recognize that a previous interactant in a social
exchange is, in fact, a particular previous interactant, and not a stranger or
a misidentified individual. Second, once an individual has been identified as
a previous interactant, information regarding that individual’s particular his-
tory of previous interactions, coded in terms of whether that individual has
been a cooperator or a cheater, must become accessible to the decision
procedures. Third, one needs an “accounting system” for keeping track of
who owes who what. As discussed in Part 1, many Pleistocene social
exchanges involved “reciprocal altruism”-exchanges in which recipro-
cation was delayed, not simultaneous. In a simultaneous, face-to-face ex-
change, if you see that the other person has come prepared to defect, you
simply withhold what that person wants.8 However, the capacity for en-

’ Organisms that lack the ability to recognize different individuals can also evolve a limited
ability to cooperate, but only because of ecological restrictions on their interactions to a very
few partners with whom they are in constant and/or exclusive physical proximity (Axelrod and
Hamilton 1981).
* One would expect people to assume, in the absence of information to the contrary, that such
intercontingent behavior occurs in face-to-face interactions. They should be more likely to
suspect someone of intending to cheat in delayed benefit transactions.
62 L. Cosmides and J. Tooby

gaging in transactions in which reciprocation is delayed requires a mental

accounting system for keeping track of who owes who what (note: Propo-
sition 5, about item-independent representations, also applies to this ac-
counting system).
The extent of the history of interaction that must become available to
the decision procedures regulating participation in social exchange (and
whether any of these facts need be consciously recalled) will depend on the
details of the particular decision procedures humans have evolved. For ex-
ample, TIT FOR TAT (Axelrod and Hamilton 1981) requires only that the
last transaction with each interactant be recalled. But TIT FOR TAT is a
successful strategy in a highly constrained and uniform universe where all
transactions are simultaneous, the same payoff matrix applies to each trans-
action, and the size of the payoffs for both players is equal within each
transaction (Axelrod 1984). In contrast, payoff matrices in the real world
are always in flux, and part of that flux is caused by the negotiative skills
of the individuals involved. Moreover, violence is possible in the real world:
Exchange situations with individuals who can reliably use violence to get
their way do not necessarily tit the constraints of a Prisoner’s Dilemma
(Tooby and Cosmides, in preparation, a). Thus, an algorithm better adapted
to conditions in the real world might assess many more factors regarding
one’s past history with an individual, such as 1) the number of transactions
one has had with that individual in the past, 2) how he or she behaved in
those transactions (reputation), 3) the size of payoffs to both parties in pre-
vious transactions, 4) whether his or her tendency to cheat varied with the
size of the payoff involved, 5) whether the conditions governing his or her
tendency to cheat have been shifting over time, 6) his or her aggressive
formidability, 7) how likely one is to meet that individual in the future (e.g.,
one party is moving away or likely to die soon), and 8) whether one has
accepted a past benefit but has not reciprocated yet. Information regarding
others’ histories of reciprocation, including circumstances and dispositions
that might explain their occasional (or systematic) failures to reciprocate,
should be intensely interesting, as should information about how others re-
gard oneself. The power of single defections to be socially communicated
through reputation may provide an especially intense disincentive to cheat,
and hence may have allowed far more reliable systems of cooperation to
develop in humans than among species whose ability to communicate such
information is limited.
A decision procedure that used such data, current behavioral cues,9 and
the payoff matrix for the current interaction to compute the conditional

9 For example, a person’s facial expression might telegraph his or her intention to cheat. All
else equal, a person’s “likeability” should be a function of his or her tendency to reciprocate,
and cues that suggest “good intentions” ought to be judged more likeable (e.g., sneers and
aggressive scowls do not suggest good intent). Although other explanations are possible, it is
interesting that people remember unfamiliar faces better when, during initial inspection, they
are asked to judge the person’s likeability than when they are asked to assign sex (Carey and
Diamond 1980).
 A Computational Theory of Social Exchange 63

probability that one’s partner will cooperate, might be better adapted to the
complexities of exchange in nature. lo If so , then the need to take such factors
into account has implications regarding the organization of human memory.
Information about one’s history of interaction with a particular person ought
to be “filed” with that person’s “identity” and activated quickly and ef-
fortlessly when an exchange-relevant situation with that person arises. When
the payoff matrix of the current interaction is such that “you” will lose a
great deal if “I” cheat you, then more of our past exchange history should
become accessible than for trivial exchanges. When you believe that I have
cheated you in a major way, there should be a flood of memories about your
past history with me: You must decide whether it is worth your while to
continue our relationship. In addition, this information will help you ne-
gotiate with me if you choose to continue our relationship: You can com-
municate how large a cost I have inflicted on you now and in the past (so
I can make amends if I want to continue the relationship), tell me how close
you came to ending our relationship (i.e., categorizing me as a permanent
defector), convince me that 1 have become increasingly untrustworthy,
threaten to injure my reputation by telling others about my past transgres-
sions, and so on.
The activation of past situations in which I have cheated you may, in
turn, activate other” affective mechanisms that communicate cost/benefit
information: They may cause you to cry, turn your back on me, scream at
me, or hit me. The extent and nature of the overt aspects of your affective
reaction communicates to me your view of the extent of my injury of you:
whether you view it as serious enough to require restitution, how much is
required and how soon, whether you intend to cut me off if I defect again,
etc. Emotion communication can be viewed as one way individuals com-
municate costs, benefits, and behavioral intentions to others in negotiative
situations (see Cosmides 1983).

Proposition 3. One Must be Able to Communicate One’s Values

to Others
To engage in an exchange with you, I must know what you want. Although
language is certainly a useful means for communicating what one values,
nonlinguistic organisms can also engage in social exchange-however, the

lo An algorithm was submitted to Axelrod’s computer tournament that computed the conditional
probability that an interactant would cooperate based on whether that individual had cooperated
or defected in past interactions (REVISED DOWNING). It cooperated only when this con-
ditional probability was greater than 50% (random). Its downfall was that it did not discount
past behavior relative to present behavior. Therefore, it was exploited by certain programs that
became more likely to cheat in later interactions. In a sense, it failed because it assumed that
competitor programs had static “personalities.”
” We say “other” because we see no principled way of drawing a dividing line between emotion
and cognition. The flood of memories commonly experienced when a person is betrayed is as
much a part of one’s “emotional reaction” as turning red and attacking (see Tooby 1985; Tooby
and Cosmides, in preparation, b).
64 L. Cosmides and J. Tooby

range of items they can exchange is necessarily more limited. For example,
chimpanzees recruit support from others in aggressive encounters and fre-
quently form long-term coalitional relationships (e.g., de Waal 1982). These
coalitions are social exchanges in which the exchanged “item” is mutual
aid in lights. A chimpanzee under attack bares its teeth, emits a fear scream,
looks at the individual from whom it wants support, and holds out its hand,
palm up, toward that individual. If the attacked chimpanzee receives the
requested support, its demeanor changes radically: Its hair stands on end,
it emits aggressive barks, and it charges its opponent-looking over its shoul-
der frequently to see if its supporter is still with it. If the chimpanzee does
not receive support, it continues cowering with hair flat and teeth bared,
screaming and holding out its hand to solicit support.
One also must be able to communicate dissatisfaction with a defector.
This also can be done without language, as is vividly illustrated by an in-
teraction between Puist and Luit, two chimpanzees in the Arnhem chim-
panzee colony in the Netherlands. Puist and Luit had a long-standing coa-
litional relationship: Puist had a long history of aiding Luit whenever he
attacked or was under attack, and Luit had a long history of extending similar
aid to Puist.
This happened once after Puist had supported Luit in chasing Nikkie [an-
other chimpanzee]. When Nikkie later displayed [aggressively] at Puist she
turned to Luit and held out her hand to him in search of support. Luit,
however, did nothing to protect her against Nikkie’s attack. Immediately
Puist turned on Luit, barking furiously, chased him across the enclosure
and even hit him. (de Waal 1982, p. 207)

The communication of desires, entitlements, and unfulfilled obligations is

possible without language, given that the communicators are both pro-
grammed to understand the signals. It requires that a gesturaVreferentia1
system be shared by the potential cooperators.
A cognitive system that can enable the communication of desires re-
quires more than the development of a few signs. The signs must be coupled
with a referential system, If I want to exchange an axe for something, how
do I indicate what I want? Let’s say I point to the pear you are holding in
your hand. What am I referring to by pointing to the pear? Do I want that
particular pear? Any pear at all? Five bushels of pears? A fruit of some kind,
not necessarily a pear? To be led to the site where you found such good
pears? Do I want you to hold a branch-ladder so I can climb into a tree that
has pears? Or a tree with some other kind of fruit? Do I want to use my axe
to core the pear, in exchange for half the pear? And so on.
The ambiguity of reference in the absence of a shared referential system
is no mere philosophical puzzle (e.g., Quine 1969; Gleitman and Wanner
1982). For example, it is not clear that the infliction of pain, in the absence
of a shared referential framework, could communicate what it is that the
individual inflicting the pain wants the other individual to stop doing. The
difficulty of communicating desires in the absence of a shared system of
 A Computational Theory of Social Exchange 65

reference is illustrated by certain “communication gaps” that occur between

two different, but closely related, species of baboons: hamadryas baboons
and Savannah baboons.
A male hamadryas baboon acquires a “harem” of females by kidnaping
juvenile females from other troops. He leads them to water holes and feeding
grounds that are widely scattered in the inhospitable Ethiopian badlands.
To keep a kidnaped female from straying, the male bites her whenever she
wanders even a few feet from where he wants her. But how does the female
know what this bite refers to, what it is that the male does not want her to
do? This may seem like a straightforward case of “narrowing hypotheses”
through conditioning. However, the same herding technique does not work
on a female Savannah baboon. When she is abducted, the hamadryas male
tries to keep her in line by biting her, to no avail. The Savannah female never
“gets” what it is he wants, and simply runs off. For males, knowing that
one can condition hamadryas females by biting them appears to be no more
“implicit in the situation” than knowing what a bite means. Savannah-ha-
madryas hybrid males who live among hamadryas baboons cannot keep a
harem-the hybrid male never “figures out” that he can herd females
through biting (Hrdy 1981).
Apparently, the learning mechanisms of hamadryas and Savannah ba-
boons include different referential systems. Hamadryas males and females
both “know” that a bite means “stay with the herd”; Savannah baboons
do not. The ability to smile, hug, or inflict pain is not enough. A gestural
system for indicating preference that is not cognitively coupled to a refer-
ential system would be inaccurate at best, and impossible at worst.
The gestural/referential system that allows members of nonlinguistic
species to signal costs, benefits, and behavioral intentions to conspecifics
can be thought of as an emotion communication system. Indeed, ethologists
have traditionally considered such signaling the primary function of emo-
tional expression, studying intention movements, courtship dances, agonis-
tic displays, and aggressive interactions in mammals, birds, reptiles, fish,
and insects. Like modern nonhuman primates, our prelinguistic hominid
ancestors undoubtedly had such a system and used it to communicate about
social exchange. For example, to this day, humans all over the globe share
the same facial expressions of emotion (Eibl-Eibesfeldt 1975; Ekman 1982);
we even share many of these facial expressions with nonhuman primates
(Jolly 1972, pp. 158-159). The same is true for certain auditory signals, like
screaming and crying (Eibl-Eibesfeldt 1975). We can think of no reason why
the appearance of language would cause this more ancient system to be
selected out. Moreover, to the extent that such signals are universally
shared, they have some interesting properties that spoken language lacks:
1. Because they are universally shared, emotion signals can be recognized
by anyone. By aiding “translation,” such signals expand the range of
possible interactants to individuals who speak a different language, in-
66 L. Cosmides and J. Tooby

dividuals who cannot yet speak a language (small children), and distant
individuals beyond the reach of speech but not of sight.
2. Emotion signals can function like intersubjective metrics, permitting an
observer to scale the values of the person emitting the signal: A very
loud scream indicates a greater cost to the screamer than a moderately
loud scream. Signals like screams, smiles, and trembling are “analog”:
The louder the scream, the wider the smile, the more noticeable the trem-
ble-the more strongly the person can be presumed to feel about the
situation causing her to scream, smile, or tremble. Words do not provide
such convenient indicators of magnitude, precisely because they are ar-
bitrary and discrete symbols. Verbal expressions indicating size of cost
or benefit are more ,“digital”: One might reasonably use “very much”
to describe the degree of one’s desire in both these sentences: “I want
very much for my child’s cancer to go into remission” and “I want that
apple very much”-yet in these two cases the degree of desire is, pre-
sumably, vastly different.
3. Emotion signals allow the incidental communication of values to potential
interactants. By observing “your” emotional reactions to various situ-
ations, even though they are not directed at me, “I” can learn what you
value, and hence what sort of exchange you are likely to agree to (see
Proposition 4). The verbal alternative is a process akin to writing to Santa
Claus: Reciting a long list stating one’s preference hierarchy, with pe-
riodic updates. I2

However, the very properties that make a natural language a poor me-
dium for communicating intensity of affect make it an excellent system for
indicating “items” of exchange. The variety of “items” that can be ex-
changed is severely limited in a species that uses only emotion signals. Pri-
mates appear to exchange primarily acts of aggression, protection, food,
sex, alarm, and grooming. The use of language does not, of course, eliminate
the problem of ambiguous reference. In the absence of a shared referential
semantics, knowing what a word refers to is no less problematic than know-
ing what a gesture refers to. l3 But a natural language permits a potentially
infinite number of arbitrary, discriminable symbols to be attached to a po-
tentially infinite number of discriminable classes or entities. As new situa-
tions arise, new words can be opportunistically created to refer to them.
Consequently, language permits a range and specificity of reference impos-
sible in the purely gestural systems of most primates.

‘* Actually, a list stating that you want X, Y, and Z is not sufficient. Your preferences, including
items you already have, would have to be hierarchically ordered using some sort of interval
scale or indifference curves, because the salient issue is: What would you be willing to give up
in order to get X, Y, and Z?
I3 This problem has prompted developmental psycholinguists to posit that children have innately
specified “hypotheses” about what sorts of entities are likely to have words attached to them.
When coupled with articulated models of the world, this hypothesis and model system amounts
to a referential semantics (Gleitman and Wanner 1982).
 A Computational Theory of Social Exchange 67

This property of language opens the vast realm of human adaptations

associated with planning and tool-use to social exchange. Tool technology
continually changes, l4 with new tools being invented constantly. New tech-
nologies enable new and constantly changing opportunities for coordinated,
cooperative behaviors that can themselves become “items” of exchange.
Great benefits can be had by exchanging tools and by participating in the
complex and opportunistically shifting cooperative enterprises these allow-
but only if the tools and behavioral routines can be named. The expanded
power of reference that language affords in social exchange may have been
one factor selecting for its evolution. It is not clear that any but the simplest
tool-using cooperative enterprises could be accomplished with a nonlin-
guistic gestural system-routines like the chimpanzees’ ladder expedition
(see footnote 6), which are discovered quite publicly in the context of an
emotionally salient event” and don’t require long periods of planning.
The evolution of language does not obviate the ability to communicate
cost/benefit information through emotion signals. In fact, the more items
that members of a species can name and exchange, and the more the in-
strumental value of these items varies between individuals and over time,
the more one needs an “item-independent” yet universally understood sys-
tem for communicating how much one values an item.
Because the variety of items exchanged by nonlinguistic primates is
limited, some sets of items whose marginal value did not vary could (hy-
pothetically) have unique cost/benefit weightings associated with them that
are shared by most other members of the species (e.g., ten grooms deserves
one assist in a tight, a season of protection by a male deserves exclusive
sexual access at the height of estrus, etc.). Theoretically, such items could
have a preprogrammed, universally recognized “exchange rate.”
But there can be no preprogrammed, universally acknowledged “ex-
change rate” for a constantly changing array of tools and coordinated be-
havioral routines. Language combined with emotion signaling affords a
uniquely powerful communicative system for social exchange in a planning,
tool using, and opportunistically cooperative species. A wide variety of items
can be precisely specified through language, and their relative value to an
individual can be simultaneously communicated, either incidentally16 or in-

I4 At least for Homo sapiens sapiens. The Homo erectus tool kit appears surprisingly constant
over a wide range of different environments, from Asia to Africa, for over 1 million years
(Pilbeam, pers. comm.). Of course, this observation applies only to tools that are recognizable
as such in the fossil record. For example, a branch used as a ladder would not show up in the
fossil record.
” The Arnhem chimpanzees discovered the ladder trick when one screaming chimpanzee,
fleeing from a very public attack, bounded up a broken branch that happened to be resting
against a tree.
I6 Because the incidental communication of cost/benefit information is important (see Propo-
sition 4), one might predict that, all else equal, individuals are more likely to emit emotion
signals in the presence (or suspected presence) of potential reciprocators than when they are
alone. Similarly, they should be more likely to suppress emotion signals in the presence of
potential aggressors-value information helps aggressors; it tells them what they should threaten
to kill, destroy, or prevent.
68 L. Cosmides and J. Tooby

tentionally, via emotion signals. Indeed, there is preliminary evidence sug-

gesting that some aspects of the acoustic expression of emotion in humans
have been integrated into our species-specific language capacity in ways that
facilitate the communication of values and intentions (Cosmides 1983).

Proposition 4. One Must be Able to Model the Values of Other

Individuals
In some ways, Proposition 4 is just the other side of Proposition 3: One must
have a cognitive system capable of decoding communications of the sort
described in Proposition 3. In addition to this, however, one ought to have
learning mechanisms that are specialized for picking up incidental infor-
mation about the values of potential interactants-for doing “marketing re-
search.” In order to propose an exchange for mutual benefit, one must have
some notion of what kind of “item” the other individual is likely to value.
The individual who is well-equipped to do “marketing research” on potential
interactants will be able to initiate far more exchanges than the individual
who waits for potential interactants to intentionally announce their prefer-
ence hierarchies.
Because emotion signals flag cost/benefit information, they should au-
tomatically recruit attention and be difficult to ignore. An ear-splitting
scream should be more difficult to ignore than an equally loud train whistle;
soft sobbing from the next room should be harder to ignore than the loud
honk of a car horn outside. A broad smile should recruit more attention than
the sound of a motor starting up or than waves in tall grass as it is blown
by the wind.” Attention should be more sustained for emotion signals emit-
ted by a potential interactant: The cry of a friend should recruit more sus-
tained attention than the cry of a stranger.
Not only should attention be drawn to emotion signals, but one’s learn-
ing mechanisms should be quick to pick up what the signal refers to-what,
exactly, the person emitting the signal is reacting to. This implies that our
referential semantics (see footnote 13) includes “hypotheses” about what
kinds of events emotion signals are likely to refer to-hypotheses about
what other individuals are likely to value. Having such hypotheses is all the
more important because many negative emotion signals refer to valued items
that are not present or have not happened, vastly complicating the task of
assigning a referent. When a person is hungry, he or she may moan because
the thing valued-food-is not present. Others must infer the desire for food
from the moan, even though there is no spatio-temporally contiguous event
in which the signal (moan) and the referent (food) are both present.
Evolutionary theory provides a rich heuristic base for developing the-
ories about what kinds of preference information are included in our ref-

” Conditioned stimuli linked to events producing large costs or benefits should also recruit
attention, e.g., a fire engine siren on your street.
 A Computational Theory of Social Exchange 69

erential semantics. Because humans are tool users, planners, and cooper-
ators who can invent many alternative means for realizing a particular goal,
many specific items of human preference will differ from culture to culture
in ways that depend on that culture’s technology, ecology, social system,
and history. This does not mean, however, that desires are random. Evo-
lutionary theory is rife with hypotheses regarding what states of affairs the
typical human is likely to prefer (see Cosmides 1985, pp. 165-167). A cog-
nitive “list” of typical human preferences would still be inadequate, how-
ever, because there are complex interactions between competing prefer-
ences that evolutionary theory speaks to (e.g., what do you do if your spouse
beats you, but he is your only source of income?). Therefore, the algorithms
that guide our “marketing research” must include cost/benefit analysis pro-
cedures that allow one to take such complexities into account in modeling
other people’s values.
Although researchers from Bartlett (1932) to Schank and Abelson (1977)
have posited that pragmatic inference is guided by “schemas,” “frames,”
or “scripts’‘-domain specific inference procedures-they have provided
little insight into their specific content. Using evolutionary biology as a
guide, the system so far proposed (default hypotheses about typical human
preference hierarchies plus procedures for combining factors) provides a
starting place for elucidating the content of “motivation scripts”-algo-
rithms that guide pragmatic inference about human preference and
motivation.
Motivation scripts should be powerful and sophisticated, for the ability
to model other people’s values is useful in a wide variety of evolutionarily
important social contexts, from social exchange to aggressive threat to mate
choice to parenting. They should prove to be strong organizational factors
in the construction and reconstruction of memories. Details that are normally
considered insignificant should be more easily recalled when activated mo-
tivation scripts allow them to be perceived as causally linked to biologically
significant variables. I8 Veridical recall of stories that violate the assumptions
about human preference instantiated in our motivation scripts should be
difficult (as, indeed it is: e.g., Bartlett 1932). Motivation scripts should
guide the reconstruction of such stories during recall, distorting the original
story in ways that make motivational sense. Implicit motivational assump-
tions are so pervasive in human communication that motivation scripts will
probably be an essential component of any artificial intelligence program
that can usefully converse in a natural language.
An emotion signal should not only recruit attention and activate one’s
own motivation scripts, it should arouse one’s curiosity. One would expect
increased tendencies to observe the emotion-arousing event and ask ques-

” Owens, Bower, and Black (1979) present evidence of this kind. Interestingly, the most bi-
ologically significant motivational theme (an unwanted pregnancy) elicited the highest recall of
mundane details about a character’s day.
70 L. Cosmides and J. Tooby

tions about it. Crowds gather around fights, children follow fire trucks to
the scene of a fire, onlookers bombard police with questions at the scene
of a crime. Journalists make a profession of gathering information about the
values and behavior of people who have a large (real or perceived) impact
on our lives. Motivation scripts may guide inferences about what exactly a
given emotion signal refers to, but it can do this only if it is fed concrete
information. The concrete information one acquires by witnessing an emo-
tion-arousing event tills in parameter values in motivation scripts, deter-
mining which data structures and inference procedures are appropriate in
decoding the reacting person’s values.”
Acquiring information about the values of potential interactants is, in
itself, valuable. Decoding the value systems of potential interactants is there-
fore likely to become a cooperative enterprise in itself. We even have a name
for such exchanges of information and “analysis’‘-gossip. Gossip is usually
about situations that cause emotional reactions in potential interactants-
exactly the kind of situations that provide a window into someone’s values.
The more biologically significant the information, the “hotter” the gossip:
Events involving sex, pregnancy, fights, windfalls, and death should be par-
ticularly “hot” topics, especially when they signal a change in someone’s
needs, values, or capacity to confer benefits. Hot gossip should be partic-
ularly interesting and easily remembered. Gossip about people who can have
a large impact on one’s well-being should be especially interesting; gossip
about people one does not know should be comparatively boring. Similarly,
cues or indicators of the character of potential interactants, including their
disposition to cheat or defect, are themselves extremely valuable informa-
tion, differentially attended to and exchanged. Reputation is an important
kind of social information, and it plays a significant role in the social life of
any stable group of humans.
The learning mechanisms that guide such “marketing research” should
produce person-specific models of the preferences and motivations of po-
tential and actual interactants. General motivation scripts help build person-
specific preference models; these become more elaborated the more contact
one has with that particular person. As this happens, inferences drawn from
a person-specific model will generate more accurate interpretations of that
person’s behavior and emotion signals than inferences drawn from the gen-
eral motivation scripts.
It would be useless for information about the preferences of different
individuals to be stored together in a semantic network, filed under “pref-
erences” or “values.” Like information about an individual’s history of
reciprocation, a model of an individual’s preferences and motivations should

r9 There are, of course, other good reasons for being curious about biologically significant
events, e.g., you yourself might be confronted with the same situation at some point. However,
when such events impact potential interactants, they should be especially interesting: A fist
fight in your academic department provokes more interest than one among strangers in another
city.
 A Computational Theory of Social Exchange 71

be “filed” under his or her identity. When the opportunity to acquire more
preference information about an individual arises, the model appropriate to
that individual must be easily retrieved, not just a model of average pref-
erences. “Averaging” the fact that one person prefers Z to W but another
person prefers W to Z into one model of “average” preference does not
enhance one’s ability to engage in social exchange.” In contrast, learning
that “Smith values W more than X more than Y more than Z” and that
“Jones values Z more than X more than Y more than w” increases your
ability to make offers that benefit you given the limits imposed by what Smith
or Jones are willing to accept. Offering W to Smith is more likely to induce
him to give you Y than offering him Z; exactly the reverse is true of Jones.
If you value Z more than W, you are better off making Smith an offer; if
you value W more than Z, then strike a deal with Jones. The proper decision
can be made only if person-specific preference information can be conven-
iently retrieved.

Proposition 5. Human Algorithms Regulating Social Exchange

Should Not be Closely Tied to Particular Items of Exchange
That tools, information about tool making, and participation in opportun-
istically created, coordinated behavioral routines were important items for
exchange has implications for the structure of human cognitive algorithms
regulating social exchange. The more limited the range of items exchanged,
the more specific the algorithms regulating exchange can be. For example,
the items exchanged in a cleaning fish symbiosis can be directly specified
in the algorithms regulating the exchange. The host fish is specifically pro-
grammed to discriminate cleaner fish from similar-looking prey items, and,
upon recognizing one, to refrain from eating it. The cleaner fish is specifically
programmed to discriminate a host fish from other large, predatory fish, and,
upon recognizing one, to approach and eat its ectoparasites (Trivers 1971).
Whereas the exchange algorithms of other organisms can be specific to the
relatively few items they exchange, human algorithms regulating social ex-
change should be able to take a wide variety of input items, as long as these
items are perceived as costs and benefits to the individuals involved in the
exchange.
However, despite our remoteness from the Pleistocene and the range
of items commonly traded then, some items should be more readily perceived
as costs and benefits: those for which the perceiver can ascertain a clear
causal link to adaptively significant variables like offspring, kin, sex, food,
safety, shelter, protection, aggressive formidability, and dominance. Evi-
dence for this is so universal that there seems little point in belaboring it.

*’ Although noting that most people in your culture prefer W to Z might enhance your ability
to recognize and participate in social exchanges with new interactants. One might expect such
culture-specific information to be incorporated into the “typical human” motivation scripts.
72 L. Cosmides and J. Tooby

For example, a Mr. Michael Pastore of Dallas recently made the following
comment in an interview in the Wall Street Journal:

“I never pay for dinner with anything other than my [American Express]
Platinum Card when I’m on a first date,” says the 30-year-old seafood im-
porter, flashing his plastic sliver inside the glitzy Acapulco Bar. “Women
are really attracted to the success that my card represents.” (The Wall
Street Journal, April 17, 1985, p. 35)

Mr. Pastore perceives a clear causal link between his “plastic sliver” and
a biologically significant variable: the ability to attract sexual partners. His
perception that a Platinum Card can attract sexual partners is based in turn
on the perception that owning one is causally linked to a variable that is
biologically significant to females in choosing male sexual partners-the
ability to accrue resources.*’ Knowing this, one can infer that Mr. Pastore
perceives owning an American Express Platinum Card as a benefit, and that
if he did not own one, he might well be willing to give up other items in
order to acquire one. It is a suitable item for social exchange.
The prediction, then, is that the algorithms regulating social exchange
in humans will be item-independent. Furthermore, they will operate on COSC-
benefit representations of the interaction. As we will argue in the next sec-
tion, any interaction that is interpreted as having a particular, characteristic,
cost-benefit structure will be categorized as an instance of social exchange
and will call up procedural knowledge specialized for reasoning about this
domain.

THE GRAMMAR OF SOCIAL CONTRACTS

A grammar of social contracts specifies the properties that must be embodied

by Darwinian algorithms for reasoning about social exchange. It incorporates
the strategic constraints outlined in Part 1 and the ecological constraints
outlined in Part 2.
Just as a grammar of the English language is a set of rules for distin-
guishing well-formed sentences from ill-formed sentences, a grammar of
social contracts is a set of rules for distinguishing well-formed social con-
tracts from ill-formed social contracts. It includes the set of assumptions
about the rules governing social exchange that must somehow be incarnated
in the psychological mechanisms of both participants. Without these as-
sumptions, much of what people say, mean, and intend to do in exchange
situations could not be understood or anticipated. This grammar creates the
“cohesion of discourse” (Wason and Johnson-Laird 1972, p. 92), and the
cohesion of behavior, in interactions involving uncoerced exchange. It con-

” In fact, cross-cultural evidence is accumulating that indicates that a potential mate’s ability
to accrue resources is more important to women than to men, just as evolutionary theory predicts
(Buss 1987).
 A Computational Theory of Social Exchange 73

stitutes the procedural knowledge that individuals must share in drder to

communicate their intentions to others in this particular kind of negotiative
interaction (see Cosmides 1983).
Moreover, in establishing what conditions must hold if a social contract
is to be recognized as well-formed, the grammar also provides a framework
for understanding what deviations from these conditions mean. For example,
certain types of deviations will indicate “dishonorable” intentions (see dis-
cussion of “baseline fraud” below), others will indicate an insult (e.g., “I’d
pay a whole nickel to sleep with your mother”), and still others will indicate
that the speaker is either joking (e.g., “I’d sell my firstborn for a cigarette”)
or eccentric (“I’ll give you my paycheck for your gumwrapper”). The gram-
mar of social contracts thus provides a framework for understanding the
many shades and colors of meaning in situations involving exchange. The
ability to formally specify such “grammars,” and hence analyze conformity
and systematic deviations from them, provides a basis for assimilating the
“interpretation of cultures” (Geertz 1973) into a cogent evolutionary frame-
work. Cultural systems of meaning have been widely regarded among social
anthropologists as being entirely beyond the capability of an evolutionary
perspective to address or explain (Sahlins 1976). However, because such
“symbolic productions” are both produced and interpreted by adaptively
structured cognitive mechanisms, “interpretation” can be converted from
a literary exercise to an exercise in evolutionary cognitive psychology. As
an increasing number of adaptive “grammars” of cognitive mechanisms are
mapped, the nature and systematic properties of cultural meaning systems
may become increasingly apparent, and tractable to an evolutionary and
empirical treatment.
We shall describe the grammar of social contracts in pedantic detail,
because it serves several important empirical and theoretical functions.
First, it is a productive source of hypotheses for uncovering the cognitive
mechanisms involved. Second, such explicitness is required to rigorously
construct a cognitive theory, e.g., one specified enough to implement in an
artificial intelligence program. The creation of such an AI program is the
test of whether one really has a complete cognitive theory, and at a minimum,
any AI program capable of understanding and reasoning about social ex-
change would have to embody this grammar of social contracts. This exercise
is useful because it exposes the hidden assumptions and complexities that
theoretical handwaving misses. Because we have such innate algorithms,
these tasks seem transparently simple, and this illusion of simplicity con-
stantly leads researchers to underestimate the intricacy of the cognitive ma-
chinery necessary to perform these tasks. And third, no psychological theory
claiming to account for how people learn to engage in social exchange can
be considered adequate unless it can be shown that the learning mechanism
proposed is powerful enough to permit the acquisition of the grammar of
social contracts. Thus, a well-specified computational theory of social ex-
74 L. Cosmides and J. Tooby

change provides a crucial test of adequacy that any proposed psychological

theory must be able to pass (Cosmides and Tooby 1987).
The items valued by our hominid ancestors were correlated with costs
and benefits in their inclusive fitness; otherwise social exchange could not
have evolved. The strategic exigencies of exchanging items that had real
effects on the inclusive fitness of the exchangers selected for algorithms
programmed with a particular set of cost/benefit relations (see Fig. 1, Part
1). These relations can be expected to regulate how we think about social
exchange, even if the items we value today are no longer correlated with
our inclusive fitness. The grammar of social contracts specifies these cost/
benefit relations.
Unlike the exchange algorithms of cleaner fish or even baboons, human
algorithms for regulating social exchange should be item-independent: They
should represent items of exchange as costs and benefits to the participants,
and operate on those representations (see Proposition 5). The proposed
grammar of social contracts is therefore expressed in cost/benefit
terminology.
What must P and Q stand for if the sentence “If P then Q” is to in-
stantiate a well-formed social contract?
To make the discussion concrete, let’s fill in some values for P and Q
in the offer “If P then Q.” Let’s say “I” offer “you” the following contract:
“If you walk my dog, then I’ll give you a million dollars.” P stands for “you
walk my dog” and Q stands for “I’ll give you a million dollars.” Likewise,
not-P stands for “you do not walk my dog” and not-Q stands for “I do not
give you a million dollars.”
At the time of this offer, but independent of it, you have a certain level
of “well-being” and certain expectations about the future, all of which play
some part in determining what you would, at this point, consider to be of
value. This baseline will be termed your zero level utility. For simplicity’s
sake, let us assume that 1) value is subjective, and 2) the individual is the
final arbiter of what he or she finds valuable. Natural selection theory has
much to say about what kinds of items and states most humans will consider
valuable (i.e., about preferences and motivations; see Propositions 4 and 5),
but such considerations are unnecessary for this analysis.

What Conditions Must Hold for You to Accept This Offer?

Let us consider what conditions must hold for you to accept this offer. Your
zero level utility baseline is derived from a vast number of conditions and
expectations about the state of the world. In the absence of this offer, one
of those expectations about the future must be not-Q-you do not expect
to be receiving $1 ,OOO,OOO from me. If not-Q comes to pass, your utility level
will not have moved from your zero level baseline, O(you).
Q-receiving $1,000,000 from me-must be something that you con-
sider to be a benefit. An “item’‘-an act, entity, or state of’affairs-is a
 A Computational Theory of Social Exchange 75

benefif ro you @(you)) if, and only if, it increases your utility above your
zero level baseline.22 Assuming you value having a million dollars (Q) more
than you value not having a million dollars (not-Q), then Q-having a million
dollars-constitutes a benefit to you. You will not accept this offer unless,
at the time of acceptance, you believe that Q constitutes a benefit to you.
Using terms defined with respect to your values (rather than mine), we can
rephrase my offer as: “If P then B(you).”
An item is a cost to you (C(you)) if, and only if, it decreases your utility
below your zero level baseline. 23 In this offer, P-walking my dog-is the
item that I have made my offer of B(you) contingent upon. Usually, P will
be something that you would not do in the absence of an inducement; other-
wise, I would be incurring a cost (giving up Q, the million dollars) by making
the offer (if you were going to walk my dog anyway it would be foolish of
me to offer you the million dollars).24 If P is not something you expected to
do in the absence of my offer, then, in your value system, not-P (not walking
my dog) is part of your zero level baseline, O(you). This means that if not-
P comes to pass, you will not have moved from your zero utility baseline,
and you will be no worse off than if my offer had never been made. If we
posit that my dog is ugly and vicious, and that walking him would embarrass
you, endanger your health, and assault your aesthetic sensibilities, then P
(walking my dog) decreases your utility and is therefore a cost to you,
C(YOUh
Stated in terms of your value system, my offer can now be rephrased
as “If C(you) then B(you).” But other conditions must hold before you will
accept my offer. There is a constraint on the magnitudes (absolute values)
of B and C, namely, B(you) > C(you), or, equivalently, &you) minus C(you)
> 0. For you to accept my offer, a million dollars must be more of a benefit
to you than walking my ugly dog is a cost. If this is not the case there would
be no point in your entering into the contract; it would not increase your
utility. The greater the magnitude of B minus C, the more attractive the
contract will appear. A contract that reversed this constraint (such that
C >> B) sounds perverse. For example, the following offer strikes people
as foolish: “If you break your arm then I’ll give you a penny.” Unsurpris-
ingly, Fillenbaum (1976) found that subjects consider such offers “extraor-
dinary” 75% of the time, compared to a 13% rate for offers that fit the
constraints described above.

22 Presumably there are costs and benefits associated with any action. More precisely, B(you)
is a net benefit-the benefit to you of receiving $1 million is greater than the cost to you of
receiving $1 million.
l3 Again, this is a net cost-the cost to you of walking my dog is greater than the benefit to
you of walking my dog.
24 P does not have to be a C(you) for you to accept my contract, although I must believe that
it is a C(you) in order to offer the contract in the first place. You could be trying to defraud
me into offering this contract by dissembling about your real intentions. Perhaps you have been
planning all along to walk my dog, but led me to believe that you are not planning to walk it
so I would make you an offer. See below, on baseline fraud.
76 L. Cosmides and J. Toeby

What Conditions Must Hold For Me to be Willing To Make an

Offer?
We can also consider the contract from the point of view of the person
offering it, for simplicity’s sake identified as “me.” What conditions must
hold for me to be willing to offer a contract? First, I must believe that not-
P (your not walking my dog) will come to pass if I do not make the offer.
This means that not-P is a component of my zero level baseline: If not-P
comes to pass, my utility level will not have changed. Second, I must want
P-in my value system, having my dog walked must increase my utility, it
must be a bene$t to me @(me)). Third, not-Q-not giving you $1,000,000-
usually will be part of my zero level baseline, O(me); if you do not accept
my offer, I do not plan on giving you $l,OOO,OOO,and if not-Q comes to pass,
I will not have moved from my zero utility baseline.25 Fourth, if not-Q is
part of my zero baseline, then Q-giving you $l,OOO,OOO-represents a de-
crease in my utility and is therefore a cost to me (C(me)). Fifth, like you,
I will not enter into the contract (offer it in the first place) unless B(me) >
C(me) (unless having my dog walked is worth more to me than relinquishing
the million dollars; this example sounds eccentric precisely because most
readers would assume that it violates this constraint).
In other words, I want P, and I am willing to give up Q to get you to
do P; but I am not willing to give up Q without getting P. (I want you to
walk my dog and I am willing to give up $l,OOO,OOO to get you to walk him;
but I am not willing to give up my $l,OOO,OOOwithout your walking him.)
In your value system, “If P then Q” translates to: “If C(you) then
B(you).” (i.e., “If I incur the cost of walking her dog, then I will get the
benefit of receiving $1,000,000 from her.“) However, in my value system
the same offer translates to “If B(me) then C(me)” (i.e., “If I get the benefit
of your walking my dog, then I will incur the cost of relinquishing my
$l,OOO,OOOto you”). As you can see, P represents a different utility level
to me @(me)) than it does to you (C(you)). The same holds for Q. In a well-
formed social contract, that is, a contract that I am willing to offer and you
are willing to accept, the utility levels associated with P and Q are those
shown in Table 1.
An offer is not entirely symmetrical, however. Suppose there were some
way of equating value systems. Although B > C for both of us (or else we
would not both agree to the contract), P (walking my dog) might be a smaller
cost to you than Q (giving up $l,OOO,OOOto you) is to me (or vice versa).
Likewise, Q might be a larger benefit to you than P is to me. These asym-
metries may lead to a difference in the magnitude of our “profit margins”
(B minus C). Where it is possible to vary the effective magnitude of the acts

*’ Not-Q being part of my zero level baseline is not a necessary condition for my making an
offer, but it is necessary that you be&~ it is part of my zero baseline if you are to accept my
offer. Unknown to you, I might intend to give you $l,OOO,OOO regardless, but want to get as
much as I can in return.
 A Computational Theory of Social Exchange 77

Table 1. Cost/Benefit Translation of My Offer into Your Value System and Mine

My offer: “If P then Q ”

(“If you walk my dog then I’11give you $lm”)
Your Point of
View My Point of View
P (you walk my dog) C(YOU) B(me)
not-P (you do not walk my dog) WYOU) O(me)
Q (I give you my $Im) B(Yo@ C(me)
nor-Q (I do not give you my $lm) WYOU) Wme)

or items of exchange (together with their associated costs and benefits),

unequal profit margins invite bargaining: Each may attempt to increase his
or her “profit margin” at the expense of the other. Bargaining under such
constraints constitutes an antagonistic game, because more B(you) per unit
C(you) corresponds to more C(me) per unit B(me). (See Fig. 2; for a fuller
account of these intercontingent relations and their psychological sequelae,
see Tooby 1975). However, as long as your profit margin is greater than
zero, it is in your interest to accept my offer, regardless of how large my
pro$t margin is (and vice versa). If B > C for both parties, then both parties
have benefited from the exchange. For this reason, we consider the term
“subtle cheating,” which Trivers (1971) uses to describe an interaction in
which profit margins are unequal, to be a misnomer. “Under-reciprocating”
is a more appropriate term; “cheating” is more usefully reserved for the
violation of a contract.

Baseline Fraud
There is a joke that runs like this:
A man from out of town walks up to a woman and says “If you sleep with
me three times 1’11give you $15,000.” She is hard up for cash, so she agrees.
After each session he pays her the money he promised. The woman decides
this is an easy way to make money, so after she has been paid the full
$15,000 she asks him if he would like to continue the arrangement. He says
he can’t because he must return home the next day. She asks “Where’s
home?” “ Oshkosh,” he replies. “Oh!” she says, “That’s where my mother
lives!” He answers, “Yes, I know. She gave me $15,000 to deliver to you.”
The man in the joke has defrauded the woman by concealing information
about their zero level baselines.
A contract has been sincerely offered and sincerely accepted when each
party believes that the B > C constraint holds for the other, in this case,
when the contract has the following cost/benefit structure:
Man’s offer: “If you sleep with me three times then I’ll give you $15,000”
“If P then Q”
Woman’s point of view: “If C(woman) then B(woman)”
Man’s point of view: “If B(man) then C(man)”
78 L. Cosmides and J. Tooby

\ :B) s
b S
“room” for S
profit
margin

0 -
S -7
S price b
\ S) b

FIGURE 2. Negotiation in social exchange (Adapted from Tooby 1975).

Imagine a buyer and a seller haggling over the price of a car. V(B) represents
the value of the car to the buyer; if the buyer could get the car for free (C(buyer)
= price = 0) then &buyer) = V(B), the car’s intrinsic value to the buyer. The b-
b-b line shows how the potential buyer’s profit margin changes as a function of price;
the higher the price he pays, the lower his profit margin (V(B) - price). B, the point
where this line intersects the x-axis, is the buyer’s breakeven point, the price at
which his profit margin is zero. The buyer makes a profit if he pays any price less
than B. V(S) represents the value of the car to the seller; if the seller gives it away
@(seller) = price = 0), then C(seller) = V(S), the car’s intrinsic value to the seller.
The s-s-s line shows how the seller’s profit margin changes as a function of price;
the higher the price he gets, the higher his profit margin (price - V(S)). S is the
seller’s breakeven point, the price at which his profit margin is zero. The seller makes
a profit if he sells the car at any price greater than S. Both buyer and seller profit
if the car is sold at any price such that S < price < B. They only profit equally,
however, at price E, the point where b-b-b intersects s-s-s. The buyer will try to
push the price down the s-s-s curve to S, the seller will try to push the price up the
b-b-b-b curve to B. The price range between S and B represents “room for bar-
gaining.” The buyer might try to convince the seller that the seller’s curve is actually
steeper and the buyer’s shallower, that B is really less than it is (i.e., he threatens
to withdraw his offer at a price lower than B), that the seller “ought” to give him
a break, etc. (and vice versa).
The emotional language of Marxist economics and labor negotiations can be
understood with this graph. The worker (person selling his labor) claims he is being
“exploited” and that management is earning “excess profits” when the price of an
hour of his labor is S C=price < E (management’s “excess profit” is the difference
between their profit margin at the price they are currently paying the “exploited”
worker and their lower profit margin at B, the price the worker prefers). Management
(person buying labor) complains that labor unions are strangling the company when
workers succeed in pushing the price of labor up such that E < price I B (“stran-
gling” implies that price > B, a situation that cannot be true if the company is making
a profit greater than zero). In truth, both labor and management benefit compared
to their other options at any price between S and B.

The woman in the joke assumed that the man’s offer fit these requirements,
that he offered a sincere contract. However, the man knew that if the woman
knew what he knew about her baseline and his, they would both see the
structure of the contract as:
“If P then Q”
Woman’s point of view: “If C(woman) then O(woman)”
 A Computational Theory of Social Exchange 79

Man’s point of view: “If B(man) then O(man)”

In actuality, the man gave up nothing in exchange for B(man).
Humor is frequently based on the violation of implicit assumptions. The
punch line of this joke violates the woman’s (and the listener’s) implicit
assumption that the man had offered a “sincere” contract. Although not-Q
is usually part of the offerer’s zero level utility baseline, this is not a rzec-
essavy condition for his making an offer (see above). However, it is a nec-
essary condition of the woman’s acceptance that she believe that, in the
absence of the offer, not-Q would come to pass; otherwise, accepting the
offer would decrease her utility by C(woman).
For any proffered contract of the form: “If you do P then 1’11do Q,”
the acceptor has been the victim of baseline fraud when:
1. The acceptor believes that not-Q will come to pass if he or she turns down
the contract, and
2. This belief is false, and
3. The offerer knows the acceptor holds this false belief, and
4. The offerer either fosters the acceptor’s false belief, or does nothing to
disabuse the acceptor of this belief.
Likewise, the offerer has been the victim of baseline fraud when:
1. The offerer believes that not-P will come to pass if he or she does not
offer the contract (or if it is turned down), and
2. This belief is false, and
3. The acceptor knows the offerer holds this false belief, and
4. The acceptor either fosters the offerer’s false belief, or does nothing to
disabuse the offerer of this belief.
Had the woman wanted to sleep with the man all along, regardless of pay-
ment, she would have thought she was tricking him by getting the added
benefit of $15,000 (until she heard about her mother’s gift). This is be-
cause the offerer’s belief that the potential acceptor’s zero level baseline
includes not-P (not sleeping with him) is a necessary condition for the offerer
to make the offer, but it is not a necessary condition for the acceptor to
accept the offer. Baseline fraud is different from cheating: In baseline fraud
both parties have, technically, honored their contractual obligations. As will
be seen, this is not the case with cheating.

Summary of the Structure of Sincere Social Contracts

The conditions that hold when an individual sincerely offers or sincerely
accepts a social contract are shown in Table 2. For the sake of simplicity,
P and Q stand for the actual items exchanged (these can be actions as well
as entities). The first column shows the contract’s cost/benefit structure in
terms of the sincere offerer’s value system; the second column shows what
the sincere offerer believes the contract’s structure is in terms of the ac-
80 L. Cosmides and J. Tooby

Table 2. Sincere Social Contracts: Cost/Benefit Relations When One Party is Sincere and
That Party Believes the Other Party is Also Sincere

My offer: “If you give me P then I’ll give you Q”

Sincere Offer Sincere Acceptance

I Believe: You believe:

P B(me) C(YOU) B(m) C(YOU)

IIN-P O(me) WYOU) Wne) O(YOU)
Q C(me) B(you) C(m) B(YOU)
not-Q O(me) O(YOU) O(me) O(YOU)
Projif margin Positive: Positive: Positive: Positive:
B(me) > C(me) B(YOU) > B(me) > C(me) B(you) >
C(YON C(you)
Translation
My terms “If B(me) then C(me)” “If &me) then C(me)”
Your terms “If C(you) then B(you)” “If C(you) then B(you)”

ceptor’s value system. The third column shows the contract’s cost/benefit
structure in terms of the sincere acceptor’s value system; the fourth column
shows what the sincere acceptor believes the contract’s structure is in terms
of the offerer’s value system. The table shows that the sincere offerer and
the sincere acceptor view the contract’s cost/benefit structure in exactly the
same way.
Table 3 shows what conditions hold when one person offers or accepts
a contract sincerely, but is the victim of baseline fraud. The sincere person
believes the contract fits the conditions specified in Table 2. However, the
defrauder believes the contract tits the criteria specified in Table 3. Fur-
thermore, if the sincere person were to find out that he or she had been
tricked concerning baseline information, that person would share the de-
frauder’s view of the contract’s cost/benefit structure. An analysis of base-
line fraud is useful because it serves to distinguish the conditions that must

Table 3. Baseline Fraud: Cost/Benefit Relations When a Sincere Party Makes a Social
Contract with an Individual Perpetrating a Baseline Fraud

My offer: “If you give me P then I’ll give you Q”

I try to defraud you; you accept You try to defraud me, I offer
sincerely sincerely
If you knew what I knew. we If I knew what you knew, we
would both believe would both believe

P B(me) ayou) O(YOU) O(m)

not-P O(me) NYOU) 7
C(m)
Q O(me) O(you) B(you) C(m)
not-Q ? C(you) O(you) O(me)
Profit margin Positive: Negative: Positive: Negative:
B(me) > C(me) C(you) B(you) > C(me)
C(YOU)
Translation
M?: terms “If B(me) then O(me)” “If O(me) then C(me)”
Your terms “If C(you then O(you)” “If O(you) then B(you)”
 A Computational Theory of Social Exchange 81

hold for a contract to be offered or accepted, from those that need not hold,
but usually do.
That people represent actions as costs and benefits with reference to a
zero point based on their current expectations is a psychological prediction
that is not strictly necessitated by natural selection theory in its simplest
form. However, reciprocation theory does require that the individual realize
a net increase in its fitness from participation; this could, in principle, be
computed using an ordinal preference scale without reference to a zero point.
We use this system because we believe it provides a powerful means
by which the individual can distinguish exchanges from other kinds of in-
tercontingent behavior. For example, most people would probably recognize
the utterance “If you call the police, then I’ll shoot you” as a threat. Yet
1) it has the same linguistic form as a contract-“If P then (3,” and 2) like
the person who accepts a sincere contract, the person threatened will realize
an increase in fitness by obeying the threat instead of defying it.
However, the hypothesis that humans represent events as costs and
benefits with respect to a zero utility point based on their current baseline
expectations provides a straightforward means of distinguishing threats from
contracts. A contract has the form “If C(you) then B(you).” However a
threat has the form “If O(you) then C(you).” In the absence of the threat,
the hearer in the example intends to call the police; it is part of his or her
zero level baseline. Being shot is not in the hearer’s plans; it constitutes a
cost. This representational system allows the principled differentiation of
the various forms of intercontingent behavior. We hypothesize that recog-
nition of the form of intercontingent behavior at hand (social exchange,
threat, etc.) automatically activates the set of rules appropriate for reasoning
about it. In the next section, we sketch the rules appropriate to social
exchange.

Social Contracts as “Speech Acts”

The relations specified in the previous sections are implicit in the sincere
offer of a contract and its sincere acceptance. But to understand cheating
(a violation of the contract), we have to analyze what contractual obligations
“you” and “I” incur by entering into a contract. This calls for a brief foray
into “speech act” theory (Searle 1971).
Speech act theory is a part of analytic philosophy that grew out of the
realization that, in speaking, people frequently do more than simply refer
to something in the world. Frequently they do something by virtue of saying
something. When I say “I promise to X”, for example, I am not referring
to something in the world: I am making a promise, and thereby incurring
certain obligations--I have committed a “speech act” (e.g., Searle 1971).
“Offering a contract” and “accepting a contract” can both be considered
speech acts. Thus, we can ask the question: What do I mean when I say “If
you give me P then I’ll give you Q” and what do you mean when you say
82 L. Cosmides and J. Tooby

you “accept” my offer? Grice (1957, 1967) has provided a convenient struc-
ture for understanding the meaning of speech acts.
In committing a speech act,

something [a behavior, intention, or frame of mind] intentionally is pro-

duced in another with the intention that he realize why it was produced
and that he realize he was intended to realize all this (Nozick 1981, pp.
369-370, on Grice).

Using this structure and the cost/benefit analysis above, when an actor, “I”,
offers a contract by saying, “If you give me P then I’ll give you Q,” the
actor means:

1. I want you to give me P,

2. My offer fulfills the cost/benefit requirements of a sincere contract (listed
in Table 2).
3. I realize, and I intend that you realize, that 4-9 are entailed if, and only
if, you accept my offer:
4. If you give me P, then I will give you Q,
5. By virtue of my adhering to the conditions of this contract, my belief that
you have given (or will give) me P will be the cause of my giving you Q,
6. If you do not give me P, I will not give you Q,
7. By virtue of my adhering to the conditions of this contract, my belief that
you have not given (or will not give) me P will be the cause of my not
giving you (2,
8. If you accept Q from me, then you are obligated to give me P (alterna-
tively, If you accept Q from me then I am entitled to receive P from you),
9. If you give me P, then I am obligated to give you Q (alternatively, If you
give me P then you are entitled to receive Q from me).

These rules capture the intercontingent nature of social exchange: They

specify the ways in which the behavior of one person is contingent upon the
behavior of another person. Some philosophical refinements are discussed
in Appendix 1. However, these points are not essential to the rest of the
article.
Offering a contract is somewhat more complicated than other speech
acts (like promises; see Searle 1971) in that none of the conditions apply
unless the hearer uccepts the contract. In contrast, the conditions for a
promise (or a threat) hold regardless of whether the hearer consents. Making
a promise is a unilateral act; making a contract is not. In saying that one
accepts an offer, the acceptor means that he or she understands, and agrees
to comply with. the conditions specified in 1-9 (above).
On first inspection it might seem that a contract actually expresses a
biconditional: “Q if and only if P.” If this were the case, the terms of the
contract would be violated (someone would have cheated) if you are not in
possession of Q after having done P (I cheated you), or if you are in pos-
session of Q without having done P (you cheated me). But it is not a bi-
 A Computational Theory of Social Exchange 83

conditional because a social contract involves the twin notions of obligation

and entitlement.
What does it mean for you to be obligated to do P?
1. You have agreed to do P for me under certain contractual conditions (like
l-9), and
2. Those conditions have been met, and
3. By virtue of your not thereupon doing P, you agree that if I use some
means of getting P (or its equivalent) from you that does not involve
getting your voluntary consent, then I will suffer no reprisal from you.
Alternatively, 3 can be:
3. By virtue of your not thereupon giving me P, you agree that if I lower
your utility by some (optimal) amount X (where X > B(you)-your un-
earned gains), then I will suffer no reprisal from you.
The first formulation expresses restitution, the second, punishment. One
would expect the tendency to punish to be greatest when restitution is not
possible. To our knowledge, the conditions determining the optimal size of
X have not yet been formally analyzed. We suspect the optimal X would be
large enough to deter future cheating but small enough that it does not dis-
courage future cooperation. However, it is clear that a cheater would not
be deterred by an X less than or equal to B(cheater). With X = B(cheater),
the potential cheater will be indifferent between cheating and cooperating;
with X < B(cheater) the potential cheater will realize a net benefit by
cheating.
To take reprisal against someone trying to claim “just” restitution or
punishment is to indicate that you are no longer interested in continuing a
relationship with that person. In the contretemps between Puist and Luit,
the two chimpanzees discussed in Proposition 3, Luit allowed Puist to punish
him for his defection. The judgment that this punishment was “allowed”
can be made because Luit is far stronger than Puist, and in a direct test of
strength Puist would not have a chance against Luit (de Waal 1982). To do
otherwise would have signaled an end to their several year reciprocal
relationship.
What does it mean for “you” to be entitled to Q?
1. I have agreed to give you Q under certain contractual conditions (like I-
9>, and
2. Those conditions have been met, and
3. By virtue of my not thereupon giving you Q, I agree that if you use some
means of getting Q (or its equivalent) from me that does not involve getting
my voluntary consent, then you will suffer no reprisal from me.
As in obligation, an alternative formulation of 3 is:
3. By virtue of my not thereupon giving you Q, I agree that if you lower
my utility by some (optimal) amount X (where X > &me)-my unearned
spoils), then you will suffer no reprisal from me.
84 L. Cosmides and J. Tooby

Thus, the notions of entitlement and obligation are reciprocally related: My

being entitled to receive P from you is equivalent to your being obligated
to give me P and vice versa.
A social contract is not a biconditional because “I” must do that which
I am obligated to do, but I am not required to accept that to which I am
entitled. If I pay the cost that I am obligated to pay (C(me), which corre-
sponds to B(you)), I have fulfilled my end of the contract; I do not have to
accept the benefit @(me)) that I am entitled to (however, you must offer
it). Unless conditions have changed, failure to accept a benefit one is entitled
to is foolish (and rare-such behavior would have been strongly selected
against), but no one regards it as violating the terms of the contract. Only
selectional thinking makes sense of these otherwise arbitrary features of
human social cognition.

Looking for Cheaters

As discussed, under biologically plausible circumstances, indiscriminate co-
operation cannot be selected for. The game-theoretic structure of the natural
selection process dictates that social exchange can evolve only if it is gov-
erned by a strategy that requires reciprocation and excludes or retaliates
against cheaters. To implement such a strategy, human social contract al-
gorithms must include procedures that allow us to quickly and effectively
infer whether someone has cheated, or intends to cheat, on a social contract.
For a mechanism to detect cheating, it must have a specification of what
conditions constitute cheating.
Cheating is the violation of the conditions of a social contract. It is the
failure to pay a cost to which you have obligated yourself by accepting a
benefit. The social contract can be explicit or implicit,z6 a private agreement
or a rule of one’s social group.
Let’s assume “I” offered, and “you” accepted, the following contract:
“If you give me P then I’ll give you Q.” In your value system this translates
to: “If C(you) then B(you).” You have cheated me when you have accepted
the item that corresponds to B(you) (item Q) without giving me the item that
corresponds to C(you) (item P). In other words, you have cheated me when
you have accepted item Q from me, but you have not given me item P. This
means I have paid C(me) (item Q), but have not received B(me) (item P).
Your payoff: B(you). My payoff: C(me).
In my value system, the same contract translates to: “If B(me) then
C(me).” I have cheated you when I have accepted B(me) (item P) without
paying C(me) (item Q). In other words, I have cheated you when I have

x Given that hominids probably participated in social exchange long before they had language,
one would expect the act of accepting a benefit to frequently be interpreted as implicit agreement
to a social contract-as a signal that the acceptor feels obligated to reciprocate in the future.
(Of course, one would expect the donor to jump to this interpretation more readily than the
acceptor!) This view is formalized in British and U.S. contract law-a contract is invalid unless
some “consideration” has changed hands-even a symbolic $1 will suffice.
 A Computational Theory of Social Exchange 85

Table 4. How Do You and I Make Out When One of Us Cheats the Other?

I Cheat You You Cheat Me Contract Fulfilled

You give me P B(me) C(you) - - B(me) C(YOU)
You do not give me P - - O(me) O(you) - -
I give you Q - - C(me) B(you) C(m4 B(you)
I do not give you Q O(me) O(you) - - - -
My payoff B(me) C(me) B(me)-C(me)
Your payoff C(YOU) NYOU) B(yo~bC(yo~)

accepted item P from you, but have not given you item Q. This means you
have paid C(you) (item P), but have not received B(you) (item Q). Your
payoff: C(you). My payoff: &me). These relations are summarized in Table
4.
As mentioned in Proposition 5, social contract algorithms in humans
must represent items of exchange as costs and benefits to the participants,
and operate on those representations. The detection of cheating depends on
modeling the exchange’s cost/benefit structure from the point of view of
one’s partner, as well as from one’s own point of view. Thus, for any given
exchange, two descriptions of each item must be computed by the social
contract algorithms. For a sincere contract, “If you give me P, then I’ll give
you Q,” item P should be described as both B(me) and C(you), and item Q
should be described as both C(me) and &you) (see Table 4). The cost/benefit
structure to oneself should be easily recoverable, even if the contract is
phrased in terms of the value system of one’s exchange partner.*’ There is
a structural parallel to transformational grammars as they were initially con-
ceptualized: The “surface structure” is the way the offer is actually phrased;
the deep structure is a cost/benefit description of the surface structure from
the point of view of each participant. The deep structure of the offer incor-
porates the information shown in Table 2 (or 3, if one party is “baseline
defrauding”). A prediction of this computational analysis is that these cost/
benefit structures are the descriptions from which participants construct
paraphrases and reconstruct the course of the interaction from memory.
Inference procedures for detecting cheaters must operate on a cost/
benefit description of the contract from the potential cheater’s point of view.
These procedures must allow one to quickly and effectively infer that in-
dividual X has cheated when one sees that X has accepted B(X) but not paid
C(X). When a transaction has not yet been completed, or when one’s in-
formation about a transaction is incomplete, “look for cheaters” procedures
should lead one to:
1. Ignore individual X if X has not accepted B(X).
2. Ignore individual X if X has paid C(X).

” However, one might predict that an offer phrased in terms of the potential acceptor’s value
system might sound more attractive, indicating that the offerer really understands (has a good
model of) what the potential acceptor wants.
86 L. Cosmides and J. Tooby

3. Watch out for individual X if X has accepted B(X).

4. Watch out for individual X if X has not paid C(X).

In situations 1 and 2, individual X cannot possibly have cheated; in situations

3 and 4, individual X can cheat. One keeps an eye on X in situation 3 to
make sure X fulfills his or her obligation by paying C(X). One keeps an eye
on X in situation 4 to make sure X does not abscond with B(X), to which
he or she is not entitled.
Most people would not guess that the structure of a simple, straight-
forward social exchange is as complex as described here. But then, this
illusion is a prediction of the theory. People usually do not realize how
complex the grammar of their language is, yet they produce grammatical
sentences with ease. Similarly, people do not realize how complex engaging
in social exchange is, yet they do it with ease. Both parties implicitly un-
derstand and act on all the relations involved because both possess the same
Darwinian algorithms for reasoning about social exchange. These algorithms
perform these computations reliably and automatically, thereby shielding us
from awareness of the underlying complexity, and leaving us with the impres-
sion that what it takes to perform a social exchange is simple and self-
evident-an activity that does not require any extraordinary computations.

PART 4. EMPIRICAL EVIDENCE SUGGESTS THAT

HUMANS DO, IN FACT, HAVE A “LOOK FOR
CHEATERS” PROCEDURE

Whether the human cognitive architecture contains an array of special pur-

pose, domain specific, procedure-rich modules, or consists entirely of a few,
major, domain general information-processing mechanisms, is still at issue
in modern cognitive science. To date, this debate (involving such issues as
learnability, innateness, and so on) has been conducted primarily within the
fields of psycholinguistics and perception, and has left most other subfields
of cognition largely untouched. Human reasoning, especially, has tradition-
ally been considered domain general: The innate processes hypothesized-
whether “logical,” “inductive,” or associationistic-have been thought of
as operating uniformly, regardless of content, with content-dependent per-
formance attributed to the vagaries of differential experience.
However, the evolutionary approach to cognition makes very different
predictions. Such an approach predicts that when a category of content is
thematically linked to important recurrent adaptive problems, specialized
procedures for dealing with that category of content will be invoked, and
they will operate on that content differently than on other categories of
content. Social exchange is a domain for which the evolutionarily-predicted
computational theory is complex, and the fitness costs associated with “er-
rors” are large. Successfully conducted social exchange was such an im-
portant and recurrent feature of hominid evolution that selection would have
 A Computational Theory of Social Exchange 87

favored a reliable, efficient cognitive capacity specialized for reasoning

about social exchange. A general-purpose learning mechanism that operated
on all kinds of content indiscriminately is necessarily more inefficient at
specialized problems: Among other things, being initially “ignorant,” it will
make costly errors and will continue to do so throughout its learning phase.
(More tellingly, no one has even been able to propose a general “design”
that could “learn” how to conduct social exchange, and it remains an open
question whether such a system is possible.) Presumably, the costly trade-
offs necessary to make a learning system general would put it at a selective
disadvantage, and accordingly a phylogenetically antecedent general learn-
ing system presumably would either be supplanted by a domain-specific
system when social exchange became significant in the hominid lineage or
would be used primarily for learning in other less structured domains.
The computational theory of social exchange discussed above, derived
from the evolutionary theory of reciprocation (Trivers 1971; Axelrod and
Hamilton 1981), provides specific predictions about the human psyche,
including:
1. The human psyche includes specialized cognitive algorithms that govern
how people reason about social exchange;
2. These algorithms to operate on item-independent, cost/benefit represen-
tations of exchange interactions; and
3. These algorithms include specialized procedures that are efficient at de-
tecting potential cheaters, cheating being defined with respect to the
grammar of social contracts (Part 3).
Cosmides (1985) and Cosmides (in press) conducted a series of exper-
iments designed to test for the existence in the human psyche of these hy-
pothesized algorithms, using a test of logical reasoning known as the Wason
selection task. These experiments are briefly reviewed below.
In psychology, the study of human reasoning started from the premise
that humans reason logically, i.e., in accordance with the rules of inference
of the propositional calculus (Wason and Johnson-Laird 1972). These rules
of inference are content-independent: they generate only true conclusions
from true premises, regardless of what the propositional content of the prem-
ises is.
However, more than a decade of research has shown that people rarely
reason according to these canons of formal logic. Moreover-and contrary
to initial expectations-psychologists found that human reasoning is con-
tent-dependent: the subject matter one is asked to reason about seems to
regulate how people reason. This can be seen very clearly in experiments
using the Wason selection task (Wason 1966), a test of logical reasoning in
which one is asked to determine whether a conditional rule has been violated
(see Fig. 3). In general, performance on the Wason selection task is very
poor, when the standard of “correct” is the propositional calculus. How-
ever, the content of a few rules elicits a high percentage of “logical” re-
a Abstract Problem (AP)
Part of your new clerical job at the local high school is to make sure that student documents have been processed correctly. Your job is 10
make sure the documents conform to the following alphanumeric rule:

“II a person has a ‘D’ rating, then his documents must be marked code ‘3’.”
(I( P then Q 1’

You suspect the secretary you replaced did not catcgorirc the students’ documents correctly. The cards below have inlormation about the
documents o( (our people who are enrolled at this high school. Each card represents one person. One side of a card tells a person’s letter rating and the
other side of the card tells that person’s number code.

Indicate only those card(s) you definitely need to turn over to see if the documents of any of these people violate this rule.

.... ..a*...... . . . . . . . . . . . . . . . . . . . . . . . . ..I. . ..*..........

. . . . .
. . . .
. D ; . F ; . 3 ; . 7 i
. .
. . . ; .
..*.**........ . . . . . . . . . . . . . ;. .. . ...*...,.,,. *.............
(0 (not-P) WI (not-Q)

6. Drinking Age Problem (DAP; adapted from Criggs k Cox, 1982) ]

In its crackdown against drunk drivers, Massachusetts law enforcement officials are revoking liquor licenses lelt and right. You are a
bouncer in a Boston bar, and you11 lose your job unless you enlorcc the following law:

“I( a person is drinking beer, then he must be over 20 years old.”

(I( P then Q )
The cards below have information about four people sitting al a table in your bar. Each card represents one person. One side of a card tells
what a person is drinking and the other side of the card tells that person’s age.

Indicate only those card(s) you definitely need to turn over to see if any of these people are breaking this law.

..*......*..,. ... ..a........ .*......,...., . . . ..*........

. . . . . . .
i drinking beer i i drinking coke i i 25 years old i i I6 years old i
.s............. . . . . . .
a............. . ..a..*..*.... . . . . . ..*......
(P) (not-P) (Q) (not-Q)
 c. Structure 01 Social Contract (SC) Problems
It is your job to enforce the lollowing law:

Rule I- Standard Social Contract (STD-SC): “If you take the benefit. then you pay the ~0%”
(If P then Q )
Rule 2 - Switched Social Contract (SWC-SC): “If you pay the cost, then you take the benefit.”
(If P then Q )
The cards below have information about iour people. Each card rcprcsrn\s one person. One side of a card tells whether a person accepted
the benefit and, the other side al the card tells whether that person paid the cost.

Indicate only those card(s) you definitely need IO turn over to see if any of these people arc breaking this law.

.,.,..*.....,.. .. . . . . . . . . . . . . . .. . . . . . . . . . . . .. .,,...........
.
. Benefit : . Benefit : . cost ; cost ;
. .
.. Accepted : : NOT Accepted : . Paid -. : NOT Paid .
.... .......... .*..**..*..... . :......,...... ..............
Rule I - STD-SC: (P) (not-P) (not-QJ
Rule 2 - SWC-SC: (Q) (not-Q) (not-P)

FIGURE 3. Content effects on the Wason selection task. The logical structures of these three Wason selection tasks are identical;
they differ only in propositional content. Regardless of content, the logical solution to all three problems is the same: to see if the
rule has been violated, choose the P card (to see if it has a not-Q on the back), and choose the not-Q card (to see if it has a P on
the back). This is because only the combination on the same card of a true antecedent (P) with a false consequent (nor-Q) can
violate-and thereby falsify-a conditional rule. Therefore, the logically correct response is P and not-Q.
Only 4-25% of college students choose both these cards for abstract problems (see panel a). The most common responses are
P and Q, or P: subjects rarely see the relevance of the not-Q card. Yet about 75% see its relevance for the Drinking Age Problem
(panel 6)-a familiar “standard” social contract (see panel c, Rule l)-and choose both P and not-Q. Poor performance on the
abstract problem is not due solely to the use of “abstract” symbols; similar rates of responding are usually found for a number of
more familiar, “thematic” conditionals: relations between food and drink, cities and means of transportation, schools and fields of
study.
Panel c shows the abstract structure of a social contract problem. A “look for cheaters” procedure would lead a subject to
choose the “benefit accepted” card and the “cost not paid” card, regardless of which logical categories they represent. For a
standard social contract, like Rule 1 or the Drinking Age Problem, the correct social contract response--P and not-Q-converges
with formal logic. However, for a switched social contract, like Rule 2, the correct social contract response--not-P and Q-diverges
from formal logic. According to social contract theory, the Drinking Age Problem reliably elicits logically correct responses because
it is a standard social contract, and not because it is familiar. Note: The logical categories (Ps and Qs) marked on the rules and
cards(*) are here only for the reader’s benefit; they never appear on problems given to subjects.
90 L. Cosmides and J. Tooby

sponses (for review, see Cosmides 1985). This effect is known as the “con-
tent effect” on the Wason selection task.
When the content effect on the Wason selection task was first observed,
a number of researchers tried to account for it in terms of differential ex-
perience. Although it is difficult to do justice to the richness of these hy-
potheses briefly, fundamentally they proposed that familiarity (differential
experience) with the rule being tested increases the probability that a subject
will produce the logically correct response (however, different theorists pro-
posed different mechanisms to account for this phenomenon; see, e.g.,
Manktelow and Evans 1979; Griggs and Cox 1982; Johnson-Laird 1982; Pol-
lard 1982; Wason 1983). The problem with these hypotheses was that some
familiar content seemed to produce the content effect, whereas other familiar
content did not.
Cosmides (1985) reinterpreted the existing literature by pointing out that
virtually all of those few familiar rules that did produce a robust and repl-
icable content effect happened to have the cost/benefit structure of a social
contract, as described in Part 3: They could be translated as “If you take
the benefit, then you pay the cost.” Moreover, in reasoning about these
rules, subjects behaved as if they were “looking for cheaters”: They in-
vestigated persons who had accepted benefits (to see if they had failed to
pay the required cost) and persons who had failed to pay the required cost
(to see if they had illicitly absconded with the benefit). For “standard” social
contract rules, such as the ones that were tested, the “benefit accepted”
card and the “cost not paid” card happen to correspond to the logical cate-
gories P and not-Q, respectively (see Figure 3, panel C, Rule I). This means
that a subject who is looking for cheaters will choose the two cards that
correspond to the logically correct response, P and not-Q, by coincidence,
because of the accidental correspondence between the logical and social
contract categories. This accounts for why subjects appeared to reason log-
ically about standard social contract rules, but not about familiar rules that
were not social contracts.
However, all the standard social contract rules tested were familiar. To
experimentally determine whether the content effect is due to the hypoth-
esized “look for cheaters” procedure, rather than to familiarity, Cosmides
tested unfamiliar rules that either did or did not correspond to social
contracts.
Cosmides took highly unfamiliar rules, such as “If a man eats cassava
root, then he has a tattoo on his face,” and embedded them in two different
contexts. One context transformed the rule into a standard social contract,
by telling the subject that cassava root was a rationed benefit and that having
a facial tattoo was a cost to be paid. The other context made the rule describe
some (non-social contract) aspect of the world (e.g.. men with tattoos live
in a different place than men without tattoos; cassava root grows only where
the men with tattoos live; so maybe men are simply eating foods that are
most available to them). If the content effect is due to familiarity. then both
 A Computational Theory of Social Exchange 91

problems should yield low levels of logically correct responses, because both
test unfamiliar rules. However, if the content effect is due to the presence
of a “look for cheaters” procedure, then the unfamiliar social contract prob-
lem should yield high levels of logically correct responses, because for stan-
dard social contracts, the “benefit accepted” card and the “cost not paid”
card happen to correspond to P and not-Q, the logically correct response.
This is, in fact, what happened: while only 23% of subjects chose P and not-
Q for the unfamiliar descriptive problems, 73% made this response to the
unfamiliar standard social contracts. Moreover, the unfamiliar social con-
tract problems elicited more logically correct responses than familiar de-
scriptive problems did (the social contract effect was about 50% larger than
the effect that familiarity had on descriptive problems).
These experiments eliminated the hypothesis that familiarity alone can
account for the content effect on the Wason selection task. Furthermore,
they showed that when a rule has the cost/benefit structure of a social con-
tract, subjects are very good at “looking for cheaters,” even when the sit-
uation they are reasoning about is unfamiliar and culturally alien.
To eliminate the hypothesis that social contract content somehow en-
hances logical reasoning, Cosmides next tested unfamiliar social contracts
that were swirched: These are rules that translate to, “If you pay the cost,
then you take the benefit.” If social contract content causes subjects to
reason logically, then they would choose the logically correct response, P
and not-Q, for switched social contracts, just as they did for standard ones,
even though these cards correspond to individuals who could not possibly
have cheated (see Fig. 3, panel C, Rule 2). However, if reasoning on social
contract rules is guided by a “look for cheaters” procedure, then subjects
would choose not-P and Q, a response completely at variance with formal
logic. This is because for a switched social contract, the “cost not paid”
card corresponds to the logical category not-P, and the “benefit accepted”
card corresponds to the logical category Q (see Fig. 3, panel C, Rule 2). A
“look for cheaters” procedure should be blind to logical category: It should
cause the subject to choose the “benetit accepted” card and the “cost not
paid” card, regardless of their logical category, because these are the cards
that represent potential cheaters.
This prediction was also confirmed. The switched social contracts elic-
ited the “look for cheaters” response, not-P and Q. from 71% of subjects,
even though this response is illogical according to the propositional calculus.
In comparison, the unfamiliar descriptive problems (i.e., those not depicting
a social contract) elicited this illogical response from only 2% of subjects,
and elicited the logically falsifying response from 14.5%.
In the above experiments, social contract rules were pitted against de-
scriptive rules. However, in a further set of experiments, social contract
rules were pitted against “permission” rules that lacked the cost-benefit
structure of a social contract. Cheng and Holyoak (1985) had proposed that
the modern individual social experience of permissions causes people
92 L. Cosmides and J. Tooby

(through an unspecified mechanism employing some kind of induction) to

build a cognitive schema that would produce the same pattern of responses
we term “the social contract effect.” In their theory, permission rules have
the format, “If an action is to be taken, then the precondition must be sat-
isfied.” To test between the two theories, the rules tested were all given the
permission format, but, as before, they were embedded in different contexts.
One kind of context portrayed taking the action as a benefit and satisfying
the precondition as a cost, thereby transforming the rule into a social con-
tract, whereas the kind of other context portrayed a permission situation,
but one that was not a social contract (see Cosmides, in press).
Seventy-five percent of subjects gave the correct social contract answer
(standard: P and not-Q; switched: not-P and Q) in response to the social
contract problems, compared to only 21% of subjects for the non-social
contract permission problems. This result supports the hypothesis that hu-
mans represent social exchange in item-independent, cost/benefit terms. The
“look for cheaters” procedure is not activated by prescriptive or “permis-
sion” rules that lack an easy or obvious cost/benefit interpretation. In con-
trast, even a highly unfamiliar rule can activate the “look for cheaters”
procedure, as long as the context in which it occurs allows the subject to
map the cost/benefit valuations that the participants to the exchange assign
to the items mentioned in the rule.
The results of these experiments show that when subjects reason about
situations involving social exchange, their responses follow the distinctive
“adaptive logic” predicted for the “look for cheaters” procedure (described
in Part 3), and that this procedure is activated by item-independent, cost/
benefit representations of exchange interactions. The results are inconsistent
with alternative theories of human performance on logical reasoning tasks,
including theories proposing that humans reason in accordance with formal
logic, experience-based associational theories, and induction-based theories.
The finding that adult subjects are very adept at detecting potential
“cheaters” on a social contract, even when it is unfamiliar and culturally
alien, stands in marked contrast to the repeated finding that they are not
skilled at detecting potential violations of descriptive and permission rules,
even though such rules are commonly encountered in everyday life. The
ontogeny of the algorithms that produce these results remains an open ques-
tion. It is possible that they are, in some carefully delimited sense, learned.
However, the mental processes involved appear to be powerfully structured
for social contracts, yet weakly structured for other elements and relations
drawn from common experience. This implies that the “look for cheaters”
procedure is either itself innate, or else the product of a learning process
that is guided and structured by innate algorithms that are specialized for
reasoning about social exchange.

CONCLUSION

Although adaptive participation in social exchange depends upon the correct

solution of complex, highly structured, and highly specialized information
 A Computational Theory of Social Exchange 93

processing problems, humans in all cultures and of virtually all ages both
understand and successfully participate in such social exchanges with ease.
The costs and benefits of participation in social exchange have also consti-
tuted an intense selection pressure over a significant fraction of hominid
evolutionary history. It is implausible to expect that natural selection would
leave learning in such a domain to the vagaries of personal experience pro-
cessed through some kind of general-purpose learning mechanism. The ev-
olutionary expectation that humans do indeed have adaptively structured
social exchange algorithms receives substantial empirical support from ex-
perimental investigations of human reasoning that 1) have falsified the com-
peting domain general theories of reasoning performance on the Wason se-
lection task and 2) have confirmed the presence of the evolutionarily
predicted complex of design features that are diagnostic of adaptation in this
domain. We argue that this study of social exchange provides an example
of how evolutionary and cognitive techniques can be combined to elucidate
aspects of human culture and the psychological mechanisms that underlie
them.
It is worth noting that the kinds of “domain general” theories that were
falsified as explanations for performance on the Wason selection task are
the same kinds of theories that are more generally advanced to account for
the human “capacity” for culture (Sahlins 1976; Geertz 1973). Because no
imaginable state of human affairs is forbidden by such domain general views
of culture, there is little in the way of cultural phenomena that is inconsistent
with such models, and consequently very little that is predicted or illumi-
nated by them. The view that cultures are arbitrary symbolic productions
is widely and justly criticized by advocates of evolutionary approaches. But
by using evolutionary psychology, it is possible to go further and meet tra-
ditional anthropological theories of culture on their own ground.
As the process of identifying and mapping these innate mechanisms
proceeds, mechanism by mechanism, the differing domains of culture each
can be analyzed through reference to the highly structured information pro-
cessing algorithms that govern its expression. The “interpretation of cul-
tures” can be changed from a post hoc literary exercise about arbitrary
symbolic productions into a principled investigation grounded in the evolved
psychology of humans and its systematic impact on the cultures it produces.

We are very indebted to Jerome Barkow, Martin Daly, Irven DeVore, Roger Shepard, Donald
Symons, and Margo Wilson for many stimulating discussions of the issues explored in this
paper. We are especially grateful to Jerome Barkow, Nicholas Blurton Jones, Michael McGuire,
and an anonymous reviewer for their comments on the various drafts, and to Jason Banfield
and Lisa Bork for their help with the manuscript. We would also like to thank Jeff Wine and
Roger Shepard (and NSF grant no. BNS 85-l 1685 to Roger Shepard) for their support.

APPENDIX 1. PHILOSOPHICAL REFINEMENTS

(CATEGORIZED BY CLAUSE)

2. In other words, the cost/benefit requirements do hold for me and I believe

that they hold for you. (Note: sincere cost/benefit requirements entail “I
94 L. Cosmides and J. Tooby

value getting P from you more than I value keeping Q,” so this need not
be added as a separate statement.) Clause 2 is an implication of my offer
even if the sincere cost/benefit requirements do not hold. After all, baseline
defrauders mean their offers to be thought sincere.
3. “. . . and I intend that you realize . . .” In other words, I did not make
the offer accidentally. My having made the offer is a consequence of the
activation of my social contract algorithms (my belief that the contract would
result in a net benefit to me is a necessary condition for my making the offer;
see discussion of the meaning of “cause ” in clause 5). If my social contract
algorithms had not been activated, I would not have made the utterance.
This is presumed for a contract that is offered verbally-there are virtually
no circumstances under which one can accidentally utter a sentence. How-
ever, for nonlinguistic primate species, one can imagine scenarios in which
“gestures” are accidentally produced. For example, in the course of a fight,
a chimp is chased up a tree. The tree limb supporting him breaks, causing
him to fall with his arm stretched out. An outstretched arm in the context
of a fight is usually a request for support. However, this gesture was made
accidentally rather than intentionally; it was not made as a consequence of
the chimp’s social contract algorithms having been activated. Therefore,
“ . . . I intend that you realize . . .” is not part of the gesture’s meaning.
The fact that it was “accidentally” produced robs the “gesture” of its mean-
ing as a request for support.
5. My belief that you have given me P cannot cause me to give you Q in
just any old way. For example, the following is not the sense of causation
meant:
Let’s say you own a priceless statue, and I have some very compromising
pictures of you that you want destroyed. I keep these pictures in my car.
I make the offer “If you give me the statue (P), then I’ll destroy the pictures
(Q).” You agree, unaware that I have no intention of destroying the pictures
because I want to continue to enrich myself by blackmailing you. We ar-
range for you to leave the statue at a drop point. I retrieve it, and my belief
that you have given me this priceless statue makes me so agitated and
nervous that I have an accident, and the car blows up, destroying the pic-
tures. I have, in fact, done Q, and my belief that you gave me P caused
me to give you what you wanted-Q-but not in the right sense of “cause.”
(e.g., Nozick 1981, p. 369)
The correct notion of “cause” refers to the psychological realization
of (the algorithm instantiating) this computational theory and the fact that
it is guiding my behavior. My belief that you have given me P fills in the
parameter value in the algorithm; this triggers the set of procedures within
the algorithm corresponding to the contract’s conditions of satisfaction. Trig-
gering these procedures results in my giving you Q. This is the same sense
of “cause” as in a computer program: the information that P can cause a
computer to do something by virtue of that information’s functional relation
to various of its procedures. Let’s say I have written a program in Basic
instantiating all the conditions for making a social contract. The program
 A Computational Theory of Social Exchange 95

then offers-“If you type ‘P’ into me then 1’11print ‘Q’ for you”-and 1
accept. Part of the program would involve the computer waiting for me to
fulfill my obligation, and this part may be written thus:

10 Input “Now give me P”;A$

20 If A$ = P then go to 40
30 go to 10
40 Print “Q”

My typing P gives the variable A$ the parameter value P (analogous (?> to

the computer believing that I have typed ‘P’ into it), and this causes the
computer to print ‘e’. The same sense of cause is meant in clause 7.

REFERENCES
Axelrod, R. The Evolution of Cooperation. New York: Basic Books, 1984.
-, and Hamilton, W.D. The evolution of cooperation. Science 211: 1390-1396, 1981.
Bahrick, H.P., Bahrick, P.O., and Wittlinger, R.P. Fifty years of memory for names and faces:
A cross-sectional approach. Journal ofExperimental Psychology 104: 54-75, 1975.
Bartlett, F.C. Remembering: A Study in Experimental and Social Psychology. Cambridge,
U.K.: Cambridge University Press, 1932.
Buss, D. Sex differences in human mate selection criteria: An evolutionary perspective. In
Sociobiology and Psychology, C. Crawford, D. Krebs, and M. Smith (Eds.). Hillsdale,
NJ.: Erlbaum, 1987.
Carey, S., and Diamond, R. Maturational determination of the developmental course of face
encoding. In Biological Studies of Menral Processes, D. Caplan (Ed.). Cambridge,
Mass., The MIT Press, 1980.
Cheng, P., and Holyoak, K. Pragmatic reasoning schemas. Cognitive Psychology 17: 391-416
1985.
Chomsky, N. Reflections on Language. New York: Random House, 1975.
Cosmides, L. Invariances in the acoustic expression of emotion during speech. Journal of
Experimental Psychology. 9: 864-881, 1983.
- Deduction or Darwinian algorithms?: An explanation of the “elusive” content effect on
the Wason selection task. Ph.D. diss., Harvard University, University Microfilms,
1985.
-The logic of social exchange: Has natural selection shaped how humans reason? Cognition,
in press.
-, and Tooby, J. From evolution to behavior: Evolutionary psychology as the missing link.
In The Latest on the Best: Essays on Evolution and Opfimality, J. Dupre (Ed.). Cam-
bridge, Mass.: The MIT Press, 1987.
Cutting, J.E., Protitt, D.R., and Kozlowski, L.T. A biomechanical invariant for gait perception.
Journal of Experimental Psychology 4: 357-372, 1978.
Dawkins, R. The Extended Phenotype. San Francisco: W.H. Freeman, 1982.
de Waal, F. Chimpanzee Politics: Power and Sex Among Apes. New York: Harper and Row,
1982.
Eibl-Eibesfeldt, I. Ethology: The Biology of Behavior, (2nd ed.). New York: Holt, Rinehart
and Winston, Inc., 1975.
Ekman, P. (Ed.) Emotion in the HumanFace, 2nded., Cambridge, U.K.: Cambridge University
Press, 1982.
Fillenbaum, S. Inducements: On the phrasing and logic of conditional promises, threats, and
warnings. Psychological Research 38: 231-250, 1976.
96 L. Cosmides and J. Tooby

Fodor, J.A. Modularity of Mind, Cambridge, Mass.: MIT Press, 1983.

Gardner, H. The Sharrered Mind, New York: Random House, 1974.
Geertz, C. The Inrerpretation of Cuhures, New York: Basic Books, 1973.
Gleitman, L.R., and Wanner, E. Language acquisition: The state of the state of the art. In
Language Acquisition: The State of the Art, E. Wanner and L.R. Gleitman (Eds.).
Cambridge, U.K.: Cambridge University Press, 1982.
Goodall, J. van Lawick. The behaviour of free-living chimpanzees in the Gombe Stream Re-
serve. Animal Behaviour Monograph 3, 1968.
- In the Shadow of Man, Boston: Houghton-Mifflin, 1971.
Grice, H.P. Meaning. Philosophical Review 66: 377-388, 1957.
-Logic and conversation. Unpublished William James Lectures, Harvard University, 1967.
Griggs, R.A., and Cox, J.R. The elusive thematic-materials effect in Wason’s selection task.
British Journal of Psychology 73: 407-420, 1982.
Hall, K., and DeVore, I. Baboon social behavior. In Primare behavior, I. DeVore (Ed.). New
York: Holt, 1965.
Hamilton, W.D. The genetical evolution of social behaviour. Journal of Theoretical Biology 7:
l-52, 1964.
Hrdy, S. Blaffer. The Woman that Never Evolved, Cambridge, Mass.: Harvard University
Press, 1981.
Issac, [Link]. The food-sharing behavior of protohuman hominids. Scientific American 238: 90-
108, 1978.
Johnson-Laird, P.N. Thinking as a skill. Quarrerly Journal of Experimental Psychology 34A:
l-29, 1982.
Jolly, A. The Evolution of Primate Behavior, New York: Macmillan, 1972.
Kinzey, W.G. (Ed.) Primate Models for the Origin of Human Behavior, New York: SUNY
Press, 1987.
Kozlowski, L.T., and Cutting, J.E. Recognizing the sex of a walker from a dynamic point-light
display. Perception and Psychodynamics 21: 575-580, 1977.
Lee, R.B. and DeVore, I. (Eds.) Man the Hunter, Chicago: Aldine.
Lute, R.D., and Raiffa, H. Games and Decisions: Introduction and Critical survey, New York:
Wiley, 1957.
Manktelow, K.I., and Evans, J. St B.T. Facilitation of reasoning by realism: Effect or non-
effect? British Journal of Psychology 70: 477-488, 1979.
Mat-r, D. Vision: A Computational Investigation into the Human Represenration and Processing
of Visual Information. San Francisco: Freeman, 1982.
-, and Nishihara, H.K. Visual information processing: Artificial intelligence and the sen-
sorium of sight. Technology Review: 28-49, October 1978.
Maynard Smith, J. 1982. Evolufion and the Theory of Games, Cambridge, U.K.: Cambridge
University Press.
McCracken, G.F., and Bradbury, J.W. Social organization and kinship in the polygynous bat
Phyllostomus hastatus. Behavioral Ecology and Sociobiology 8: 1l-34, 1981.
Nozick, R. Philosophical Explanations, Cambridge, Mass.: Harvard University Press, 1981.
Owens, I., Bower, G.H., and Black, J.B. The “soap opera” effect in story recall. Memo01 and
Cognition 7: 185-191, 1979.
Pilbeam, David, Department of Anthropology, Harvard University, personal communication.
Pollard, P. Human reasoning: Some possible effects of availability. Cognition 10: 65-96, 1982.
Quine, W.V.O. Ontological Relativity and Other Essays, New York: Columbia University
Press, 1969.
Sahlins, M. The Use and Abuse of Biology, Ann Arbor: University of Michigan Press, 1976.
Schank, R., and Abelson, R.P. Scripts, Plans, Goals, and Understanding. Hillsdale, N.J.:
Erlbaum, 1977.
Searle, J.R. (Ed.) The Philosophy oflanguage, Oxford: Oxford University Press, 1971.
Smuts, B. Special relationships between adult male and female olive baboons (Pupio anubis).
Ph.D. diss., Stanford University, 1982.
Strum, S.C. Baboons may be smarter than people. Animal Kingdom 88: 12-25, 1985.
Tooby, J. Prospects for an evolutionary psychology. Unpublished manuscript, Harvard Uni-
versity, 1975.
- The emergence of evolutionary psychology. In Emerging Syntheses in Science, D. Pines
(Ed.). Santa Fe: Santa Fe Institute, 1985.
 A Computational Theory of Social Exchange 97

-, and Cosmides, L. Evolutionary psychology and the generation of culture, part I. Theo-
retical considerations. Ethology and sociobiology 10: 29-49, 1989.
-, and - The evolution of revenge. In preparation, a.
-, and - Evolution and cognition. In preparation, b.
-, and DeVore, I. The reconstruction of hominid behavioral evolution through strategic
modeling. In Primate Modelsfor the Origin of Human Behavior, W.G. Kinzey (Ed.).
New York: SUNY Press, 1987.
Trivers, R.L. The evolution of reciprocal altruism. Quarterly Review of Biology 46: 35-57,
1971.
Wall Street Journal. Prestige cards: For big bucks and big egos. April 17, 1985, p. 35.
Wason, P.C. Reasoning. In New Horizons in Psychology, B.M. Foss (Ed.). Harmondsworth,
U.K.: Penguin, 1966.
- Realism and rationality in the selection task. In Thinking and Reasoning: Psycho/ogical
Approaches, J. St B.T. Evans (Ed.). London: Routledge and Kegan Paul, 1983.
-, and Johnson-Laird, P.N. Psychology of Reasoning: Structure and Content, London:
Batsford, 1972.
Wilkinson, G.S. Reciprocal food sharing in the vampire bat. Nature 308: 181-184, 1984.
Williams, G.C. Adaptation and Natural Selection: A Critique of Some Current Evolutionary
Thought, Princeton, N.J.: Princeton University Press, 1966.
Wrangham, R.W. War in evolutionary perspective. In Emerging Syntheses in Science, D. Pines
(Ed.). Santa Fe: Santa Fe Institute, 1985.