Neuron, Vol. 35, 395­405, July 18, 2002, Copyright 2002 by Cell Press
A Neural Basis for Social Cooperation
expertise, information, opportunities, and a host of ma-James K. Rilling,1,2
David A. Gutman,
Thorsten R. Zeh, Giuseppe Pagnoni, terial resources.
On the other hand, cooperation based on reciprocalGregory S. Berns, and Clinton D. Kilts
Department of Psychiatry and Behavioral Sciences altruism is rare in the rest of the animal kingdom and
has only been convincingly demonstrated for a handfulEmory University
Atlanta, Georgia 30322 of species (Axelrod and Hamilton, 1981; Ridley, 1996;
Trivers, 1971). In attempting to provide an explanation
for its scarcity, evolutionary biologists have theorized
that two or more preconditions must be satisfied forSummary
reciprocal altruism to evolve in a species: (1) individuals
must interact repeatedly with social partners over theCooperation based on reciprocal altruism has evolved
course of their lifetime, and (2) individuals must be ablein only a small number of species, yet it constitutes
to recognize conspecifics and discriminate againstthe core behavioral principle of human social life. The
those who do not reciprocate altruism (Axelrod anditerated Prisoner's Dilemma Game has been used to
Hamilton, 1981; Trivers, 1971). As an additional precon-model this form of cooperation. We used fMRI to scan
dition, there must also be a mechanism that enables36 women as they played an iterated Prisoner's Diindividuals
to inhibit the temptation to accept but notlemma Game with another woman to investigate the
reciprocate altruism; a mechanism that weights long-neurobiological basis of cooperative social behavior.
term rewards and punishments over immediate andMutual cooperation was associated with consistent
transient, short-term gains (Frank, 1988). Only with suchactivation in brain areas that have been linked with
a mechanism can the long-term benefits of sustainedreward processing: nucleus accumbens, the caudate
mutual cooperation be realized.nucleus, ventromedial frontal/orbitofrontal cortex, and
The iterated Prisoner's Dilemma Game has been usedrostral anterior cingulate cortex. We propose that actiby
investigators from a wide range of disciplines tovation of this neural network positively reinforces remodel
social relationships based on reciprocal altruismciprocal altruism, thereby motivating subjects to resist
(Axelrod and Hamilton, 1981; Axelrod, 1984; Boyd, 1988;the temptation to selfishly accept but not reciprocate
Nesse, 1990; Trivers, 1971). To elucidate the neural sub-favors.
strates of the emotional and cognitive processes that
support cooperative, reciprocally altruistic relationships,Introduction
we investigated game-related neural activations with
fMRI as subjects played an iterated Prisoner's DilemmaEvolutionary biologists have long theorized about how
Game with other subjects outside the scanner. In thisaltruistic behavior can exist, given that natural selection
game, two players independently choose to either coop-is based primarily on the differential survival and reproerate
with each other or not, and each is awarded a sumductive success of individual organisms rather than
of money that depends upon the interaction of bothgroups of organisms. W.D. Hamilton's kin selection theplayers'
choices in that round. There are four possibleory nicely accounts for altruism among relatives (Trivers,
outcomes of a round: player A and player B cooperate1985). But cooperation among nonrelatives is pervasive
(CC), player A cooperates and player B defects (CD),in human society and must also be explained. Reciprocplayer
A defects and player B cooperates (DC), or playerity, including both direct and indirect reciprocity, has
A and player B defect (DD). The payoffs for the outcomesbeen proposed to account for altruism toward nonrelaare
arranged such that DC CC DD CD, and CCtives. In direct reciprocity, individuals dispense favors,
(CD DC)/2. Each cell of the payoff matrix (Figure 1A)and these favors are likely to be returned by the recipicorresponds
to a different outcome of a social interac-ent, in one form or another, in times of future need
tion. DC represents the situation where player A opts(Sahlins, 1972; Trivers, 1971). In indirect reciprocity, the
for noncooperation and player B cooperates so thatfavor is returned by a third party (Nowak and Sigmund,
player A benefits at player B's expense. CD is the con-1998). This study sought to define the neural basis of
verse. CC involves mutual cooperation, and DD involvesdirect reciprocity.
mutual noncooperation.The paradigmatic example of reciprocal altruism is
In two separate experiments, we scanned a total of 36food sharing, which human beings engage in far more
women with fMRI as they played the Prisoner's Dilemmadeliberately and pervasively than any other species (RidGame.
Experiment 1 was designed to isolate the neuralley, 1996), and which was almost certainly essential to
correlates of cooperation and noncooperation in socialthe survival of our hominid ancestors in their African
and nonsocial contexts, and of monetary reinforcementsavannah niche (Lee and DeVore, 1968). This deeply
of behavior. Nineteen subjects were scanned duringingrained tendency manifests itself in myriad ways in
each of four game sessions. The results of the first ex-modern human social life, including the exchange of
periment revealed different patterns of neural activation
depending on whether the playing partner was identified1
Correspondence: jrilling@princeton.edu
as a human or a computer. This motivated a second2
Present address: Green Hall, Princeton University, Princeton, New
Jersey 08544. experiment in which 17 subjects were scanned during
Neuron
396
(Figure 2A). This pattern of behavior, switching to defection
as the end of the game approaches, has been predicted
on theoretical grounds (Axelrod, 1984) and observed
empirically in previous studies (Andreoni and
Miller, 1993). In games played with the provocative human
confederate, the frequency of mutual cooperation
was lower and mutual defection higher (Figure 2B). Because
the "tit-for-tat" computer strategy initiated the
game with defection, mutual cooperation was uncom-Figure 1. Study Design
mon in early rounds but rebounded to levels observed(A) Payoff matrix used for the four outcomes in the Prisoner's Diwith
the unconstrained human partner as the game pro-lemma Game. Scanned subjecťs choices (C or D; player A) are listed
atop columns and nonscanned subjecťs choices (C or D; player B) gressed before declining sharply on the very last round
are listed aside rows. Dollar amounts in bold are awarded to player (Figure 2C).
A. Amounts in parentheses are awarded to player B.
As for Experiment 1, mutual cooperation was the most
(B) Time course of a single round of the Prisoner's Dilemma Game.
common outcome in games played with presumed human
partners for Experiment 2 (Figures 2D and 2E). In
these games, the observed reduction in mutual cooperaeach
of three game sessions, focusing specifically on
tion in rounds 18­20 was forced by the computer strathuman
versus computer interaction.
egy, which defected automatically in these rounds. The
In both experiments, the players interacted via a netrebound
of mutual cooperation in rounds 20­23 (Figure
worked computer that accepted the responses from the
2D) was induced by programmed cooperation by the
player inside the scanner (player A) and the nonscanned
computer in these three rounds. When subjects in Explaying
partner (player B). Each Prisoner's Dilemma
periment 2 were instructed that they were playing the
game consisted of at least 20 rounds, with each round
game with a computer rather than another person, mulasting
21 s (Figure 1B). During the first 12 s of each
tual cooperation was less common throughout the game
round, both players independently selected either to
(Figure 2F; paired t 4.90, 19 df, p 0.001), even though
cooperate or defect. At 12 s, the square of the matrix
subjects were actually playing against exactly the same
where the two choices intersected was highlighted to
computer strategy.
reveal each player's choice and the resulting payoff for
In both experiments, there was a tendency for subject
that round. Subjects were compensated in direct propairs
who arrived at a CC outcome to persist with mutual
portion to their accumulated earnings. The outcome was
cooperation so that a CC outcome in the current round
displayed for 9 s, and then the next round began (Figure
was most likely to be followed by a CC outcome in the
1B). Functional images were collected every 3 s. We
next (Table 2).
analyzed both the BOLD response to the game outcome
and the BOLD response during the decision-making pefMRI
Datariod of each round. For the former, we examined the
Neural Activations Related to the Reaction to theresponse for the epoch between 12 and 21 s. For the
Game Outcome (Seconds 12­21 of Each Round)latter, we examined the 6 s epoch preceding the button
The BOLD response to a given outcome type (i.e., CC,press signaling a choice to cooperate or defect in each
CD, DC, or DD) could be attributable to an effect ofround.
either the partner's choice, the player's choice, or to an
interaction between the two that exceeded the sum ofResults
their respective main effects. The statistical interaction
is of special interest because it relates specifically toPrisoner's Dilemma Game Behavior
the social interaction rather than to independent effectsThe number of occurrences of each outcome type was
of player and partner decisions. Therefore, we begana function of the two players' choices and so was not
by testing for main effects of player and partner choicesspecified in advance. Table 1 shows the average number
and for an interaction between the two. More specifi-of each outcome type per session for both experiments.
cally, for the games in which player A assumed she wasIn Experiment 1, mutual cooperation was the most
playing with a human partner, we examined the maincommon outcome when the playing partner was a beeffect
of player A's decision (irrespective of player B'shaviorally unconstrained woman (see Experimental Prodecision)
on neural activity in player A during the reac-cedures for details of experimental design). However,
tion epoch (seconds 12­21 for each round), the mainin the final rounds of the game, the frequency of mutual
effect of player B's decision (irrespective of player A'scooperation decreased, and mutual defection increased
decision) on neural activity in player A, and the interaction
effect of player A and player B's decisions on neural
activity in player A.
Table 1. Average Number of Outcome Types per Session, for
Experiments 1 and 2 were analyzed separately. TheSessions with Presumed Human Playing Partners
following procedure was used to identify brain regions
Experiment Partner CC CD DC DD Total
that were activated in both experiments. For a given
1 unconstrained 11.2 2.3 3.2 3.2 20 contrast, we masked the thresholded (p 0.01) t statis1
confederate 6.4 4.6 4.2 4.7 20 tic map for that contrast from Experiment 1 and limited
2 open ended 11.9 3.8 3.6 4 23 our analysis of Experiment 2 to voxels within the mask.
2 closed 9.9 2.8 2.5 5 20
We then calculated the same contrast for Experiment
A Neural Basis for Social Cooperation
397
Figure 2. Round by Round Depiction of the
Proportion of All Subjects Pairs Who Mutually
Cooperated in Three Sessions from Each of
Two Experiments
Results for Experiment 1 are shown for sessions
with (A) unconstrained human playing
partner, (B) provocative confederate playing
partner, (C) computer partner playing "tit-fortat".
Results for Experiment 2 are shown for
sessions with (D) an assumed human partner
with the number of rounds unspecified in advance,
and an assumed (E) human and (F)
computer partner with the number of rounds
specified in advance.
2. Voxels within the mask that survived at p 0.01 were Event-Related Plots
Event-related plots were constructed to determinethen reported as replicated activations.
which outcome or outcomes were responsible for theMain Effects
interaction effect. Plots were made for the peak voxelsNo main effects replicated across both experiments.
in the anteroventral striatum and OFC ROIs for everyInteraction Effects
subject. Across both experiments, for each session withConsistent interaction effects were observed across the
a human partner in which all four outcomes occurredtwo experiments (Table 3). These effects were restricted
(n 61 sessions), we examined which of the four out-to one side of the interaction, namely ([CC DD])
comes had the largest amplitude-fitted response in the[CD DC]) and not the opposite ([CD DC] [CC
general linear model and whether that response wasDD]). That is, for the regions listed in Table 3, the neural
positive or negative. The plot for one subject is shownresponse to game outcomes CC and DD combined was
in Figures 4A and 4B. For the peak voxel of the antero-greater than activation following outcomes CD and DC
ventral striatal ROI, CC had the largest fitted responsecombined. This is of interest because CD and DC outfor
30 of the 61 sessions, whereas DD had the largestcomes are typically aversive to at least one of the two
response for 19 sessions, and both CD and DC had thesubjects and are consequently unlikely to be repeated.
largest response for only 6 sessions each (Figure 4C).On the other hand, CC and DD are more stable in the
This distribution differed significantly from chance (chi-sense that subjects often persist with these outcomes.
square 26.8, p 0.001). In 25 of the 30 sessions whereHence, CC and DD outcomes might be considered beCC
had the largest fitted response, that response washaviorally reinforcing. In terms of spatial extent, the
positive. In other words, CC was associated with in-largest activation for this interaction involving symmetric
creased activation relative to the other conditions, rathersocial behavior is in the anteroventral striatum and subthan
less deactivation. Though not as pronounced, theregenual anterior cingulate cortex (BA 25). The striatal
was also evidence of deactivation for the CD and DCactivation includes the caudate nucleus and nucleus
outcomes at this location (Figure 4C). In 19 of the 23accumbens (Nac), both of which receive midbrain dopacases
where CD had the smallest fitted response, thatmine projections known to be involved with processing
response was negative, and the DC response was nega-reward (Schultz, 1998). The ventromedial/orbitofrontal
tive in 21 of the 23 cases where DC had the smallestcortex (OFC), another brain area involved in reward profitted
response.cessing (Rolls, 1999), was also activated for the interacFor
the peak voxel in the OFC ROI (see Figure 3), CCtion (Figure 3).
had the largest fitted response for 32 of the 61 sessions
(versus 15 for DD, 7 for CD, and 7 for DC), and 24 of
these were positive in amplitude. This distribution alsoTable 2. Transition Probabilities Following CC Outcomes in
differed significantly from chance (chi-square 27.9,Experiments 1 and 2
p 0.001). Thus, for both ROIs, the interaction effect
Experiment Partner CC CD DC DD
was dominated by the positive response to CC.
1 unconstrained human 0.79 0.06 0.11 0.04 CC versus the Other Outcomes
1 confederate human 0.47 0.34 0.13 0.06 Given that the BOLD response to CC was largely respon2
assumed human 0.82 0.11 0.05 0.01
sible for the interaction effect, we decided to focus more
open ended
specifically on this outcome by contrasting the BOLD2 assumed human 0.77 0.14 0.06 0.03
response to the CC outcome with the average responseclose ended
of the other three outcomes combined. Masking the
The probability of each outcome, given a CC outcome in the previous
results of the Experiment 2 with Experiment 1 revealed
round, is listed as a function of experiment and partner type.
larger and more significant activations in ventromedial
Neuron
398
Table 3. Reaction Epoch: Location of Brain Activations in Player A Related to the Interaction of Player B's and Player A's Decision to Cooperate
or Defect
Brain Region Coordinates Peak t Statistic Number Voxels
Player A player B interaction
(CC DD) (CD DC)
R caudate 6 18 0 4.94 8
L post-central gyrus (BA 1/3) 27 39 60 3.86 5
R central sulcus (BA 4) 18 30 72 3.35 7
R medial frontal gyrus (BA 11) 6 51 18 3.26 7
Activations are for Experiment 2 (p 0.01, n 17 subjects) after limiting the search volume to voxels that survived a statistical threshold
p 0.01 in Experiment 1 (n 19 subjects). Activations consisting of fewer than five contiguous voxels are not reported. L, left hemisphere;
R, right hemisphere.
frontal cortex and anteroventral striatum than were next round. In our experiments, CC outcomes tended
to occur in consecutive strings so that a CC outcomefound for the interaction analysis (Table 4; Figure 5A).
In contrast to the interaction t map, the activation in the in one round was most likely to be followed by a CC
outcome in the next (Table 2). Thus, the intervals follow-ventromedial frontal cortex extended dorsally into the
rostral anterior cingulate cortex (BA32). In Figures 5C ing CC outcomes typically involved a decision to continue
cooperating, rather than defect. To more systemat-and 5D, the statistical parametric map for this contrast
is displayed on a spatially normalized EPI image to dem- ically investigate neural activity related to opting for
social cooperation, a model was specified that com-onstrate that the observed ventromedial frontal/orbitofrontal
activation is not within an area of high magnetic pared the BOLD signal in the 6 s interval immediately
preceding the choice to cooperate or defect (as markedsusceptibility artifact. Figure 6 is an event-related plot
for one subject for the peak voxel of the OFC ROI. by a button press), and analyzed as a function of the
partner's decision in the previous round. The four condi-CC Compared with Monetary Reinforcement
To investigate the possibility that this pattern of activa- tions were XC,CX (i.e., choosing to cooperate after the
partner had cooperated in the previous round), XC,DXtion was simply a consequence of monetary reinforcement
($2 for a CC outcome), we tested for a condition (i.e., choosing to defect after the partner had cooperated
in the previous round), XD,CX (i.e., choosing to cooper-(human partner versus control) by monetary outcome
($2 versus others) interaction in Experiment 1. That is, ate after the partner had defected in the previous round,
and XD,DX (i.e., choosing to defect after the partner hadwe asked whether earning $2 when playing with a human
partner produced more activation than earning $2 in defected in the previous round).
The decision to cooperate following a cooperativethe nonsocial control condition. The test for interaction
revealed activation in the anteroventral striatum and choice by one's partner in the previous round activated
the left anterior caudate and the right post-central gyrusOFC (Figure 5E). Thus, the anteroventral striatum, rACC,
and OFC were activated more by reciprocated social (Table 6; Figure 7). The decision to reciprocate cooperation
was also associated with activation in two regionscooperation than by a $2 reward in a nonsocial context.
CC with Computer versus Human that were activated following mutual cooperation in the
reaction epoch: the rostral anterior cingulate cortex andPlaying Partners
Finally, sessions with computer playing partners were the anteroventral striatum (Table 6; Figure 7).
included in both experiments to determine whether activations
detected with human partners were specific to
human social interaction. In both experiments, mutual
cooperation with a computer playing partner activated
regions of the ventromedial/orbital frontal cortex (BA
11) that were also activated with human playing partners
(Table 5), although for Experiment 1, the overlap was
only observed if the t statistic threshold was decreased
to p 0.05. In neither of the two experiments did mutual
cooperation with a computer activate the rostral anterior
cingulate or the anteroventral striatum observed for human
playing partners.
Neural Activation Related to Social Decision
Making (6 s Epoch Preceding the C or D Choice)
Given that subjects make their choices early (mean
Figure 3. Reaction Epoch
3.4 s) within the 12 s decision-making period of each
Activation in player A when playing with an assumed human partner.round, it seems likely that the 9 s period during which
Voxels activated for the interaction of player A and player B's
the game outcome was displayed (and over which neural
choices (CC CD) (DC DD) in Experiment 2 (p 0.01), after
activity was sampled for the reaction epoch) involves masking the results with voxels activated for the same contrast in
not only the reaction to the outcome of the current round Experiment 1 (p 0.01). OFC ventromedial frontal/orbitofrontal
cortex.but also decision making related to the choice for the
A Neural Basis for Social Cooperation
399
Figure 4. Reaction Epoch: Event-Related Plot for the Peak Voxel in the Anteroventral Striatum
(A) Fitted response for all four outcomes in a single subject, CC red, CD blue, DC green, and DD cyan.
(B) Raw data for CC outcome for a single subject. The outcome is revealed at t 0 s and displayed for 9 s.
(C) Distribution of outcome types having the largest and smallest amplitude fitted response in the GLM, across all 61 sessions.
Discussion relate to the rewarding effects of arranging and/or experiencing
a mutually cooperative social interaction.
Recent evidence indicates that reward-related neuralReaction Epoch
Postscan subject interviews revealed that mutual coop- activity is greater for unpredicted than predicted rewards
(Schulz et. al., 1997). Our results are consistenteration was typically considered the most personally
satisfying outcome. The more profitable DC outcome with this observation insofar as subjects exerted no control
over their partners' decisions so that the game out-was typically described as less desirable than CC outcomes
either because it provoked guilt over having prof- come always had an element of unpredictability. A subject
could never know for certain if her cooperativeited at the partner's expense, or because subjects realized
that the outcome would likely provoke defection choice would be reciprocated. However, when subjects
choose to cooperate, they are guessing that their partnerby the partner, thereby destabilizing the relationship and
leading to lower cumulative earnings. Combined with the will do the same; and when their cooperation is met with
defection, an anticipated reward is omitted. Schulz et al.neuroimaging and electrophysiological evidence linking
the orbitofrontal cortex (Francis et al., 1999; O'Doherty (1997) have demonstrated that the omission of expected
rewards deactivates midbrain dopamine neurons (de-et al., 2001; Rolls, 1999; Schultz et al., 2000; Thut et al.,
1997) and ventral striatum (Berns et al., 2001; Breiter et creases spike production), an observation that leads to
the prediction that the CD outcome should be associ-al., 2001; Koepp et al., 1998; Pagnoni et. al., 2002;
Schultz, 1998) to reward processing, this suggests that ated with deactivation of the midbrain and perhaps the
striatal neurons to which it projects. Indeed, CD wasthe orbitofrontal and anteroventral striatal activations
associated with the CC outcome in our experiment may often associated with deactivation of the anteroventral
Table 4. Reaction Epoch: Location of Significant Brain Activations for the Contrast Comparing the CC Outcome with the Average of the Other
Three Outcomes when Playing with a Human Partner
Region Coordinates Peak t Statistic Number Voxels
CC versus all other choices
L paracentral lobule (BA 7) 18 39 54 6.45 * 22
R caudate 3 18 0 5.35 * 14
L postcentral gyrus (BA 1) 39 30 60 4.3 7
R medial frontal gyrus (BA 11) 3 48 12 4.03 28
rostral anterior cingulate gyrus (BA 32) 3 51 6 3.65 6
L superior temporal gyrus (BA 22/42) 51 30 12 3.57 7
R paracentral lobule (BA 5/7) 18 45 60 2.99 5
Activations are for Experiment 2 (p 0.01, n 17 subjects) after limiting the search volume to voxels that survived (p 0.01) in Experiment
1 (n 19 subjects). Voxels surviving a corrected p value 0.05 after small volume correction with the mask from Experiment 1 are marked
with an asterisk. Activations consisting of fewer than five contiguous voxels are not reported in the table. L, left hemisphere; R, right hemisphere.
Neuron
400
Figure 5. Reaction Epoch: Activation in Player A in Response to CC Outcomes
(A) Voxels activated more by mutual cooperation (CC) than the mean of the other three outcomes in Experiment 2 (p 0.01) after masking
the results with voxels activated for the same contrast in Experiment 1 (p 0.01).
(B) Plot of contrast value for CC versus others in the peak voxel of the anteroventral striatal ROI against the probability of CC repeating in
consecutive rounds, for the 17 subjects in Experiment 2.
(C and D) Statistical parametric map for the contrast in (A) displayed on a normalized EPI image for (C) Experiment 1 and (D) Experiment 2.
Statistical t images are thresholded at p 0.01 (uncorrected).
(E) Voxels showing a significant condition (human versus control) by monetary outcome ($2 versus others) interaction (p 0.01) in which the
response to $2 is greater for the social than the control condition. Data are for Experiment 1 only because the control condition was not
included in Experiment 2. OFC ventromedial frontal/orbitofrontal cortex.
striatum in our experiment (Figure 4C). DC was also often evolved in the service of preserving social relationships
based on reciprocity (Trivers, 1971; Frank, 1988). Thisassociated with striatal deactivation, an observation that
could be reconciled with predictions if subjects find DC agrees with the everyday observation that we often behave
altruistically toward others simply because we likemore aversive than DD (they defect to protect themselves
from potential exploitation by a defecting partner them, not because we consciously calculate that they
are likely to reciprocate in the future.but experience guilt upon realizing a DC outcome).
Cooperating is always risky given the unpredictability Subjects who find the CC outcome rewarding would
be expected to persist with CC outcomes more thanof the intentions of another person in a social dyad. So,
it is possible that the observed pattern of activation other subjects. We were therefore interested in whether
the magnitude of the activation in the anteroventral stria-relates more generally to a realization of success following
a risky decision and not specifically to a reciprocated tum and OFC was related to subjects' tendencies to
persist with CC outcomes. Indeed, subjecťs who wereact of altruism. Alternatively, it may be the case that the
observed activation is associated with positive feelings more likely to experience consecutive CC outcomes had
greater activation in the peak voxel of the anteroventraltoward one's partner; that activation of anteroventral
striatum and OFC can result in feelings of trust and striatum ROI (r 0.70; p 0.002, Figure 5B). There was
no such behavioral correlation for the peak voxel of thecomradery that reinforce the cooperative act, superseding
any conscious recognition that material gains will OFC ROI.
Comparisons between human and computer activa-flow from mutual cooperation. Indeed, some theorists
have proposed that many of the social emotions have tion patterns show that the orbitofrontal activation asso-
A Neural Basis for Social Cooperation
401
Figure 6. Reaction Epoch: Event-Related Plot for the Peak Voxel in the OFC for the Contrast CC versus Other Outcomes
(A) Fitted response for all four outcomes in a single subject, CC red, CD blue, DC green, and DD cyan.
(B) Raw data for CC outcome for a single subject. The outcome is revealed at t 0 s and displayed for 9 s.
ciated with CC outcomes is not specific to rewarding tions are formed. Thus, one possible interpretation of
this activation is that it is related to a representation ofhuman social interaction but can also be elicited by
interactive computer programs, at least when the latter a somatic state of an emotional experience that follows
mutual cooperation.are programmed to be responsive to their partner's behavior.
On the other hand, cooperation with a human
partner may be a more effective stimulus for striatal Decision Making
The decision to cooperate following cooperation bymechanisms related to reward since we did not observe
striatal activation in association with CC for computer one's partner in the previous round was associated with
activation in the right post-central gyrus. The post-cen-partners.
Finally, we note that the most significant activation in tral gyrus activation is in primary somatosensory cortex
and could be a neural representation of a somatic re-association with the CC outcome was in neither the OFC
nor striatum, but in somatosensory association cortex sponse to an imagined decision to reciprocate cooperation
(Damasio, 1994; Aziz et. al., 2000).in the medial posterior parietal lobe (BA 7; see Table 4
and Figure 5A). A prominent theory of emotion processing The anteroventral striatum was also activated for this
contrast (i.e., XC,CX versus others). Our social decision-proposes that a neural representation of an organism's
somatic state is an important referent of emotional expe- making epoch (for round n 1) trails but overlaps with
the reaction epoch (to round n), raising the possibilityrience (Bechara et. al., 2000), and that somatosensory
association cortex is largely where these representa- that the anteroventral striatal activation represents proTable
5. Reaction Epoch: Location of Significant Brain Activations for Contrast Comparing the CC Outcome with the Average of the Other
Three Outcomes when Playing with a Computer Partner
Region Coordinates Peak t Statistic Number Voxels
Experiment 1 (n 19 subjects)
No activations
Experiment 2 (n 17 subjects)
L insula 39 3 18 4.7 18
L OFC (BA 11) 3 36 12 4.39 12
L anterior insula 27 9 6 4.15 6
L frontal pole (BA 10) 6 66 6 3.86 6
R OFC (BA 11) 6 48 18 3.35 18
Activations for computer partners (p 0.01) were masked with the results of the same contrast for human partners (p 0.01) to show areas
of overlap. Activations consisting of fewer than five contiguous voxels are not reported in the table. L, left hemisphere; R, right hemisphere.
Neuron
402
Table 6. Decision-Making Epoch: Location of Significant Brain Activations for Contrast Comparing Cooperation Following a Cooperative
Choice by One's Partner in the Previous Round (XC,CX) with the Average of the Other Three Outcomes, (XD,CX), (XC,DX), (XD,DX), when
Playing with a Human Partner
Region Coordinates Peak t Statistic Number Voxels
(XC,CX) versus all other conditions
L anterior caudate 12 24 12 5.23* 10
R post-central gyrus 36 27 54 4.76* 5
R anterior cingulate gyrus (BA32) 3 36 6 4.06 5
R collateral sulcus 39 45 6 3.87 5
R caudate 6 21 6 3.79 5
Activations are for Experiment 2 (p 0.01, n 17 subjects) after limiting the search volume voxels that survived in Experiment 1 (p 0.01,
n 19 subjects). Voxels surviving a corrected p value 0.05 after small volume correction with the mask from Experiment 1 are marked with
an asterisk. Activations consisting of fewer than five contiguous voxels are not reported in the table. L, left hemisphere; R, right hemisphere.
longed responses to the CC outcome that extend into to emotions (Davidson, 2000). Its involvement with emotion
is also supported by multiple neuroimaging studiesour decision-making epoch. However, it is also possible
that some of the activations in Table 6 relate specifically (Drevets and Raichle, 1998). Thus, the observed rostral
anterior cingulate activation may reflect the emotionalto social decision making. For example, the anterior
cingulate cortex is involved in the detection of cognitive tone of social decision making.
The decision to continue cooperating following a CCconflict (Cohen et. al., 2000). The decision to persist with
cooperation may involve conflict given the ever present outcome in the previous round also requires overcoming
a putative bias that humans and other animals have totemptation to defect and earn an extra dollar. However,
processing of cognitive conflict has been linked with weight the attractiveness of a reward in inverse proportion
to its delay (Chun and Herrnstein, 1967), a bias thatthe caudal anterior cingulate, known as its cognitive
division, whereas the cingulate activation we report here would encourage our subjects to value the immediate
reward of defection and its $3 payoff more than theis in rostral anterior cingulate cortex (Bush et. al., 2000).
Nevertheless, conflict based on emotional interference delayed reward from sustained mutual cooperation. In
other words, persisting with mutual cooperation re-reportedly activates the rostral ACC (Whalen et. al.,
1998), and it has been hypothesized that this region may quires restraining the impulse to defect and achieve
immediate gratification. Accumulating evidence impli-be generally involved with processing conflict related
Figure 7. Decision-Making Epoch: Activation
Related to the Decision-Making Epoch
(A) Voxels activated more when player A
chose cooperation following a cooperative
choice by her partner in the previous round
(XC,CX) than for the average of the other three
conditions: cooperation following partner defection
(XD,CX), defection following partner
cooperation (XC,DX), and defection following
partner defection (XD,DX). Results are for Experiment
2 (p 0.01) after masking with voxels
activated for the same contrast in Experiment
1 (p 0.01).
(B) Results from (A) further masked by voxels
that were activated in both experiments for
the contrast CC versus others during the reaction
epoch (i.e., Table 4; Figure 5A) to show
areas activated during both reaction and decision-making
epochs. rACC rostral anterior
cingulate cortex.
A Neural Basis for Social Cooperation
403
made to clarify the game. Only after the investigators concludedcates the ventromedial frontal/orbitofrontal cortex in this
that subjects understood the task were subjects positioned in therole (Grafman et. al., 1996). Although ventromedial/orbitscanner.
Players were instructed to adopt a strategy that would
ofrontal activation was not detected in our combined
maximize their earnings (with the exception of the constraints imanalysis
of the decision-making epoch, it was activated posed on the confederate) and were compensated in direct propor(p
0.001) in Experiment 2. Patients with damage to the tion to their accumulated total.
ventromedial frontal lobe are characterized by impaired
personal and social decision making (Damasio, 1994; Experimental Design
The game matrix was projected onto a screen that player A viewedBechara et. al., 2000) and have been described as lacking
through a mirror mounted on the head coil and player B viewed onthe ability to delay gratification. Analogously, subjects who
a computer screen in an adjacent room. Player A indicated her
defect out of mutually cooperative social interactions in
decision to cooperate or defect by pressing one of two buttons on
the Prisoner's Dilemma Game opt for immediate gratifi- a fiber optic button box. Player B chose to cooperate or defect using
cation (attaining the maximum payoff for that round) and two keys on the computer keyboard. When either player pressed a
may overlook or fail to consider the future consequences button or key, their choice was indicated by a color change of the
corresponding selection above the column (Figure 1). Their partner'sof defection (partner retaliation and lower cumulative
choice would not be revealed until 12 s after the round started,earnings). The corollary is that subjects who resist the
when the game outcome for that round was displayed. The outcome
temptation to defect for short-term gain and instead
of each round was recorded and saved to a computer file that was
persist in mutual cooperation may be better guided by used to specify the general linear model design matrices for each
the future consequences of their decisions. Thus, our subject.
findings are consistent with the notion that the ventromedial
frontal cortex is involved with increasing sensitiv- Experiment 1
For Experiment 1, subjects were informed that each game wouldity to distant rewards and punishments (Rogers et. al.,
consist of 20 rounds. In one game, the subject played 20 rounds1999).
with an unconstrained human player. In another game, the playing
partner was a provocative human confederate who was constrained
Summary in her choices by having to cooperate on round 1 and defect if both
In summary, mutually cooperative social interactions players mutually cooperated on three previous rounds. Scanned
subjects were unaware of these constraints. In a third session, sub-in the Prisoner's Dilemma Game were associated with
jects played the game with a preprogrammed computer strategy.activations in anteroventral striatum, rostral ACC, and
The computer defected on round 1 of the game and subsequently
OFC that were not observed in response to monetary
played a "tit-for-tat" strategy in which it mimicked the human subreinforcement
in a nonsocial control condition. OFC, but jecťs selection from the previous round. The remaining session was
not rostral ACC or anteroventral striatum, activation was a control task to determine brain activation related to monetary
also observed for mutual cooperation with a computer reward in a nonsocial context. For the control task, subjects pressed
one of four buttons to select one square of an empty payoff matrix,partner, suggesting that the ACC and striatal activations
during the first 12 s of each of 20 rounds. Each round, the computermay relate specifically to cooperative social interactions
randomly assigned $0, $1, $2, or $3 to each square of the matrix.
with human partners.
At 12 s, the random payoff for the selected square was revealed
Cooperative social interactions with nonkin are perva- and displayed for 9 s.
sive in all human societies and generally emerge from Prior to each run, subjects were reminded whom they would be
relationships based on reciprocal altruism. Such rela- playing with (the partner's name, a "preprogrammed computer strategy,"
or the "control task"). We hypothesized that the confederatetionships arguably lay the foundation for the interdepenrun
would be more provocative if it followed a run with a typicallydence upon which societal division of labor is based.
less provocative (i.e., more cooperative) human partner. Therefore,
We have identified a pattern of neural activation that
we used a fixed order for runs. In attempting to control for the
may be involved in sustaining cooperative social rela- potential confounds related to task novelty (e.g., anxiety associated
tionships, perhaps by labeling cooperative social inter- with the very first run of the experiment), the control scan was placed
actions as rewarding, and/or by inhibiting the selfish first rather than last for 9 of the 16 subjects.
impulse to accept but not reciprocate an act of altruism.
Experiment 2
In each of three sessions, subjects played against the same prepro-Experimental Procedures
grammed computer strategy that made cooperate or defect choices
according to probabilities derived from the behavior of the uncon-Subjects
The mean age of the 19 female participants in Experiment 1 was strained human subjects from the first experiment. That is, behavioral
data from the unconstrained human subjects who played out-28.8 years (range 20­60 years). The mean age of the 17 female
participants in Experiment 2 was 23.8 years (range 20­30). The sub- side the scanner in Experiment 1 were used to calculate the
probability that a person would cooperate, as a function of theject pool was restricted to women because of published reports
that men and women play the game differently, particularly in the outcome of the previous two rounds of the game. Thus, a different
probability was calculated for each of the 16 possible contingenciespresence of a male experimenter (Hottes and Kahn, 1974; Rapoport
and Chammah, 1965; Skotko et al., 1974). (e.g., CC,CC; CC,CD; ... DD,DD). In all three games, the computer
was programmed to defect automatically in rounds 18­20 in orderPrior to scanning, all participants completed a 10 min computer
tutorial, complete with examples, intended to familiarize them with to ensure sufficient non-CC outcomes for statistical analysis. To
protect against the possibility of subjects recognizing a predictablethe Prisoner's Dilemma game and with appropriate strategies for
maximizing earnings. Specifically, it was pointed out that two play- strategy that always defected on the last three rounds, game one
included an additional three rounds (21­23) in which the computerers would both earn $40 if they both cooperated each round, but
only $20 if they both defected each round. They were also told that always cooperated. In two of the three sessions, subjects were told
that their playing partner was one of two women whom they hadone would earn $60 and the other $0 in the unlikely event that one
player cooperated each round and the other defected each round. just previously met. In a third session, they were told the playing
partner would be a computer. The first game was open ended inSubsequently, all players completed a two question multiple-choice
quiz designed to assess their comprehension of the game. For sub- the sense that subjects were not told how many rounds the game
would consist of. We included an open-ended game to control forjects who answered one or both questions incorrectly, efforts were
Neuron
404
brain activations related to anticipating the game's end. For the two ing partners. The latter includes data for both the session with the
unconstrained and confederate human partners from Experiment 1remaining games, subjects were told in advance that each would
consist of 20 rounds, with one game played with a human playing and both the open-ended and closed sessions from Experiment 2.
partner and the other with a computer partner. The order of the two
sessions was counterbalanced. The identity of the playing partner Acknowledgments
was announced before each game.
For both experiments, subjects were introduced to two human We thank Drs. Hui Mao and Stephan Hamann for assistance with
partners prior to scanning in order to reinforce the belief that they various aspects of this study. This research was supported by a
would be playing the game with real people. In both experiments, Markey Center for Neurological Sciences Fellowship (to J.K.R.),
for games against the "computer," subjects were told they would NIDA (DA00367 to G.S.B.), NIMH (MH61010 to G.S.B), and NARSAD
play the game with a "preprogrammed computer strategy that does (to G.S.B.).
not play a fixed sequence of choices. Instead, it responds to your
choices from earlier rounds with specified probabilities," but they Received: October 5, 2001
were not told what strategy the computer would play. Revised: May 3, 2002
References
Image Acquisition and Analysis
A 1.5 Tesla Philips NT scanner was used to acquire T1-weighted
Andreoni, J., and Miller, J.H. (1993). Rational cooperation in the
structural images and gradient echo, echoplanar T2*-weighted imfinitely
repeated prisoner's dilemma: experimental evidence. Econ.
ages with blood oxygen level-dependent (BOLD) contrast (Ogawa
J. 103, 570­585.
et al., 1992). For Experiment 1, we acquired 28 axial slices (5 mm
Ashburner, J., and Friston, K.J. (1999). Nonlinear spatial normaliza-thick) in a plane parallel to the anterior-posterior commissural line
tion using basis functions. Hum. Brain Mapp. 7, 254­266.that included the entire brain volume (TR 3000 ms, TE 40 ms,
Axelrod, R., and Hamilton, W.D. (1981). The evolution of cooperation.flip angle 90 , 64 64 matrix). For Experiment 2, we acquired
Science 211, 1390­1396.scans coronally with a reduced TE in an attempt to minimize magnetic
susceptibility artifacts in orbitofrontal and medial temporal Axelrod, R.M. (1984). The Evolution of Cooperation (New York: Basic
lobe regions (O'Doherty et al., 2001). 27 slices (6 mm thick) were Books).
collected perpendicular to the anterior-posterior commissural line
Aziz, Q., Thompson, D.G., Ng, V.W.K., Hamdy, S., Sarkar, S., Bram(TR
3000 ms, TE 28 ms, flip angle 90 degrees, 64 64
mer, M.J., Bullmore, E.T., Hobson, A., Tracey, I., Gregory, L., et al.
matrix). The most caudal aspect of the occipital lobe was excluded
(2000). Cortical processing of human somatic and visceral sensain
those cases where we could not cover the entire brain volume.
tion. J. Neurosci. 20, 2657­2663.
For Experiment 1, functional images were acquired in four runs of
Bechara, A., Damasio, H., and Damasio, A.R. (2000). Emotion, deci-145 volumes. For Experiment 2, functional images were collected
sion making and the orbitofrontal cortex. Cereb. Cortex 10, 295­307.in a single run of 480 volumes in which the three games were preBerns,
G.S., McClure, S.M., Pagnoni, G., and Montague, P.R. (2001).sented in succession, with intervening 1 min rest periods. Head
Predictability modulates human brain response to reward. J. Neu-movement was minimized by padding and restraint.
rosci. 21, 2793­2798.Data were analyzed using statistical parametric mapping SPM 99
(Wellcome Department of Cognitive Neurology, London, UK). Motion Boyd, R. (1988). Is the repeated Prisoner's Dilemma a good model
correction of images to the first functional scan was performed of reciprocal altruism? Ethol. Sociobiol. 9, 211­222.
within subject using a 6 parameter rigid-body transformation (Friston Breiter, H.C., Aharon, I., Kahneman, D., Dale, A., and Shizgal, P.
et al., 1995a). Images were then spatially normalized to the Montreal (2001). Functional imaging of neural responses to expectancy and
Neurological Institute (MNI) template by applying a 12 parameter experience of monetary gains and losses. Neuron 30, 619­639.
affine transformation followed by nonlinear warping using basis
Bush, G., Luu, P., and Posner, M.I. (2000). Cognitive and emotional
functions (Ashburner and Friston, 1999). Images were subsequently
influences in anterior cingulate cortex. Trends Cogn. Sci. 4, 215­222.
smoothed with a Gaussian kernel of 8 mm FWHM.
Chun, S.-H., and Herrnstein, R. (1967). Choice and delay of reinforce-A random-effects, event-related, statistical analysis was perment.
J. Exp. Anal. Behav. 10, 67­74.formed with SPM (Friston et al., 1995b). The experiment was analyzed
as a 2 2 factorial design. First, a separate general linear Cohen, J.D., Botvinick, M., and Carter, C.S. (2000). Anterior cingulate
model (GLM) was specified for each subject with four conditions and prefrontal cortex: who's in control? Nat. Neurosci. 3, 421­423.
representing the four possible choice pairs of a trial: CC, CD, DC, Damasio, A.R. (1994). Descartes' Error: Emotion, Reason, and the
and DD. This was done for each of the game sessions (e.g., uncon- Human Brain (New York: G.P. Putnam).
strained human, confederate, computer, and control for Experiment
Davidson, R.J. (2000). Affective style, psychopathology, and resil1),
yielding a 16 column design matrix if all four outcomes occurred
ience: brain mechanisms and plasticity. Am. Psychol. 55, 1196­1214.
at least once in all four runs. Global differences were controlled by
Drevets, W.C., and Raichle, M.E. (1998). Reciprocal suppression ofproportional scaling (Friston et al., 1995a), high-frequency noise was
regional cerebral blood flow during emotional versus higher cogni-removed by temporally filtering the data with a hrf low pass filter,
tive processes: implications for interactions between emotion andand linear trends were removed by entering scan number as a covarcognition.
Cognition Emotion 12, 353­385.iate in the design matrix. For each of the game sessions, we calculated
three two-sided contrast images that corresponded to the Francis, S., Rolls, E.T., Bowtell, R., McGlone, F., O'Doherty, J., Browmain
effects of the subjecťs decision (contrast vector [1 1 1 1]), ning, A., Clare, S., and Smith, E. (1999). The representation of pleaspartner's
decision (contrast vector [1 1 1 1]), and the interaction ant touch in the brain and its relationship with taste and olfactory
term (contrast vector [1 1 1 1]). Contrasts were also performed areas. Neuroreport 10, 453­459.
for all possible pair-wise comparisons of the four outcomes, and Frank, R.H. (1988). Passions within Reason: The Strategic Role of
for each outcome relative to the mean of the other three outcomes. the Emotions, First Edition (New York: Norton).
These individual contrast images were entered into a second-level
Friston, K., Ashburner, J., Frith, C., Poline, J.-B., Heather, J., and
analysis, using a separate one-sample t test. Each experiment was
Frakowiak, R. (1995a). Spatial registration and normalization of imanalyzed
separately. The resulting statistical parametric map from
ages. Hum. Brain Mapp. 2, 1­25.
Experiment 1 was thresholded at p 0.01 (uncorrected) to generate
Friston, K.J., Frith, C.D., Frackowiak, R.S.J., and Turner, R. (1995b).an initial brain map of activations related to social cooperation. A
Characterizing dynamic brain responses with fMRI: a multivariatemask was made of voxels surviving this threshold, and this mask
approach. Neuroimage 2, 166­172.was used to constrain the anatomical search space in Experiment
2. Voxels within the mask that also survived a threshold of p 0.01 Grafman, J., Schwab, K., Warden, D., Pridgen, A., Brown, H.R., and
Salazar, A.M. (1996). Frontal lobe injuries, violence, and aggression:(uncorrected) in Experiment 2 were reported as activations. We
present analyses for computer partners and assumed human play- a report of the Vietnam head injury study. Neurology 46, 1231­1238.
A Neural Basis for Social Cooperation
405
Hottes, J.H., and Kahn, A. (1974). Sex differences in a mixed-motive
conflict situation. J. Pers. 42, 260­275.
Koepp, M.J., Gunn, R.N., Lawrence, A.D., Cunningham, V.J., Dagher,
A., Jones, T., Brooks, D.J., Bench, C.J., and Grasby, P.M. (1998).
Evidence for striatal dopamine release during a video game. Nature
393, 266­268.
Lee, R.B., and DeVore, I. eds. (1968). Man the Hunter (Chicago:
Aldine).
Nesse, R.M. (1990). Evolutionary explanations of emotions. Hum.
Nat. 1, 261­289.
Nowak, M.A., and Sigmund, K. (1998). Evolution of indirect reciprocity
by image scoring. Nature 393, 573­577.
O'Doherty, J., Kringelbach, M.L., Rolls, E.T., Hornak, J., and Andrews,
C. (2001). Abstract reward and punishment representations
in the human orbitofrontal cortex. Nat. Neurosci. 4, 95­102.
Ogawa, S., Tank, D.W., Menon, R., Ellermann, J.M., Kim, S.G., Merkle,
H., and Ugurbil, K. (1992). Intrinsic signal changes accompanying
sensory stimulation: functional brain mapping with magnetic
resonance imaging. Proc. Natl. Acad. Sci. USA 89, 5951­5955.
Pagnoni, G., Zink, C.F., Mantague, P.R., and Berns, G.S. (2002).
Activity in human ventral striatum locked to errors of reward prediction.
Nat. Neurosci. 5, 97­98.
Rapoport, A., and Chammah, A.M. (1965). Prisoner's Dilemma; a
Study in Conflict and Cooperation (Ann Arbor, MI: University of
Michigan Press).
Ridley, M. (1996). The Origins of Virtue (London: Viking).
Rogers, R.D., Owen, A.M., Middleton, H.C., Williams, E.J., Pickard,
J.D., Sahakian, B.J., and Robbins, T.W. (1999). Choosing between
small, likely rewards and large, unlikely rewards activates inferior
and orbital prefrontal cortex. J. Neurosci. 19, 9029­9038.
Rolls, E.T. (1999). The Brain and Emotion (Oxford, UK: Oxford).
Sahlins, M.D. (1972). Stone Age Economics (Chicago: Aldine-
Atherton).
Schultz, W. (1998). Predictive reward signal of dopamine neurons.
J. Neurophysiol. 80, 1­27.
Schulz, W., Dayan, P., and Montague, P.R. (1997). A neural substrate
of prediction and reward. Science 275, 1593­1599.
Schultz, W., Tremblay, L., and Hollerman, J.R. (2000). Reward processing
in primate orbitofrontal cortex and basal ganglia. Cereb.
Cortex 10, 272­283.
Skotko, V., Langmeyer, D., and Lundgren, D. (1974). Sex differences
as artifact in the Prisoner's Dilemma game. J. Confl. Resolut. 18,
707­713.
Thut, G., Schultz, W., Roelcke, U., Nienhusmeier, M., Missimer, J.,
Maguire, R.P., and Leenders, K.L. (1997). Activation of the human
brain by monetary reward. Neuroreport 8, 1225­1228.
Trivers, R.L. (1971). The evolution of reciprocal altruism. Q. Rev.
Biol. 46, 35­57.
Trivers, R. (1985). Social Evolution (Menlo Park, CA: Cummings).
Whalen, P.J., Bush, G., McNally, R.J., Wilhelm, S., McInerney, S.C.,
Jenike, N.A., and Rauch, S.L. (1998). The emotional counting stroop
paradigm: a functional magnetic resonance imaging probe of the
anterior cingulate affective division. Biol. Psych. 44, 1219­1228.