31. R. F. Anderson, Z. Chase, M. Q. Fleisher, J. P. Sachs,
Deep-Sea Res. II 49, 1909 (2002).
32. M. A. Brzezinski et al., Geophys. Res. Lett. 29, 1564
33. D. A. Hutchins, K. W. Bruland, Nature 393, 561 (1998).
34. M. Ikehara et al., Paleoceanography 15, 170 (2000).
35. E. Calvo, C. Pelejero, G. A. Logan, P. De Deckker,
Paleoceanography 19, 10.1029/2003PA000992 (2004).
36. P. J. Mueller, M. Cepek, G. Ruhland, R. R. Schneider,
Paleogeogr. Paleoclimatol. Paleoecol. 135, 71 (1997).
37. M.-A. Sicre et al., Org. Geochem. 31, 577 (2000).
38. P. Martinez et al., Org. Geochem. 24, 411 (1996).
39. D. Budziak et al., Paleoceanography 15, 307 (2000).
40. A. Sanyal, N. G. Hemming, G. N. Hanson, W. S.
Broecker, Nature 373, 234 (1995).
41. N. R. Catubig et al., Paleoceanography 13, 298 (1998).
42. J. Bijma, B. Hoenisch, R. E. Zeebe, Geochem. Geophys.
Geosys. 3, 10.1029/2002GC00038 (2002).
43. D. M. Anderson, D. Archer, Nature 416, 70 (2002).
44. J. R. Toggweiler, Paleoceanography 14, 571 (1999).
45. B. B. Stephens, R. F. Keeling, Nature 404, 171 (2000).
46. J. R. Toggweiler, R. Murnane, S. Carson, A. Gnanadesikan,
J. L. Sarmiento, Global Biogeochem. Cycles 17, 1027
47. D. E. Archer et al., Paleoceanography 18, 10.1029/
48. W. Broecker et al., Global Biogeochem. Cycles 13,
49. J. R. Toggweiler, A. Gnanadesikan, S. Carson, R.
Murnane, J. L. Sarmiento, Global Biogeochem. Cycles
17, 1026 (2003).
50. S. Levitus et al., Science 292, 267 (2001).
51. L. Bopp, C. Le Que ´re ´, M. Heimann, A. C. Manning, P.
Monfray, Global Biogeochem. Cycles 16, 1022
52. L. Xie, W. W. Hsieh, Fish. Oceanogr. 4, 52 (1995).
53. M. E. Conkright et al., NOAA Atlas NESDIS 46 World
Ocean Database 2001, vol. 5, Temporal Distribution
of Nutrient Profiles (U.S. Government Printing Office,
Washington, DC, 2002).
54. A. H. Orsi, T. Whitworth, W. D. Nowlin, Deep-Sea
Res. I 42, 641 (1995).
55. We thank A. Ridgwell, S. Kienast, I. Tegen, and two
anonymous reviewers for helpful comments. We
gratefully acknowledge the work of all authors cited
in the Supporting Online Material for their invaluable
data contributions. The ideas presented here emerged
from stimulating discussions with I. C. Prentice and the
participants of the Green Ocean Project: www.bgc-jena.
Supporting Online Material
Materials and Methods
Figs. S1 to S5
Tables S1 to S3
17 September 2004; accepted 31 January 2005
Getting to Know You: Reputation
and Trust in a Two-Person
Brooks King-Casas,1Damon Tomlin,1Cedric Anen,3
Colin F. Camerer,3Steven R. Quartz,3P. Read Montague1,2*
Using a multiround version of an economic exchange (trust game), we report
that reciprocity expressed by one player strongly predicts future trust expressed
by their partner—a behavioral finding mirrored by neural responses in the dorsal
striatum. Here, analyses within and between brains revealed two signals—one
encoded by response magnitude, and the other by response timing. Response
magnitude correlated with the ‘‘intention to trust’’ on the next play of the game,
by 14 seconds as player reputations developed. This temporal transfer resembles
a similar shift of reward prediction errors common to reinforcement learning
models, but in the context of a social exchange. These data extend previous
model-based functional magnetic resonance imaging studies into the social
domain and broaden our view of the spectrum of functions implemented by the
The expression and repayment of trust is an
important social signaling mechanism that
influences competitive and cooperative behav-
ior (1–6). The idea of trust typically conjures
images of complex human relationships, so it
would seem to be a difficult part of social
cognition to probe rigorously in a scientific
experiment. Nevertheless, instances of trust can
be stripped of complicating contextual features
and encoded into economic exchange games
that preserve its essential features (7–9). For
example, in a game in which two players send
money back and forth with risk, trust is oper-
ationalized as the amount of money a sender
(9). Such trust games now enjoy widespread
use both in experimental economics (10) and
neuroscience experiments (11–17).
To measure neural correlates of trust using
we firstmade a simple modification to a single-
exchange trust game in order to improve the
we changed the single-round format to a mul-
tiround format in which the same two indi-
viduals (one designated the Binvestor,[ and
the other the Btrustee[) played 10 consecutive
rounds. This modification reflects the fact that
significant social exchanges are rarely single-
shot, and the assumption that algorithms in
our brains are tuned to this fact (1–6). Thus,
by adapting the multiround format, (i) trust
becomes bidirectional, in that both the inves-
tor and trustee assume the risk that money
sent might not be reciprocated by their part-
ner; and (ii) reputation building can be probed,
as players develop models of one another
through iterated exchange (10, 11). Partic-
ipants were informed that individual rounds of
the trust game would be implemented as fol-
lows: One player (investor) could invest any
portion of $20 with the other player (trustee),
the money appreciated (three times the invest-
ment), and the trustee then decided how much
of the tripled amount to repay (Fig. 1) (18).
Players maintained their roles throughout the
entire 10-round game. Responses were encoded
only in monetary units and player identities
were never revealed, thus stripping away many
of the confounding elements of context and
communication known to influence trust (10).
Volunteers were recruited from separate sub-
ject pools at Baylor College of Medicine in
Houston, TX, and California Institute of Tech-
nology in Pasadena, CA. Volunteers were in-
structed identically, but separately, at each
institution (instructors read a script describ-
ing the task).
We used event-related hyperscan-fMRI (h-
fMRI) to monitor homologous regions of two
subjects_ brains simultaneously as they played
the multiround trust game (19) (fig. S1). The
motivating idea behind this approach is simple:
we scan the brains of multiple subjects engaged
in a social interaction. Social decision-making
els of social partners. In principle, such covert
knowledge might be inferred from behavioral
observations. However, behavioral signals are
intrinsically lower dimensional than their under-
lying neural responses, and so behavior alone is
an insufficient signal source for inferring neural
representations. Put another way, an inference
partner ignores many observable neural
processes that give rise to that behavior. The
measurement of both interacting brains directly
cross correlation of internal models—replacing
inference with a measurement.
Reciprocity predicts trust. Linear re-
gression analyses of the behavior of 48 pairs of
trust (20). Reciprocity is defined as a fractional
change in money sent across rounds by one
1Human Neuroimaging Laboratory, Department of
Neuroscience,2Menninger Department of Psychiatry
and Behavioral Sciences, Baylor College of Medicine,
One Baylor Plaza, Houston, TX 77030, USA.3Social
Cognitive Neuroscience Laboratory, Division of
Humanities and Social Sciences 228-77, California
Institute of Technology, Pasadena, CA 91125, USA.
*To whom correspondence should be addressed.
R E S E A R C H A R T I C L E S
1 APRIL 2005VOL 308SCIENCEwww.sciencemag.org
player in response to a fractional change in
money sent by their partner. This definition is
simply an operationalized version of tit-for-tat,
that is, a repayment in kind. Deviations from
neutral reciprocity (perfect tit-for-tat) act as a
strong social signal in the context of this game.
In particular, strong deviation in investor
reciprocity was the best predictor of changes
our analysis (20, 21). Investor reciprocity on
round j was quantified as DIjj DRjj1, where
DIjis the fractional change in investment from
round j j 1 to j and DRjj1is the last frac-
tional change repayment (Rjj1j Rjj2).
Forty-eight subject pairs were scanned in
this study (21), and we divided the exchanges
into three approximately equal-sized groups:
(i) benevolent reciprocity, (ii) neutral reciproc-
ity, and (iii) malevolent reciprocity (22).
These behavioral exchange data are summa-
rized in Fig. 2A. For benevolent reciprocity,
investors are actually being generous (sending
more) in response to a defection by the trustee
(decrease in repayment) (left panel). Con-
versely, for malevolent reciprocity, the inves-
tor repays the trustee’s generosity with a
breach of trust (right panel).
Using a general linear model analysis, we
first sought trustee brain regions whose blood
oxygenation level–dependent (BOLD) re-
sponse was greater for benevolent or malevo-
lent investor reciprocity than for neutral
investor reciprocity (21). This analysis identi-
fied four significant regions: inferior frontal
sulcus, superior frontal sulcus, thalamus, and
inferior/superior colliculli (23). These findings
Fig. 1. Timeline for the
two-person trust game.
Trust experiments were
carried out in 48 pairs of
subjects. Each pair of sub-
utive trust exchanges. Each
exchange began with a
screen that indicated the
beginning of the round,
followed by a cue to in-
vest. The investor then
entrusted the trustee with
any amount between 0
and 20 monetary units.
During this first free re-
sponseperiod, the trustee
saw a blank screen for 8 s
after the investor’s deci-
sion was submitted. The investment was revealed to both players simul-
taneously. Amounts kept and given were represented both graphically (by
a bar graph) and numerically. After the investor’s decision was revealed, the
trustee was then prompted to split three times the invested amount in any
proportion between themselves and the investor. Eight seconds after the
trustee repayment decision was submitted, the repayment was revealed to
both players in the same graphical and numerical fashion. After another 8 s
were separated by a variable 12- to 42-s interval. Except for the periods of
free response, both players viewed the same visual stimuli simultaneously.
Fig. 2. Correlates of reciproc-
ity in a multiround economic
exchange. (A) Behavioral sum-
mary. Mean T SE of investor
(DI, red) and trustee (DR,
black) behavior of rounds con-
tributing to benevolent (n 0
125), neutral (n 0 134), and
malevolent (n 0 125) investor
reciprocity categories. In each
round j, investor reciprocity
was defined as rj0 DIjj
DRjj1; that is, the difference
between the current change in
payment DIjby the investor in
response to the previous
change in repayment DRjj1
by the trustee. In the case of
benevolent reciprocity, inves-
tors are being generous (send-
ing more) in response to a
defection by the trustee (de-
crease in repayment). Likewise,
in the case of malevolent
reciprocity, the investor repays
the trustee’s generosity (in-
crease in previous repayment)
with a breach of trust (20). (B)
Response of trustee brain to
investor reciprocity. A general linear model analysis identified four regions in
the trustee brain that showed responses that were greater for the revelation
of malevolent and benevolent investor reciprocity than for neutral reci-
procity (21). Only one region, the head of the caudate nucleus, showed a
response that was greater for benevolent relative to malevolent reciprocity
(statistical parametric map shown alongside pseudo-color legend). No re-
gion showed greater responses to malevolent relative to benevolent inves-
tor reciprocity. (C) Region-of-interest analysis of head of caudate in trustee
brain. Average activity 6 to 10 s after the investor’s decision is revealed to
trustee shows that the brain response to benevolent reciprocity was signif-
icantly greater from neutral (two-tailed t test, P G 0.05) and malevolent
reciprocity (two-tailed t test, P G 0.005) (21).
R E S E A R C H A R T I C L E S
www.sciencemag.orgSCIENCEVOL 3081 APRIL 2005
are consistent with a surprise signal—an
behavior of one’s partner. A second analysis,
comparing BOLD response for benevolent
reciprocity to BOLD response for malevolent
reciprocity, identified significant differences
only in the head of the caudate nucleus (Fig.
2, B and C): (i) BOLD response was greater
for instances of benevolent reciprocity relative
to malevolent and neutral reciprocity; and (ii)
responses to malevolent reciprocity did not
differ from those to neutral reciprocity. These
voxels were subsequently subjected to a
region-of-interest (ROI) analysis (21).
‘‘Intention to trust’’ signals. Weexpected
correlated with the trustee’s next choice to
repay, and we expected that such signals might
for this expectation derived from the fact that
reciprocity expressed by the investor (DIjj
DRjj1) strongly predicted (r 0 0.56) future
changes in trust (repayment, DRj) by the
trustee. For example, benevolent reciprocity by
theinvestorisexpected to generate theintention
to increase repayment (trust) in the brain of the
trustee. A similar intention to decrease trust
(repayment) would be expected in the trustee
brain following malevolent reciprocity by the
investor. Some part of the investor’s brain
should anticipate the neural consequences of
changes in their own reciprocity on the trustee’s
brain; therefore, we also expected that such
‘‘intention to trust’’ signals would show strong
cross-brain correlations. Indeed, they did.
Model building of partner: Cross-brain
analysis. To carry out this analysis, we
separated the hemodynamic responses in the
caudate of the trustee brain into three groups
according to whether their next repayment
was larger, smaller, or the same as their last
repayment. We were particularly interested in
the net neural response to the intention to
increase trust (repayment), because this act
embodies risk on the part of the trustee and
signals to the investor a degree of willingness
to cooperate. We computed the net intent-to-
trust signal in the ROI of the trustee caudate as
Hðincreased repayment next roundÞ j
Hðdecreased repayment next roundÞ
Using this difference signal in the trustee brain,
we computed cross-brain correlations with the
investor brain and sought regions with the
largest correlations. We were particularly
interested in how the cross-brain correlations
might change as the task developed and the
subjects built better models of one another.
Consequently, changes in this signal were
examined across early (3 and 4), middle (5
and 6), and late (7 and 8) rounds using cross-
brain and within-brain correlational analysis.
Figure 3 illustrates the cross-correlograms of
this signal with activity in two regions: the
the anterior cingulate cortex (ACC) of trustees
in the investor brain and ACC activity in the
trustee brain were most strongly correlated (r 9
in time by 14 s. The important point here is that
the strongest cross-brain correlation did not
shift significantly in time from early to late
rounds; that is, neural responses in both brains
to fiducial markers of the task did not change
relative to each other. However, the peak of the
cross-correlogram between investor MCC ac-
tivity and the trustee ‘‘intention to trust’’ signal
in the caudate showed a pronounced 14-s shift
from early to late rounds (green traces). A
similar finding resulted for the within-brain
analysis of the trustee, using ACC activity and
the same ‘‘intention to trust’’ signal in the
caudate (red traces).These analysesshow thata
dramatic change in the relative timing of the
in the ‘‘intention to trust’’ signal of the trustee
caudate or in both the trustee ACC and investor
MCC. As shown in Fig. 4, the source of the
shift is in the ‘‘intention to trust’’ signal of the
Figure 4 shows the time traces of the
hemodynamic responses in the head of the
trustee caudate segregated according to future
changes in trust (increases are shown in black,
Fig. 3. Correlograms of
the ‘‘intention to trust’’
with activity in investor
MCC and trustee ACC.
(A) Regions of correlation.
The ‘‘intention to trust’’ sig-
nal in the trustee caudate
was correlated within-
and between-brains with
regions that responded
strongly to basic behav-
ioral events within each
round: The middle cingu-
late cortex (MCC) of the
investor was strongly ac-
tive when the investor
lodged a decision, and the
anterior cingulate cortex
(ACC) of the trustee was
strongly activated when
an investor’s decision was
revealed (21). (B) Correlo-
grams of caudate, ACC,
and MCC. The caudate
signal between rounds of
increased and decreased
repayment isolated an
‘‘intention to trust’’ sig-
nal in trustees. Average
‘‘intention to trust’’ signal
was correlated with aver-
age ACC signal of trustee and average MCC signal of investors during the
investment phase of each round (21) and is plotted with different time shifts.
Correlograms are shown for early (rounds 3 and 4) and late (rounds 7 and 8)
periods of the game. Blue traces indicate that the strongest cross-brain cor-
relation for responses to basic behavioral events of the game did not shift
significantly in time from early rounds to late rounds. The peak of the cross-
correlogram between investor MCC activity and the trustee ‘‘intention to
trust’’ signal in the caudate shows a pronounced 14-s shift from early to late
rounds (green traces). A similar result is evident in the within-brain analysis of
the trustee, using ACC activity and the same signal in the caudate (red traces).
R E S E A R C H A R T I C L E S
1 APRIL 2005 VOL 308 SCIENCEwww.sciencemag.org
decreases in red) (21). The amplitude and
time effects associated with the 14-s time shift
are shown in Fig. 4A and summarized in the
bar graphs in Fig. 4B. In early rounds of the
task (rounds 3 and 4), the peak of the response
for intended increases in trust (i.e., an increase
in next repayment) occurs after the investor’s
decision is revealed. In middle rounds (rounds
5 and 6), this response begins to drop back
toward baseline and begins to grow at a time
just before the revelation of the investor’s
decision. By late rounds (rounds 7 and 8), this
peak is anticipatory and occurs before the
revelation of the investor’s decision. These
data are consistent with a signal for intended
increases in trust changing from being reactive
to anticipatory and suggest that the trustee is
building a model of the investor’s likely next
move. To test this model-building idea direct-
ly, we performed a separate version of the trust
game and queried the trustees on each round
about their expectation of the next investment.
Figure 5 illustrates the results of this
additional experiment (n 0 21 pairs, behavior
only). On each round, both the investor and
trustee were simultaneously prompted. The
investor was cued to make their investment
and the trustee was cued to guess the inves-
tor’s decision (Fig. 5A). Timings were other-
wise kept the same. The results of these
experiments are summarized as the fraction
of highly accurate guesses (to within T$1) by
the trustee as a function of round. Notice that
the increase in the trustee’s accuracy across
rounds parallels the time during which the
temporal transfer of the neural signal cor-
related with future increases in trust.
Discussion. We used an anonymous trust
game in conjunction with event-related fMRI
to probe neural correlates of the expression
and repayment of trust between interacting
human subjects. Important social relationships
are rarely a single expression of trust between
two strangers; thus, we made the game
multiround instead of one-shot. Specifically,
we sought to examine trust in a context in
which (i) trust was expressed by both partners
in the relationship, and (ii) trust could change
over time and with experience (25).
Using a multiround trust game and a large
sampleofsubjects(n 0 48pairs),weidentifieda
social signal (reciprocity) expressed by the
investor that strongly predicted changes in trust
by the trustee. This social signal elicited two
notable effects in the trustee brain: (i) brain
regions whose activity correlated with large
changes in reciprocity in a manner consistent
with a surprise response; and (ii) a specific brain
region, the head of the caudate nucleus, where
the BOLD response was greater for benevolent
reciprocity than for malevolent reciprocity. The
strong relation between investor reciprocity and
subsequent changes in trustee repayment led us
to probe the ‘‘intention to trust’’ in the caudate
nucleus. Rounds were segregated on the basis of
whether trustees subsequently increased or de-
creased their repayment, representing a signal of
the ‘‘intention to trust.’’ Cross- and within-brain
correlations of this intended-trust signal with
neural responses to fiducial markers of the task
(investment submitted and investment revealed)
identified a remarkable temporal transfer of the
‘‘intention to trust’’ signal from a time just after
the revelation of the investor’s decision (a re-
active signal) to a time just before this same
revelation (an anticipatory signal). This shift
development of a model of the investor in the
trustee’s brain. To examine this latter possibility,
we ran a separate behavioral experiment (n 0 21
pairs) to test the trustee’s ability to accurately
guess (to within T$1) the decision by the
investor. The error rate of these accurate guesses
the temporal transfer of the future trust signal
shifted from reactive to anticipatory. This ob-
thedevelopmentofa reputation fortheirpartner.
Lastly, we address an important detail about
the amplitude differences between the caudate
response to impending increases (black traces,
Fig. 4) and impending decreases in trust (red
Fig. 4. Neural correlates
of reputation building in
trustee brain. (A) ROI
time series. An ROI anal-
ysis was performed on
voxels identified by the
contrast illustrated in Fig.
2B (21). We segregated
in response to the reve-
lation of the investment
(time 0 0 s) according to
the next decision made
by the trustee (trustee’s
decision period begins at
t 0 22 s). Hemodynamic
amplitudes for future
increases in trust (DR 9
5%; black trace) were
greater (P G 0.05) than
future decreases in trust
(DR G j5%; red trace) in
early rounds (top). As the
game progressed (middle
and bottom), the peak of
this differentiated re-
sponse underwent a tem-
poral transfer from a time
after the revelation of the
investor’s decision (t 0
10 s; a reactive signal)
to a time before this
same revelation (t 0
j4 s; an anticipatory
signal). Traces represent
subsamples of 144 rounds in which repayment increased or decreased
Q5% (mean 0 20; SD 0 4.4). (B) ROI bar plot. The difference between the
intention to increase trust [black trace of (A)] and the intention to de-
crease trust [red trace of (A)] is plotted for t 0 j4 s and t 0 10 s. The 14-s
temporal transfer from reactive to anticipatory is consistent with the de-
velopment of a reputation for the investor within the trustee brain.
R E S E A R C H A R T I C L E S
www.sciencemag.orgSCIENCEVOL 308 1 APRIL 2005
traces, Fig. 4). One explanation, supported by
the behavioral data, is that increases in trust
(DR) may have a greater effect on their
partner’s subsequent behavior (DI) than
decreases in trust. If this were the case, an
efficient computational system would devote
more computational steps, and hence more
energy, to deciding the magnitude of an increase
in trust relative to a decrease. In this particular
trustee were correlated positively with changes
in investment on the subsequent round by the
investor(r 0 0.27) (fig. S6A). This was not true
for decreases in trust, where there was no such
correlation (r 0 0.00) (fig. S6B). The absence of
predictive information associated with a de-
crease in trust suggests that no analogous
energetic investment should be made.
Taken together, these results suggest that
the head of the caudate nucleus receives or
the game, the ‘‘intention to trust’’ is evident
only after an investment is revealed. With ex-
perience, this signal shifts to a time preceding
the revelation of the investment.This finding is
reminiscent of analogous shifts of reward pre-
(25–27) that have recently been identified by
fMRI in human caudate and putamen (28–32)
and are thought to involve outputs of midbrain
dopaminergic systems. These prediction error
signals were identified using simple condi-
tioning experiments in which lights predict
the future delivery of rewards (e.g., squirt of
juice or delivery of monetary return) (33, 34).
The scheme is simple: An initially neutral light
is flashed; it causes no change in dopaminergic
activity, but the later (surprising) arrival of
juice causes a burst of activity in the dopamine
neurons. Repeated pairing of light followed at
a consistent time later by juice causes two
dramatic changes: (i) The response to juice
delivery drops back to baseline and (ii) a burst
response occurs just after the light is flashed.
This temporal transfer of the burst response to
the light is thought to represent the future
value predicted by the light. The simplicity
of these experiments is somewhat beguiling.
The temporal transfer in the conditioning
experiments is directly analogous to the tempo-
ral shift that we observe in the trustee brain as
they build a model of the investor’s response,
but framed in the context of a social exchange.
In the trustee brain, the analog to the light is the
cue for the social partner to invest, and the
‘‘social juice’’ is change in investment. We
know that positive changes in investment
correlate with subsequent positive changes in
repayment; a correlation that grows over the
rounds of the task (fig. S5). Early in the
occurs after revelation of the investor’s decision
to increase investment (Fig. 4A and fig. S5);
that is, the increased investment is surprising.
The intention to increase repayment therefore
follows this revelation. As the game proceeds,
this ‘‘intention to trust’’ response transfers to a
time before the revelation of the investor
decision to increase investment. The only open
issue for this speculation is why the signal
transferred to this particular time. There are
in time to occur just before this. This social
prediction error interpretation is provocative
and consistent but leaves this important ques-
tion unanswered. The more general hypothesis
is that the dopaminergic system can be used to
establish more complex goal states (‘‘rewards’’)
and make more complex predictions through
connections from prefrontal cortex onto
midbrain and other subcortical structures (35).
It is possible that similar economic ex-
change tasks could be used to explore social
processing deficits in a variety of neuro-
psychiatric disorders. These include popula-
tions that have faulty or missing capacities
for building correct models of others (e.g.,
schizophrenia or autism spectrum disorders)
(36, 37), as well as individuals who mis-
attribute motivations and intentions to others
(e.g., borderline personality disorder) (38).
References and Notes
1. R. L. Trivers, Q. Rev. Biol. 46, 35 (1971).
2. R. Axelrod, W. Hamilton, Science 211, 1390 (1981).
3. J. S. Coleman, in Foundations of Social Theory (Harvard
Univ. Press, Cambridge, MA, 1990), pp. 177–179.
4. H. Rachlin, Behav. Brain Sci. 25, 239 (2002).
5. R. Adolphs, Nat. Rev. Neurosci. 4, 165 (2003).
6. E. Fehr, U. Fischbacher, Nature 425, 785 (2003).
7. C. Camerer, K. Weigelt, Econometrica 56, 1 (1988).
8. E. Fehr, G. Kirchsteiger, A. Riedl, Q. J. Econ. 108, 437
9. J. Berg, J. Dickhaut, K. McCabe, Games Econ. Behav. 10,
10. C. Camerer, Behavioral Game Theory: Experiments in
Strategic Interaction (Princeton Univ. Press, Princeton,
NJ, 2003), pp. 60–62.
11. K. McCabe, D. Houser, L. Ryan, V. Smith, T. Trouard,
Proc. Natl. Acad. Sci. U.S.A. 98, 11832 (2001).
12. D. J. de Quervain et al., Science 305, 1254 (2004).
13. J. Decetey, P. L. Jackson, J. A. Sommerville, T. Chaminade,
A. N. Meltzoff, Neuroimage 23, 744 (2004).
14. N. I. Eisenberger, M. D. Lieberman, K. D. Williams,
Science 302, 290 (2003).
15. P. Glimcher, A. Rustichini, Science 306, 447 (2004).
16. J. Rilling et al., Neuron 35, 395 (2002).
17. A. G. Sanfey, J. K. Rilling, J. A. Aronson, L. E. Nystrom,
J. D. Cohen, Science 300, 1755 (2003).
18. ‘‘$20’’ refers to 20 monetary units (MU). Subject pay-
ment at the end of the experiment varied between
$20 and $40, depending on the number of MU the
subject accumulated over 10 rounds. Payment sched-
ule was as follows: G68 MU 0 $20, 68 to 133 MU 0
$25, 134 to 200 MU 0 $30, 201 to 300 MU 0 $35, and
9300 MU 0 $40. Before the game, participants were
informed that they would receive between $20 and
$40, scaled by their performance. However, they had
no knowledge of the step payoff function until the
game was completed. Notice that the perfectly selfish
Nash equilibrium strategy (in which the investor
keeps all $20 each round) results in 200 MU; no
subject adopted this strategy.
19. P. R. Montague et al., Neuroimage 16, 1159 (2002).
20. Investments (I) and repayments (R) were scaled by the
amount available to be sent ($20 for I; three times the
amount invested for R). See fig. S2 for a description of
investments and repayments over the course of the
game. Linear regressions identified significant predic-
tors of change in trust for investors (DIj) and trustees
(DRj). Three predictors of DIjwere examined: (i) pre-
vious repayment (Rjj1: r 0 0.02), (ii) change in re-
payment (DRjj1: r 0 0.10), and (iii) previous trustee
reciprocity (DRjj1j DIjj1: r 0 0.31). Three predictors
of DRjwere examined: (i) previous investment (Ij: r 0
0.10), (ii) change in investment (DIj: r 0 0.26), and (iii)
previous investor reciprocity (DIjj DRjj1: r 0 0.56).
Thus, reciprocity was a stronger predictor than either
amount previously sent (Ijor Rjj1) or change in
amount previously sent (DIjor DRjj1). However, it is
noteworthy that reciprocity expressed by the investor
(r 0 0.56) was more strongly related to change in trust
than reciprocity expressed by the trustee (r 0 0.26).
This difference is likely accounted for by an
asymmetry in the structure of the exchange: In each
round, the investor can accumulate money ($20
endowment) without the cooperation of the trustee,
Fig. 5. Model building by
trustee brain In a separate
anonymous trust game (n 0
21 pairs), trustees were
queried to ‘‘guess the amount
invested’’ just before the rev-
elation of the investor’s pay-
ment decision to both brains;
otherwise, the task was iden-
tical to that of the original
game (n 0 48 pairs) from
which scanning data were
derived. (A) Timeline for
queries to each player (inves-
tor and trustee). During the
investment phase of the ex-
change, the trustees were
prompted to guess the
investor’s decision. The trust-
ee response to this query was
not revealed to the investor.
(B) Model building—highly
accurate guesses by trustee
of investor’s next payment. A
highly accurate guess was defined as T1 monetary unit from the actual investment (T5%). These data
show that a model of the investor’s next move is available to the trustee by the middle to late rounds
of the exchange and is not available in the early rounds.
R E S E A R C H A R T I C L E S
1 APRIL 2005 VOL 308SCIENCE www.sciencemag.org
whereas the trustee is wholly dependent on the
investor’s cooperation. This dependency of the trustee
on the investor likely results in greater responsivity by
the trustee to changes in investor reciprocity.
21. A description of methods is available as supporting
material in Science Online.
22. Each dyad contributed eight behavioral events to this
analysis (48 pairs ? 8 rounds 0 384 rounds). Investor
reciprocity cannot be calculated for the initial two
SD of j0.01T 0.35,skewness of j0.19(SE 0 0.12),and
kurtosis of 2.55 (SE 0 0.25). Rounds were divided into
approximately equal-sized categories: 125 malevolent
reciprocity rounds (x G j0.025), 134 neutral reciproc-
ity rounds (j0.025 e x e þ0.05), and 125 benevolent
reciprocity rounds (x 9 þ0.05). For additional descrip-
tion of reciprocity categories, see figs. S3 and S4.
23. Regions with Q10 significant voxels were identified
using t tests. Z values and statistical parametric
mapping (SPM) coordinates for each region are
available in table S1.
24. The correlation of change in investment (DIj) and
subsequent change in repayment (DRj) grew as
experience between players accrued (fig. S5).
25. P. Dayan, L. F. Abbott, Theoretical Neuroscience (MIT
Press, Cambridge, MA, 2001).
26. K. C. Berridge, in The Psychology of Learning and
Motivation, D. L. Medin, Ed. (Academic Press, New
York, 2000), pp. 223–278.
27. A. Dickinson, B. W. Balleine, in Steven’s Handbook of
Experimental Psychology, C. R. Gallistel, Ed. (Wiley,
New York, 2002), vol. 3, pp. 26–72.
28. G. Pagnoni, C. F. Zink, P. R. Montague, G. S. Berns,
Nat. Neurosci. 5, 97 (2002).
29. S. M. McClure, G. S. Berns, P. R. Montague, Neuron
38, 339 (2003).
30. J. P. O’Doherty, P. Dayan, K. Friston, H. Critchley, R. J.
Dolan, Neuron 38, 329 (2003).
31. J. O’Doherty et al., Science 304, 452 (2004).
32. B. Seymour et al., Nature 429, 664 (2004).
33. W. Schultz, P. Dayan, P. R. Montague, Science 275,
34. P. R. Montague, S. E. Hyman, J. D. Cohen, Nature 431,
35. R. C. O’Reilly, T. S. Braver, J. D. Cohen, in Models of
Working Memory: Mechanisms of Active Maintenance and
Executive Control, A. Miyake, P. Shah, Eds. (Cambridge
Univ. Press, New York, 1999), chap. 11, pp. 375–411.
36. K.-H. Lee, T. F. D. Farrow, S. A. Spence, P. W. R.
Woodruff, Psychol. Med. 34, 391 (2004).
37. E. L. Hill, U. Frith, Philos. Trans. R. Soc. London Ser. B
358, 281 (2003).
38. P. A. Johnson, R. A. Hurley, C. Benkelfat, S. C. Herpertz,
K. H. Taber, J. Neuropsychiatry Clin. Neurosci. 15, 397
39. This work was supported by the Center for Theoretical
Neuroscience at Baylor College of Medicine (P.R.M.),
National Institute on Drug Abuse (NIDA) grant
DA11723 (P.R.M.), National Institute of Neurological
Disorders and Stroke grant NS045790 (P.R.M.), Na-
tional Institute of Mental Health grant MH52797
(P.R.M.), NIDA grant DA14883 (G. Berns), The Kane
Family Foundation (P.R.M.), The David and Lucile
Packard Foundation (S.R.Q.), and The Gordon and
Betty Moore Foundation (S.R.Q.). We thank P. Dayan,
J. Li, T. Lohrenz, C. Stetson, and two anonymous
referees for comments on this manuscript. We thank
the Hyperscan Development Team at Baylor College of
Medicine for Network Experiment Management Object
(NEMO) software implementation (www.hnl.bcm.tmc.
edu/nemo) and G. Berns for early discussions and
efforts leading to the development of hyperscanning.
We also thank A. Harvey, S. Flaherty, K. Pfeiffer, R.
Pruitt, and S. Gleason for technical assistance.
Supporting Online Material
Materials and Methods
Figs. S1 to S6
30 November 2004; accepted 7 February 2005
Trafficking Underlying a Form of
Simon Rumpel,1Joseph LeDoux,2Anthony Zador,1
To elucidate molecular, cellular, and circuit changes that occur in the brain
during learning, we investigated the role of a glutamate receptor subtype in
fear conditioning. In this form of learning, animals associate two stimuli, such
as a tone and a shock. Here we report that fear conditioning drives AMPA-
type glutamate receptors into the synapse of a large fraction of postsynaptic
neurons in the lateral amygdala, a brain structure essential for this learning
process. Furthermore, memory was reduced if AMPA receptor synaptic incor-
poration was blocked in as few as 10 to 20% of lateral amygdala neurons.
Thus, the encoding of memories in the lateral amygdala is mediated by AMPA
receptor trafficking, is widely distributed, and displays little redundancy.
Animals continually adapt their behavior in
response to changes in the environment. It
has long been held that selective modifica-
tions in synaptic efficacy represent the phys-
ical substrate for this behavioral plasticity
(1, 2). Long-term potentiation (LTP), a cel-
lular model of synaptic plasticity, has emerged
as a leading candidate mechanism underlying
associative forms of learning in the central
nervous system (3–12). Much is now known
about the molecular mechanisms during LTP
that translate a brief change in electrical ac-
tivity patterns to a modification in synaptic
efficacy (13–23). Recent studies indicate that
synaptic addition of GluR1 subunit–containing
AMPA-type glutamate receptors (GluR1-
receptors) mediates the synaptic strengthening
observed during LTP (24, 25). An attractive
1Cold Spring Harbor Laboratory, Cold Spring Harbor,
NY 11724, USA.2New York University, New York, NY
*To whom correspondence should be addressed.
Fig. 1. Viral infection with
amplicon vectors does not
alter basic electrophysiolog-
ical properties. (A) Schemat-
ic of recombinant proteins
used in this study: GluR1-
GFP, a fusion protein of
GFP and the GluR1 sub-
unit; GluR1-C-tail–GFP, a
fusion protein of GFP and
the last C-terminal 81
amino acids of the GluR1
subunit; and GFP alone. (B
and C) Low magnification
transmitted light (B) and
epifluorescence (C) images
of a coronal section of the
right hemisphere including
the amygdala. Note the
area of GFP-expressing cells
within the lateral amygdala
(dotted line) 1 day after in-
jection. d, dorsal; m, medial.
(D and E) Highly magnified
image of the lateral amyg-
dala by infrared-differential
interference contrast mi-
croscopy (D) and epifluo-
rescence (E), which contains
a neuron expressing (upper
arrow) or not expressing
(lower arrow) GFP. (F) Su-
perimposed current-clamp recordings of an infected (green traces) and noninfected (black traces)
neuron during 300-ms current injections of –100, 0, þ100, þ200, and þ550pA. Rp, resting po-
tential of neurons indicated next to traces.
R E S E A R C H A R T I C L E S
www.sciencemag.orgSCIENCEVOL 3081 APRIL 2005