SEPARATING TRUST FROM COOPERATION

IN A DYNAMIC RELATIONSHIP

PRISONER’S DILEMMA WITH VARIABLE

DEPENDENCE

Toshio Yamagishi, Satoshi Kanazawa, Rie Mashima and

Shigeru Terai

ABSTRACT

In this article we introduce a new experimental game called

Prisoner’s Dilemma with Variable Dependence (PD/D), which
allows players to separate their trust in their exchange partners
from their cooperation with them in an ongoing relationship. The
game allows researchers to observe the emergence of trust and co-
operation separately, and ascertain the causal relationship between
them. In six studies that use the PD/D design, we find that the
players of PD/D consistently achieve very high cooperation rates,
sometimes mean cooperation rates of about 95%, which are
higher than in standard PD games sharing similar design features.
These findings demonstrate that separating trust from cooperation
is critical for building trust relations. They also show that the GRIT
(Graduated Reciprocation In Tension reduction) strategy helps
build such relations in the absence of mutual trust. Our results
suggest that it is cooperation which leads to trust, not the other
way around.

KEY WORDS . trust . cooperation . prisoner’s dilemma . risk

taking . trust game

Introduction

Interest in trust and cooperation has dramatically increased in psy-

chology, sociology, economics, and related fields in the last few
years. Just since 2001, PsycInfo culls 3461 references with the key-
word ‘trust’, and 2169 with ‘cooperation’ (as of October 2004).
A comparable search on EconLit for the same time period produces

Rationality and Society Copyright & 2005 Sage Publications. Vol. 17(3): 275–308.
[Link] DOI: 10.1177/1043463105055463

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

276 RATIONALITY AND SOCIETY 17(3)

616 references with ‘trust’ and 1409 with ‘cooperation’. Various

writers have implicated trust as a necessary ingredient for economic
prosperity (Fukuyama 1995; Knack and Keefer 1997), democracy
(Putnam 1993; Braithwaite and Levi 1998), and, of course, coopera-
tion itself (Gambetta 1988; Cook 2001). Yet the precise relationship
between trust and cooperation, whether trust leads to cooperation
or the other way around, remains elusive, and leading theorists dis-
agree on the causal direction (Hardin 2002; Macy 2002).
This disagreement is largely due to the fact that investigators of
trust and cooperation often treat the two concepts interchangeably.
For instance, the iterated Prisoner’s Dilemma (PD) game, the most
popular experimental paradigm for studying trust and cooperation,
entirely conflates the two. A player in a repeated PD, faced with the
binary choice of cooperation or defection, cannot show trust for the
other player without cooperating and cannot show distrust without
defecting; one cannot cooperate but distrust, or trust but defect.
In this article we introduce a new experimental game called
‘Prisoner’s Dilemma with Variable Dependence’, or PD/D, which
allows the players to make separate decisions to trust and to co-
operate, and which thus allows us to study the two processes
separately. We will present the initial results from a series of experi-
ments with PD/Ds, which demonstrate that subjects can sustain
much higher levels of cooperation (often above 90%) than in ordin-
ary iterated PDs, when they can separate their trust from their
behavioral choices. In PD/Ds players can initially cooperate without
trust, thereby minimizing the negative externalities of potential
betrayal by the other player; when cooperation is reciprocated,
mutual trust eventually emerges and leads to sustained trustful
cooperation. These results therefore allow us to specify the precise
mechanism by which initial cooperation without trust leads to trustful
cooperation. The findings reported later also demonstrate that
cautious unconditional cooperators (those who cooperate uncon-
ditionally but adjust their levels of trust for the other player) out-
perform those who employ contingent strategies like TFT. We will
discuss the social welfare implications of our findings.

Trust and Cooperation in Experimental Research

There is no consensus on the definitions of trust, and mutually

inconsistent technical definitions, for instance trust as a psycho-

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 277

logical state or as choice behavior, abound in behavioral sciences

(Kramer 1999: 571–4). Worse yet, many scientists use the term for
its everyday meaning without clearly defining it. We define trust,
or to be precise, an act of trust, as an act that voluntarily exposes
oneself to greater positive and negative externalities by the actions
of the other(s). This is in fact the definition of trust adopted in the
trust game literature (Dasgupta 1988; Kreps 1990). Confiding in a
friend about a personal problem, loaning money to a colleague
(without a legally binding loan agreement and a collateral), and
asking an unknown fellow theatergoer to watch a coat on your
seat while stepping out to go to the restroom are all instances of
trustful behavior. In each case, if the trust placed in the other
actor is reciprocated by trustworthy behavior, then the actor is
better off than before as a result of positive externalities of the
other actor’s cooperation. The actor is better off having placed the
trust than not having done so. For example, the actor may receive
friendly and useful advice from the confidant (or at least have some-
one to talk to) about the personal problem, generate a reputation as
a good colleague and expect the financial favor to be returned in the
future, or have the opportunity to go to the restroom without having
to carry the coat or losing the seat.
However, trustful behavior necessarily carries the risk of negative
consequences if the trust is misplaced and met by untrustworthy
behavior. The confidant might break the confidence and share the
confidential information with others as gossip, the borrower might
default on the loan and not repay it, and the fellow theatergoer
might walk away with the coat. In each case, if the trust is misplaced,
the actor would have been better off not placing the trust at all.
Having to deal with the personal problem alone is better than
becoming a laughing stock among friends, being known as a stingy
colleague is better than losing a large sum of money, and having to
go to the restroom with the coat and losing the seat in the theater is
better than losing the coat. The possibility of betrayal (misplaced
trust met by defection) makes placing trust in another actor always
a risky or even uncertain proposition.
Cooperation is an act that increases the welfare of the other(s) at
some opportunity cost where the former is greater than the latter.
The forgone opportunity cost (potential gains from defection) is the
hallmark of cooperation; without it, an act does not represent genuine
cooperation. In the above examples, the confidant forgoes the oppor-
tunity to spread juicy gossip about the actor among mutual friends,

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

278 RATIONALITY AND SOCIETY 17(3)

the borrower forgoes the opportunity to keep the money by default-

ing on the loan, and the fellow theatergoer forgoes the opportunity to
possess a new coat. Defection results when the other actor succumbs
to the temptation not to forgo the opportunity cost.
The welfare of an actor is a multiplicative function of trust and
cooperation: welfare ¼ trust
cooperation. The more trust the
actor places in the other, the greater the positive externalities of
the other’s cooperation (and thus the actor’s welfare), but at the
same time the greater the negative externalities of the other’s defec-
tion. Further, the actor’s trust in the other is a prerequisite for the
other’s cooperation (or defection). If the actor does not trust, the
other actor cannot choose to cooperate or defect. The confidant
cannot choose to keep or break the confidence if the actor does
not confide, the borrower cannot choose to repay or default on
the loan if the actor does not loan the money, and the fellow theater-
goer cannot choose to watch the coat or walk away with it if the
actor does not ask for the favor.
All of these considerations point to the need to study trust and
cooperation separately, to examine the causal effect of one on the
other. Given our definitions of trust and cooperation above, how-
ever, there is no element of trust in a one-shot (non-iterated)
Prisoner’s Dilemma (PD) game that is simultaneously played by
the two players. This is because a player’s choice between C and
D cannot affect the size of externalities of their partner’s choice.
Trust becomes confounded with cooperation in a sequentially
played one-shot PD, in which player A makes a decision between
C and D, and then player B, after being informed of the decision
of A, makes the same decision. (Both A and B make the decision
only once.) This is because the choice for the second player, B, is
now between reciprocating A’s choice and simple-mindedly pursu-
ing her own self-interest. Which choice, C or D, produces a better
outcome for A depends on whether B is a reciprocator or an eco-
nomically rational actor. If B is a reciprocator, A should choose C
since her choice of C makes B also choose C, and A’s outcome is
then 3 (in a PD in which T ¼ 3, R ¼ 2, P ¼ 1, and S ¼ 0). If A
chooses D, B will also choose D, and A’s outcome is 2. On the
other hand, A should choose D if B is an economically rational
actor, since B always chooses D regardless of the choice of A. By
choosing D, A gets 2, whereas she gets 1 by choosing C. Without
knowing exactly which type of a person B is, A has to choose
whether or not to trust B to be a reciprocator. The difficulty here

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 279

is that A may choose C because A trusts B to be a reciprocator or

because A is an altruist who cooperates regardless of the expected
choice of B. Similarly, A may choose D because she does not trust
B to be a reciprocator or because she is a competitor who is moti-
vated to outperform B, disregarding the absolute level of outcome.
There is a large body of experimental evidence that players of one-
shot PD games do behave reciprocally (to the expected choice of
their partners) and expect their partners to behave reciprocally,
even in simultaneously played one-shot games (Hayashi et al. 1999;
Kiyonari et al. 2000). This implies that the choice of cooperation
and defection in PD games can be partly based on the players’ moti-
vation to cooperate, to reciprocate, and their trust in their partners
to be reciprocators. In short, both trust and cooperation can poten-
tially be operating behind PD players’ choices, and it is difficult to
isolate one from the other.
Defection out of fear (the prospect of being exploited by a defect-
ing partner) is a case in point. Player A who cannot trust Player B
may still want to cooperate in order to signal her own trust-
worthiness to B, yet must nonetheless defect in order to protect her-
self from negative externalities of B’s choice and not to receive the
‘sucker’s payoff ’ when B defects. Pruitt and Kimmel’s (1977)
review of over 1000 studies of PD by the mid-1970s concludes that
the lack of trust in the partner, rather than greed, constitutes the
major factor preventing PD players from taking a cooperative
choice. From B’s perspective, however, A’s behavioral choice of
defection is identical to what she would have done if she were
driven by greed and the desire to free ride. From A’s perspective,
her defection is necessitated by B’s perceived untrustworthiness,
but from B’s perspective, A’s defection is a proof of her own untrust-
worthiness. Such miscommunication easily leads to mutual recrimi-
nation, the ‘echo’ of mutual defection, in iterated PD games (Axelrod
1984). Further, the experimenter cannot tell from the players’ actions
whether the mutual defection results from the players’ own untrust-
worthiness (preference for defection) or their inability to trust each
other despite their own trustworthiness (preference for cooperation).

Prisoner’s Dilemma with Variable Dependence (PD/D)

Game theorists who realize the undesirable implications of conflat-

ing trust and cooperation prefer the trust game (Dasgupta 1988;

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

280 RATIONALITY AND SOCIETY 17(3)

Kreps 1990) to the PD in studying trust and cooperation separately.

The trust game (TG) is played by two players, A and B. Player A is
given a choice between trusting (T) and not trusting (NT). If A does
not trust B, then the game ends there and the result is a status quo
with no changes in the player’s welfare. Since the actor’s trust is a
prerequisite for the other’s cooperation, B is given a choice between
cooperation and defection only when A chooses to trust B. When A
trusts B, then B is given a choice of cooperation and defection. B’s
choice of defection (not to honor A’s trust or NH) produces greater
personal welfare for him than his choice of cooperation (to honor
A’s trust or H): NH > H for B. The confidant forgoes (H) the
opportunity to spread juicy gossip (NH), the borrower forgoes (H)
the opportunity to keep the money by defaulting on the loan
(NH), and the fellow theatergoer forgoes (H) the opportunity to
possess a new coat (NH). For A, however, the welfare provided
by B’s defection is less than that of the status quo, whereas the wel-
fare provided by B’s cooperation is greater than status quo. TG thus
nicely captures all the elements involved in the earlier examples of
trust and cooperation. According to the logic of backward induction
used by game theorists, B should not honor A’s trust since NH > H,
and A, knowing this, should not place trust in B. While successfully
capturing the critical elements involved in trust and cooperation, TG
has two limitations: it is static, and one-sided.
The first limitation is recently removed when researchers started
using repeated TG rather than one-shot TG. An early example of
repeated TG is the Centipede Game introduced by Rosenthal (1981).
In the Centipede Game (CG), two players keep choosing between
Take and Pass. Each time a player Passes, the total amount of
money to be split between the two players increases, and the other
player is given another choice between Take and Pass. The game
ends when one of the player Takes; then, that player earns the
lion’s share of the total amount of money and the other player
earns very little. Thus, each segment of the CG represents a one-
sided TG. One important aspect of CG is that trust and cooperation
are inseparably confounded. As McKelvey and Palfrey (1992) point
out, Take in CG may be a means to maximize a player’s own self-
interest or may be lack of the (psychological) trust that the other
player will Pass. Similarly, the player may Pass because she is an
altruist who wants to improve the other player’s welfare (i.e. Pass
is based on cooperation using our terminology) or based on her
trust that the other player will also Pass and thus she can earn a

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 281

better payoff later. CG is thus not appropriate for studying the

dynamic relationship between trust and cooperation.
Bolton et al.’s (2003) study, in one condition, development (or,
more precisely, maintenance) of trust and trustworthiness by letting
their subjects play TG repeatedly between the same two players.
They also lift the second limitation of TG when they let the players
alternate between the roles of truster and cooperator. Their subjects
started with and sustained very high rates of trusting behavior (83%)
and cooperation (89%).
Another variant of TG is the investment game or IG (Berg et al.
1995). IG is also played between two players, A and B. As in TG,
Player A decides to trust or not to trust B, and B decides to honor
A’s trust or not. The difference between TG and IG is in the
nature of A’s and B’s choices. In TG, both A and B make binary
choices, A between trusting and not trusting, B between honoring
and not honoring A’s trust. In IG, they make continuous choices:
Player A decides how much trust she places in B, and Player B
decides how much of the trust placed by A to reciprocate. Specifi-
cally, Berg et al. (1995) provided A and B with an endowment of
$10 each, and asked A to transfer any of the $10 to B. The money
transferred to B is tripled by the experimenter. If, for example,
A transfers $4 to B, B receives $12. Player B, who receives the trans-
ferred and tripled money in addition to the endowment of $10, then
decides whether to send some money back to A. IG thus captures the
same elements of trust and cooperation as TG with an additional
benefit of allowing the researcher to study varying levels of trust
and cooperation. Berg et al. (1995) use IG in a static and one-
sided manner, however. It has not been used to study development
of mutual trust and cooperation between the same two partners.
The PD/D we are introducing in this article may be conceived of
as a mutual and repeated version of the investment game.
At about the same time as Berg and her colleagues were develop-
ing the investment game, and long before economists started apply-
ing TG in mutual and repeated manner, we were independently
developing a new game, PD/D, that could be used to study develop-
ment of trust relationship between partners. The game allows each
player separately to choose the level of trust she wants to place in
the other player, and the behavioral choice to cooperate or defect
with the other. In PD/D, it is possible for players to choose to co-
operate with the others without trusting them. Further, and more
importantly, the game allows researchers to observe the emergence

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

282 RATIONALITY AND SOCIETY 17(3)

of, and changes in, mutual trust between players apart from their
behavioral choices to cooperate with or defect on each other.
In PD/Ds, we measure A’s trust in B as the extent to which A
voluntarily chooses her outcome (welfare) to be dependent on B’s
behavior, or the amount of ‘fate control’ (Thibaut and Kelley 1959)
that A voluntarily gives B. If A trusts B, she voluntarily chooses
her outcomes to be dependent on B’s behavior and allows him to
exercise greater fate control over her outcomes. We will describe
the two different types of PD/D with two different operationaliza-
tions of dependence, and summarize the initial results from experi-
ments. All experiments reported here are conducted in a fully
computerized laboratory. Subjects participate in the experiment in
isolation, each in a small room equipped with a computer. Instruc-
tions, information, and decision prompts are displayed on the sub-
ject’s computer monitor, and the subject’s decision is entered into
the computer, which is connected to the control computer and
other subjects’ computers through LAN.
Most of these experimental studies have been published in
Japanese and thus have not been accessible to anyone who does
not read Japanese. The main purpose of this article is to extract
major findings that have been discovered by the use of the PD/D
experimental design and discuss some of their implications. In addi-
tion, it provides non-Japanese-speaking researchers with research
findings that have not before been made available in English. We
take this opportunity to re-analyze the data when our discussion
requires analysis that is not reported in the original studies.

PD/Dm: Trust via Matrix Changes

In the first type of PD/D, subjects choose the amount of their depen-
dence on the other player by choosing a particular payoff matrix,
which determines the joint outcomes of their behavior. Subjects
can choose to gradually increase or decrease their dependence on
the other player by increasing or decreasing the absolute size of
their payoffs associated with each outcome, thereby increasing or
decreasing their ‘fear’ of exploitation and temptation to free ride.
In the first round of the iterated PD/Dm (trust via matrix
changes), the two players play the game with the initial symmetrical
payoff matrix (T ¼ 30, R ¼ 10, P ¼ 10, S ¼ 30; see Figure 1).
Then, in the second round, before they make their behavioral

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 283

choice to cooperate or defect, each player can choose to increase

or decrease their dependence on the other player (or remain at the
current level of dependence). If Player A chooses to increase her
dependence on Player B, for instance, her payoff matrix will shift
to the right and the absolute values of all of her payoffs increase
by 10% of the original values. Her payoffs in the second round
will thus be: T ¼ 33, R ¼ 11, P ¼ 11, S ¼ 33. Note that
Player A’s payoff in the case of mutual cooperation has increased,
but the potential damage by Player B’s betrayal has also increased.
By choosing to increase her dependence on B, A has increased both
the positive and negative externalities of B’s actions on her. Such a
choice requires trust in B that B will cooperate. After choosing the
payoff matrix, both players make the binary behavioral choice to
cooperate or defect.
If, on the other hand, Player A chooses to decrease her depen-
dence on B, then the payoff matrix for the second round will shift
to the left and the absolute values of all of her payoffs decrease by
10%.1 Her payoffs in the second round will be T ¼ 27, R ¼ 9,
P ¼ 9, S ¼ 27. Player A has chosen voluntarily to decrease her
payoff in the event that her partner cooperates (when her trust in
B is met with his cooperation), but at the same time decreasing the
damage that B can potentially cause her when her trust is misplaced.
By choosing to decrease her dependence on B, A has decreased both
the positive and negative externalities of B’s actions on her. The two
players’ behavioral choices are informed to them before they decide
to increase or decrease their matrix size in the next round.
After each round, Player B can similarly choose to increase or
decrease his dependence on A. However, note that his decisions
are not known to Player A. Player A only knows of B’s behavioral
choice (cooperation or defection), but not how much he trusts her.
That is why B’s payoffs in all matrices except the initial one have
a ‘?’ in them. Player A knows that B’s payoffs can be more or less
than the initial values, but not the current sizes of B’s payoffs. We
designed our first PD/Dm this way so that the two players cannot
inadvertently communicate their expectations of each other’s beha-
vior through their levels of trust; in this version of PD/Dm (trust via
matrix changes), the two players can communicate their trustworthi-
ness only by their own behavioral choices to cooperate or defect. We
change this feature of the experiment in PD/Dc discussed below,
however. The design of our PD/Dm is essentially the same as Van
Lange and Visser’s (1999) experiment on ‘locomotion’ (which is

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

284 RATIONALITY AND SOCIETY 17(3)
Figure 1. Prisoner’s Dilemma with Variable Dependence (PD/Dm)
(Changes in Player A’s payoffs)
Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015
YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 285

their term for the choice of matrix and thus of the level of depen-
dence on the other player). However, our PD/Dm was invented inde-
pendently of Van Lange and Visser (1999) and was first published in
a Japanese journal a couple of years before theirs (in Kakiuchi and
Yamagishi 1997).
In an experiment (Kakiuchi and Yamagishi 1997; first experiment
in Yamagishi and Kakiuchi 2000, in English) using PD/Dm, subjects
are first grouped into ‘high trusters’ and ‘low trusters’ depending on
their responses to two questions on the pre-experimental question-
naire (median-split), administered to potential participants at least
a few weeks before the experiment. These questions measure their
levels of general trust: ‘most people are basically honest,’ and ‘most
people are trustworthy.’ Then, two players, unbeknownst to them,
are matched on their level of general trust. In the PD/D condition
(n ¼ 40), the subjects play the PD/Dm described earlier, and in the
PD condition (n ¼ 40), they play the ordinary PD with the initial
payoff matrix on all rounds without the ability to increase or
decrease their dependence by the choice of payoff matrix. Subjects
in both conditions play the game for 48 rounds (although they do
not know the exact number of rounds ahead of time). The experiment
takes place in Japan with Japanese subjects. Subjects interact with
their partners in their own compartment, and they have no oppor-
tunity to meet other participants either before, during or after the
experiment. Their decisions are completely anonymous. They start
with the matrix located at the center of Figure 1, and play the game
for 48 trials. They are paid one yen for each point they earn in the
experiment.
In the PD/D condition, both high trusters and low trusters
increase their levels of dependence on the other player throughout
the experiment, although high trusters do so much more quickly
than do low trusters. Kakiuchi and Yamagishi (1997) break down
the 48 trials into six trial blocks of eight trials. The difference in the
matrix size between low trusters and high trusters started to emerge
in the fourth trial block in which the increase in the matrix size was
145.3% of the original matrix among high trusters and 96.8%
among low trusters. During the last trial block, the increase reached
254.5% of the original matrix among high trusters, and 133.4%
among low trusters. As Table 1 shows, the mean level of cooperation
in the PD/D condition is higher among high trusters than among
low trusters, whereas the difference between low and high trusters
is small in the PD condition. More importantly, however, regardless

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

 286
Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

Table 1. Trust and cooperation in Prisoner’s Dilemma (PD) and Prisoner’s Dilemma with Variable Dependence (PD/D) games

Experiment Condition n N of trials Trust Cooperation**

and subjects

KY97 PD high trust 20 48 .45

RATIONALITY AND SOCIETY 17(3)

Japanese PD low trust 20 48 .53
PD/Dm high trust 20 48 24.42* .82
PD/Dm low trust 20 48 18.80* .49

MY2001 Random PD 150 24 .40

Japanese PD 52 36 .66
PD/Dc 98 36 7.35 coins .76
CYCCMM2003 Random PD 106 25 .39
Americans PD 32 45 .58
PD/Dc 28 45 8.92 coins .90
Random PD/Dc 46 45 6.81 coins .53

Japanese Random PD 192 24 or 25 .42

Random PD/Dc 42 45 5.06 coins .47
 YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION
TMY2003a PD/Dcc 40 22 minutes .99 .95
Japanese
TMY2003b, nPD/Dcc 14 45 minutes .93 .83
Experiment 1 Naı̈ve subjects
Japanese nPD/Dcc, Naı̈ve and expert 11 45 minutes .99 .79
subjects
Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

TY unpublished nPD/Dcc 14 18 minutes .97 .91

Japanese Cost Condition
nPD/Dcc 14 18 minutes .98 .91
No-cost Condition
MYM2003 nPD/Dcc 38 30 minutes .97 .89
Japanese 44 30 minutes .98 .84
Americans

* *Average net increase in matrix size.

** Proportion of the times entrusted money is returned for PD/Dc and PD/Dcc.

287
288 RATIONALITY AND SOCIETY 17(3)

of their level of general trust, players of PD/D are able to sustain

much higher levels of mutual cooperation than players of PD,
Fð1; 36Þ ¼ 4:21, p < :05. The mean cooperation rate among high
trusters in PD/D is 82%, whereas that among low trusters is 49%,
about the same as the mean cooperation rates of high and low
trusters in the PD condition (45% and 53%). The average coopera-
tion rate in the last trial block reaches 88% among high trusters and
67% among low trusters in PD/D, whereas the average cooperation
rate in PD in the last trial block is 72% among high trusters and 73%
among low trusters. It appears that the ability to separate players’
behavioral choices from the amount of trust they place in the other
player increases the likelihood of mutual cooperation, especially
when high trusters play each other (the interaction effect of general
trust and game type, Fð1; 36Þ ¼ 6:33, p < :05). Van Lange and
Visser (1999), who used a similar design, also report a high level of
cooperation and reach the same conclusion.

PD/Dc: Trust via Coin Entrustment

Upon completion of the first experiment with PD/Dm, we reached
the conclusion that the experimental task for the subjects, where
they make separate decisions first to choose the payoff matrix
(only one half of which was visible to them) and then to make the
behavioral choice to cooperate or defect, might have been cogni-
tively too demanding for our subjects. We have therefore designed
an entirely different experiment, which nonetheless derives from
the same underlying logic and thus represents a conceptual replica-
tion of our first experiment.
In the PD/Dc (trust via coin entrustment), each player first
receives an endowment of 10 coins at the beginning of each round,
then each makes a decision as to how many of the 10 coins they
want to entrust to the other as in the investment game. However,
PD/Dc differs from IG in that it is symmetrical; each player
makes the trusting decision. They can choose any number of coins,
shown on their computer display, between 0 and 10 to entrust (the
minimum number of coins to entrust is one in some experiments)
and keep whatever coins they do not entrust to the other. The
number of coins entrusted by the subject appears on her partner’s
computer display. Then, in the second phase of the round, the players
make the binary decision as to whether or not to return the entrusted
coins to the other player. If they choose to return them (equivalent to

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 289

cooperation), the coins are doubled in number before being returned

to the original player. If they choose not to return them (defection),
they get to keep the entrusted coins, but they are not doubled in
number. Earlier experiments on PD (Yamagishi and Sato 1986;
Kollock 1993) have allowed subjects to choose continuous levels
of cooperation (rather than the binary cooperate–defect decision).
However, none of these experiments allow their subjects to separate
trust from cooperation.
PD/Dc is similar to IG in some respects but different in others.
The two share the same idea that the trusting choice creates an
opportunity for the trusted to cooperate or defect. The amount of
trust determines the size of the pie to be shared and cooperation
divides the pie. An important difference exists, however, between
the two games that affect interpretation of the subject’s behavior.
In IG, the money transferred from Player A triples in value before
it reaches Player B; in PD/Dc, the coins entrusted to Player B by
Player A double in value only when Player B chooses to return
them. In IG, it is the truster’s decision that increases the size of
the pie to be divided; in PD/Dc, it is the trusted player’s decision.
From the point of view of the trusted, how much to return to the
truster in IG involves a division of a fixed sum; it may thus reflect
his fairness concerns, but it is difficult to call it cooperation. In
contrast, the choice for the trusted in PD/Dc is between greater
collective welfare versus maximization of own individual welfare dis-
regarding the collective welfare. IG is thus better suited to study fair-
ness behavior on the part of the trusted and expectation of fairness
on the part of the truster. PD/Dc is better suited to study coopera-
tion on the part of the trusted and expectation of cooperation on
the part of the truster. We use PD/Dc since we are interested in the
relationship between trust and cooperation rather than between
trust and fairness.
The second important difference between the two games is
whether the choice of the trusted is binary or continuous. (The
choice of the truster is continuous in both games.) When the game
is asymmetrical, as in the original IG, the trusted’s choice must be
continuous to allow him to behave in a fair manner. For example,
let us assume that the truster sends all $30 to the trusted, and that
the trusted has a binary choice of returning all $30 or keeping all
$30. Then, the choice for the trusted is between two kinds of unfair-
ness. When he sends back the money, the truster gets $30 and the
trusted $10, whereas when he does not send back the money, the

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

290 RATIONALITY AND SOCIETY 17(3)

truster gets nothing and the trusted gets $40. This problem dis-
appears, however, when the game is symmetrical; if both transfer
all of their endowments to their partners and each chooses to send
back all, each gets $30. We use PD/Dc for its simplicity.
In PD/Dc, the choice in the entrustment stage is continuous (the
players can choose to entrust any number of coins), whereas the
choice in the behavioral stage is binary (they can either return or
keep the entrusted coins in toto but may not return some of the
coins and keep others). The game is symmetrical; each player first
makes a choice of how many coins to entrust to her partner, and,
second, whether or not to return the entrusted coins. The player’s
total earnings for each round is the sum of the number of coins
she chooses not to entrust to the other player in the entrustment
stage, plus the number of coins entrusted to her if she chooses not
to return them, and double the number of coins that she entrusts
to the other and are returned to her (if they are returned by the
other player) in the behavioral stage. The payoff structure of
PD/Dc, once the decisions of how many coins to entrust to the other
player has been made, is thus: T ¼ 10 þ m þ n, R ¼ 10 þ n,
P ¼ 10 þ m  n, S ¼ 10  n, where n is the number of coins a
player entrusts to the other player and m the number of coins the
other player entrusts to him.2

Comparison of PD with PD/Dc. The first study using the PD/Dc is

conducted by Matsuda and Yamagishi (2001), in which cooperation
rates are compared between the PD/D condition and the PD condi-
tion. Ninety-eight Japanese subjects are assigned to the PD/D con-
dition and 52 to the PD condition. In order to make the results
comparable, PD players also choose between keeping and returning
the coins entrusted by the other player. However, unlike in the
PD/Dc, players in the PD condition cannot choose how many coins
to entrust; the number of coins to be entrusted to the other player is
randomly determined by a computer program and there is nothing
the players can do to alter their ‘trust’ level. The players in the PD
condition thus only make the behavioral decision to return (co-
operate) or keep (defect) the coins. All subjects first experience the
random-partner PD condition for 24 rounds. In this initial
random-partner PD condition, pairs are randomly formed each
time among six or eight subjects, and they are not informed of the
identity of their partner. The first 24 rounds thus represent repeated
one-shot PD games. The number of coins to be entrusted to the

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 291

other player is randomly determined by the computer between 1 and

10 (uniform distribution). Then, at the beginning of the 25th round,
they are matched with one partner based on their cooperation rates
in the first 24 rounds, and they play for the remaining 36 rounds with
the same partner. The participants are not informed of the total
number of rounds in advance. The manipulation of the two games
takes place during the latter 36 rounds. Subjects in the PD/D condi-
tion receive additional instructions before the 25th round and are
told that they are now matched with one partner for the remaining
rounds, and that they can choose for themselves how many coins to
entrust to their partner. Subjects in the PD condition receive instruc-
tions only about permanent pair matching. Each coin they earned in
the experiment is converted to 2 yen at the end of the experiment and
is paid to the subject as the reward for participation.
Figure 2 presents the results of this experiment. The mean co-
operation rate during the first 24 rounds (i.e., in the random-PD

Figure 2. Trends over trial blocks of trust (number of entrusted coins) and
cooperation (proportion of coins returned by the trusted player) in the PD and the
PD/D condition.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

292 RATIONALITY AND SOCIETY 17(3)

condition) is 40%. When two subjects are permanently matched in

the PD condition (during the last 36 rounds), this cooperation rate
improves to 66%. This improvement in cooperation rate reflects a
well-established difference between one-shot games and repeated
games (Axelrod 1984); the cooperation rate in PD games is higher
when the game is repeatedly played between the same two people
than when it is played once. More importantly, the mean co-
operation rate among the PD/D players of 76% is even higher
than that observed among the PD players under similar conditions,
Fð1; 148Þ ¼ 3:16, p < :08.3 The results of this PD/Dc experiment
replicate the earlier results with PD/Dm that allowing subjects to
separate trust and cooperation decisions significantly increases the
level of cooperation among them.

Comparison of Japanese and American subjects. Since the two

studies described earlier use Japanese students as subjects, Cook et
al. (2005) run a replication experiment in the USA and Japan using
American and Japanese4 subjects to replicate the finding that
separating trust from cooperation facilitates cooperation in con-
tinuing relations. In addition, they examine the hypothesis that the
cooperation-enhancing effect of separation of trust from coopera-
tion will be more pronounced among Americans than Japanese.
This hypothesis is derived from the consistent finding that the
Japanese are more risk averse than Americans. For example, accord-
ing to Hofstede (1980), Japan is near the top of the 40 nationalities
he studied on their level of uncertainty avoidance whereas Americans
are near the bottom of the list. Yamagishi et al. (1998) report find-
ings from two cross-societal experiments in which they found that
the Japanese, compared with Americans, are less willing to engage
in cooperative behavior in the absence of a collective system that
reduces social uncertainty, and are more willing to engage in beha-
vior that reduces social uncertainty. The results of experiments pro-
vide strong evidence for the argument that makes a distinction
between commitment formation as uncertainty avoidance and trust
as risk taking. High trusters – those who were shown to have a
high level of general trust in their responses on a trust scale – are
less likely to engage in commitment formation than are low trusters
when faced with a socially uncertain situation (Yamagishi et al.
1998). In uncertain situation, those who prefer to form a commit-
ment relation with a particular partner and thus reduce the risks
within such a relation are less trustful of other people in general.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 293

The same result concerning the effect of general trust on commit-

ment formation was obtained both with American and Japanese
participants. Those low in general trust of others in both societies
are more likely to form commitment relations with those who are
trustworthy despite the opportunity cost involved. The predicted
difference between the Japanese and the Americans reflects not
only relative levels of uncertainty avoidance, but also different
levels of general trust (typically higher in the USA than in Japan).
The same computer program (with translated display materials)
that has been used by Matsuda and Yamagishi (2001) is used in this
replication study with 60 American subjects, and the results are
compared to the Japanese results. In the random-PD condition (first
24 or 25 rounds), the cooperation rate is higher among Japanese
subjects (40%) than among American subjects (30%), and the differ-
ence is statistically significant, Fð1; 206Þ ¼ 8:75, p < :01 (see Table 1).
The same difference is also observed in the PD condition (last 35 or
36 rounds), where the average cooperation rate is 66% among
Japanese and 58% among American subjects, though the nationality
difference is not statistically significant, Fð1; 82Þ ¼ 1:17, ns. This
difference reverses itself in the PD/D condition in which subjects
can choose the level of trust independent of the choice between co-
operation and defection. American subjects in the PD/D condition
entrust an average of 8.92 coins per round, which is significantly
greater than the average number of coins Japanese participants
entrust to their partners (7.35), Fð1; 124Þ ¼ 7:98, p < :01. The mean
cooperation rate among American subjects jumps from 58% in
the PD condition to 90% in the PD/D condition, while the same
increase among Japanese subjects is from 66% to 76%. Given the
separation of trust from cooperation, Americans cooperate at a
higher rate than Japanese, and the nationality difference is statisti-
cally significant, Fð1; 124Þ ¼ 5:67, p < :05. As predicted by Cook
and her colleagues, Americans are better able than Japanese to
take advantage of the opportunity to separate trust and cooperation
provided by the PD/D; the interaction effect on cooperation
between nationality and game type is significant, Fð1; 206Þ ¼ 5:67,
p < :05. The main effect of the game condition (PD versus PD/D)
in the nationality
game condition ANOVA of the cooperation
rate during the last 36 rounds is highly significant, Fð1; 206Þ ¼
19.76, p < :0001, whereas the main effect of nationality is not,
Fð1; 206Þ ¼ 0:42, ns.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

294 RATIONALITY AND SOCIETY 17(3)

Cook and her colleagues examined another condition of PD/D in

which two players were randomly matched in each trial, and found
that the positive effect of the opportunity to separate trust and co-
operation provided by the PD/D requires that the game be played
between particular partners repeatedly. While American subjects
entrusted more (6.81) coins to their partners than did their Japanese
counterparts (5.06 coins), and the difference was statistically signifi-
cant, Fð1; 86Þ ¼ 11:58, p < :01, their cooperation level was not as
high as in the PD/D with a fixed partner (see Table 1). These results
indicate that the positive effect of separating trust and cooperation
exists only when it is possible to establish a trust relationship
between particular partners in which the two players trust each
other and cooperate with each other.

PD/Dcc in a Continuous Decision Environment

In the continuous-decisions version of the PD/Dc, we eliminate dis-
crete decision rounds. Players receive an endowment (real money
rather than coins) at some unpredictable and irregular interval
(e.g. every two to three minutes) and the game continues uninter-
rupted and unpunctuated for some period of time unknown to
them (e.g. 45 minutes). Within this time period, players can freely
choose to entrust, return or keep the endowment in discrete units
at their leisure. There are no time limits; once money is entrusted
to them, for example, they can indicate their choice to keep or
return it immediately, or they can postpone the decision for as long
as they want.
There are three constraints on players’ behavior in the
continuous-decision version of PD/Dc, which we will call PD/Dcc.
First, once they entrust money to a player, they cannot further
entrust more to the same player before she decides whether to
return or keep the money entrusted by them earlier. Second, if
Player A entrusts money to Player B, Player B must first decide
whether to keep or return it before entrusting his own money to
Player A. Two players cannot simultaneously entrust money to
each other. Third, even though unentrusted money accumulates in
the players’ accounts without limit, they can entrust only up to the
size of each deposit in a single entrustment decision.
In an experiment with PD/Dcc, Terai et al. (2003a) find that both
trust (measured by the proportion of the endowed money entrusted
to the partner) and cooperation (measured by the proportion of the

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 295

time that players return the entrusted money to the entruster) are
extremely high (see Table 1). Forty Japanese subjects are paired
with each other throughout the entire experiment that lasts for
22 minutes. Subjects, however, do not know how long the experi-
ment lasts. They are given 30 yen from the experimenter every 75
to 105 seconds, and choose to entrust any of that money, in incre-
ment of 5 yen, to their partner. They may repeat the entrusting
decision as many times as the endowment money is available.
They may entrust the whole 30 yen to their partner at one time, or
entrust 5 yen, wait until it is returned, and then entrust another
5 yen, wait until it is returned, and so on. Anonymity of their choices
is tightly maintained. The subjects in this experiment entrust 99% of
the money they are given by the experimenter, and return the
entrusted money 95% of the time.

nPD/Dcc in a Continuous Decision Environment with Multiple

Partners
Terai et al. (2003b) use the PD/Dcc in an environment in which
players can simultaneously interact with multiple partners. In this
n-person PD/Dcc or nPD/Dcc, there are more than two players
simultaneously in the game, but, unlike social dilemma games, all
transactions are bilateral (only two of the players involved in the
exchange of entrustment and return of money at a time). There is
no pooling or dividing of resources by the n players. However,
players can choose to spread their endowments by entrusting them
to multiple players simultaneously (e.g. entrust 10 yen to Player B,
5 yen to Player C, and 15 yen to Player D).
A total of 14 Japanese students participated in the experiment
in two seven-person groups. Each subject receives an endowment
of 50 yen every two to four minutes, and decides how much of it
(in increments of 5 yen) to entrust and to whom. The experimental
session lasts for 45 minutes, though subjects do not know how
long it lasts. The subjects in this experiment again show extremely
high levels of trust and cooperative behavior (see Table 1). They
entrust 96% of the money they are given by the experimenter, and
return 93% of the time they are entrusted money by their partner.
Each player has six other players to whom to entrust, but they
tend to concentrate on a few. Of the entrusted amount, 32% goes
to the top trustee, and 23% goes to the second, together accounting
for more than a half of the total entrustment. And the cooperation

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

296 RATIONALITY AND SOCIETY 17(3)

rate is especially high among the top trustees. The cooperation rate
of the most preferred trustee (not necessarily the same person for all
players) is 100%, and that of the second most preferred trustee is
above 99%. In short, subjects of this experiment form trust relation-
ships in which near perfect cooperation is ensured.
These results are replicated by a follow-up experiment in which
each seven-person group consists of four naive subjects and three
experts (graduate students who are familiar with experimental
gaming). The expert subjects are paid in the same way as the naive
subjects. They are encouraged to do their best to ‘exploit’ others
to make the most money they can. While the use of ‘expert’ subjects
and encouragement for them to take advantage of their expertise
reduce both trusting and cooperative behavior, they are still at rela-
tively high levels. The 14 subjects entrust an average of 96% of their
endowment, and return 81% of the entrusted money. Interestingly,
the average entrustment level (99%) is higher among expert subjects
than among naive subjects (93%).

Mixing American and Japanese Participants

Both American (n ¼ 44) and Japanese subjects (n ¼ 38) participated

in an experiment conducted by Mashima et al. (2004), using the
same nPD/Dcc as in Terai et al. (2003b). The subjects participate
in seven- or eight-person groups, consisting of three or four Japanese
subjects and four American subjects. They participate in the experi-
ment in similar isolation rooms in each location, Hokkaido
University in Japan and Cornell University in the USA. These two
laboratories were connected via the Internet, and the program was
controlled from a server computer located in the Hokkaido labora-
tory. In one condition of the experiment, the subjects do not know
that they consist of two nationalities; they have no idea that some
of the fellow participants are in another country. In the other condi-
tion, subjects know that participants are from two nationalities, and
each of the other members displayed on the subject’s computer
screen comes with a national flag indicating whether the member
is Japanese or American. To accommodate 11 hours of time differ-
ence, the Japanese participated in the morning and the Americans in
the evening.
Mashima and her colleagues (2004) find no effect of nation-
ality information or the nationality of the subject on the trusting

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 297

behavior; both American subjects (98%) and Japanese subjects

(97%) entrust most of the endowment to the other members of
their group (see Table 1). Nor do they find any ingroup-favoring ten-
dency in the choice of trustees. The cooperation rate in this experi-
ment, while still very high, is slightly lower than in the other study
(Terai et al. 2003b) with the same design and computer program.
The cooperation rate is 89% among Japanese subjects and 84%
among American subjects. This difference between the two groups
of subjects is statistically marginal, Fð1; 78Þ ¼ 3:37, p < :08, in a
subject’s nationality
partner’s nationality
display of nationality
information ANOVA. The display of nationality information helps
improve the cooperation rate, even though the level of cooperation
with ingroup members is not higher than that with outgroup
members. The cooperation rate is 83% without nationality informa-
tion and 92% with nationality information, and the difference is
significant, Fð1; 78Þ ¼ 7:67, p < :01.

Introducing Transaction Costs

The extremely high level of cooperation in the continuously played

PD/D (PD/Dcc and nPD/Dcc), even compared to the already high
level of cooperation in PD/Dc, is very impressive. The average co-
operation rates in these games are 79–95% among Japanese subjects
and 84% among American subjects. The mean trust levels are 93–
99% among Japanese subjects, and 98% among American subjects.
These extremely high levels of trust and cooperation can be attri-
buted to two factors, in addition to the features of the PD/D
games discussed above. The first is the reduction in risk. In the dis-
crete PD/Dc game, trusting the entire endowment requires a large
risk – entrusting 50 yen involves a risk of losing 50 yen. And yet a
player has to entrust the maximum amount to maximize the poten-
tial gain. If a player entrusts only 5 yen, she can gain only 5 yen when
the entrusted money is returned. (Five yen of the returned 10 yen is
the player’s original money, so she only gains 5 yen from this trans-
action, not 10 yen). If she entrusts 50 yen, then she can potentially
gain 50 yen. In the continuous PD/Dcc, she does not need to take
a large risk for achieving the maximum gain of 50 yen. She may
simply entrust 5 yen 10 times, each following the return from the
partner.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

298 RATIONALITY AND SOCIETY 17(3)

The data from the experiment demonstrate that subjects are in

fact using this strategy of dispersing risk. Subjects in the PD/Dcc
game can minimize risk at no cost. This leads to the second factor
– reduction in temptation. Insofar as a player disperses the risk in
this way, each time entrusting the minimum possible amount, wait-
ing until the partner returns the money, and then entrusting the
minimum amount again and again, the gain for the partner from
defection is minimized as well. The most the partner earns from
defecting (i.e., not returning) is only 5 yen. The minimum amount
of 5 yen is too little for them to risk jeopardizing the relationship
and losing a partner who will otherwise keep returning the money
he entrusts in the future. The continuous version of PD/Dcc repre-
sents a social situation in which people do not need to trust their
partners. Yamagishi and Yamagishi (1994; see also Yamagishi 1998;
Yamagishi et al. 1998) call such a social situation one characterized
by assurance relations rather than trust relations.
Terai and Yamagishi conducted an unpublished experiment in
which they examined this account of extremely high levels of co-
operation in the continuous version of PD/D or PD/Dcc. In one
condition of their experiment, Terai and Yamagishi introduce a
transaction cost for each entrusting action. Specifically, the subjects
in this condition are charged 5 yen each time she entrusts money to
another player. Thus, entrusting the minimum amount of 5 yen is a
sure way of losing money. Even when the entrusted person returns
the 5 yen and the truster gains 5 yen, 5 yen ‘entry fee’ is charged
by the experimenter and her profit from this successful transaction
is zero. Thus, the best she can expect from a trustworthy trustee is
zero profit, and the worst is a loss of 10 yen (5 yen of ‘waste’ and
5 yen of the entry fee). With the entry fee, players have to entrust
a larger amount at a time to make the entrusting behavior worth-
while, compromising the risk dispersion strategy. The introduction
of the entry fee is thus expected to reduce the levels of trust and
cooperation, by making risk dispersion less profitable and thus
increasing temptation to keep the larger amount of entrusted money.
Terai and Yamagishi manipulate the cost condition (no cost
versus 5 yen for each entrustment) as a within-subjects factor.
Specifically, their subjects (n ¼ 35) play the same game in the cost
condition for the first 18 minutes, and in the no-cost condition for
the last 18 minutes. On average, they entrust 98% of the endowed
money in the no-cost condition, compared to 97% in the cost con-
dition, and the difference is not statistically significant, tð34Þ ¼

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 299

1.78, ns. However, they entrust their money less frequently in the cost
condition (21.3 times) than in the no-cost condition (28.3 times),
as expected, and the difference is highly significant, tð34Þ ¼ 3:86,
p < :001. At the same time, they entrust a larger amount per entrust-
ment decision in the cost condition (15.6 yen per entrustment
decision) than in the no-cost condition (14.0 yen). The proportion
of the times the entrusted money is returned is not different between
the two cost conditions. In either condition, the entrusted money is
returned 91% of the time. On the other hand, the proportion of the
entrusted money returned in the cost condition (87%) is significantly
lower than in the no-cost condition (93%), tð34Þ ¼ 2:16, p < :05,
suggesting that the trusted players kept the entrusted money more
frequently when the amount was larger than when the amount was
smaller. While the cooperation rate in the cost condition is still very
high, it is lower than the cooperation rate in the no-cost condition
in which players can disperse risks without cost.

Cautious Cooperation

The difference in cooperation rate between the ordinary repeated

PD games and the PD/D games is impressive. Except among the
low-trusting pairs, the subjects in the experiments presented above
cooperate at a much higher rate in PD/D games than in equivalent
PD games. In Kakiuchi and Yamagishi (1997), cooperation rate
among high-trusting Japanese subjects increases from 45% to 82%.
In Matsuda and Yamagishi (2001), the cooperation rate among
Japanese subjects increases from 66% to 76%. The positive effect
of separating trust from cooperation is even more pronounced
among American subjects. In Cook et al.’s (2005) study, which is
a replication of Matsuda and Yamagishi (2001) with American sub-
jects, cooperation rate among American subjects increases from
58% to 90%! The high rates of cooperation in the PD/D conditions
in these studies, 82%, 76%, and 90%, compared with those in the
ordinary repeated PD conditions, 45%, 66%, and 58%, are impres-
sive. And the cooperation rates are even higher in the continuous
version of the PD/D, 83–95% among Japanese subjects, and 84%
among American subjects.
Here, we address the question of why separating trust from co-
operation is so important in generating and sustaining such high
levels of cooperation. We have already provided the clue for this

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

300 RATIONALITY AND SOCIETY 17(3)

answer in the introduction. Separating trust from cooperation pro-

vides a remedy for ‘fearful defection’. In ordinary PD games in
which trust and cooperation are conflated, even those who are will-
ing to forgo the opportunity to exploit the partner unilaterally in
order to achieve and maintain mutual cooperation may still fail to
behave cooperatively to avoid the ‘sucker’s payoff’. Using Arneson’s
(1982) terminology, they are ‘nervous’ or ‘reluctant’ cooperators.
Nervous or reluctant cooperators recognize the importance of
mutual cooperation, and are willing to cooperate if they are con-
vinced that others are also willing to do so. In the absence of such
assurance, however, they hesitate to cooperate.
There are two possible reasons for this. First, they may be afraid
that their efforts for acting cooperatively may be wasted if others do
not cooperate. For example, people are reluctant to join the election
campaign of a candidate who seems unlikely to attract much
support. Arneson (1982) calls these people ‘nervous cooperators’.
Second, they may be afraid that their cooperative initiatives will
be taken advantage of by predatory non-cooperators. For example,
a price-fixing cartel (although it is illegal in many countries) will fail
if members are afraid that others would undersell. If some corpora-
tions lower their prices, those that do not will lose their share and
will eventually be driven out of the market. Arneson (1982) calls
those who are afraid that they may be exploited by non-cooperators
‘reluctant cooperators’. Nervous or reluctant cooperators will co-
operate if they are convinced that others will as well. If they cannot
trust others, they will not cooperate.
This is the story of PD, but not of PD/D. The nervous or reluc-
tant cooperators do not need to defect out of fear (of waste or of
exploitation) in the PD/D. In order to reduce their fear and feel
safe, they only need to reduce their levels of trust. In the PD/Dc,
for example, nervous or reluctant cooperators can stop sending
money to their partner while returning all the money they are
entrusted by the partner. This conveys their intentions to build a
mutually cooperative relation with the partner, and thus prevent a
vicious cycle of self-fulfilling suspicion (or using Lawler’s [1986]
terminology, a conflict spiral).
We believe that this is an important part of the story, but it is
not the whole story. The other part of the story is that a relationship
in which both parties can safely trust each other is self-sustaining
once it is achieved. It is difficult to achieve such a relation, and thus
it is a valuable resource for both parties. A friendship in which

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 301

you can safely confide in your friend is hard to develop, and thus
such a friendship is extremely valuable to you. Furthermore, such
a relationship is more valuable to you to the degree it is difficult
to achieve. A mutually trustful and trustworthy relationship is self-
sustaining since it is too valuable to destroy for a short-term gain.
Conversely, it takes a larger short-term gain to forgo such a valuable
relation. The difficulty, of course, is how to build one. Separating
trust from cooperation provides a first step toward achieving such
a difficult state, but it is just a start. Players of a PD/D game can
reduce their fear by not trusting their partners. The initial strategy
of a nervous or reluctant cooperator in a PD/D game is cooperation
without trust. The challenge is how to transform this mutually dis-
trustful relation into a mutually trustful one.
Matsuda and Yamagishi (2001) analyze the players’ strategies,
and come to the conclusion that cautious unconditional cooperation
is the best way to achieve and sustain trustful cooperation. They
analyze what kind of strategy each of their subjects uses in their
experiment, by classifying them into types based on their trusting
and cooperation behaviors in response to their partner’s behaviors
in the previous round. Then, they compare the earnings associated
with each strategy. According to their analysis, the most frequently
observed players are the ones who unconditionally cooperate (regard-
less of the partner’s behavior in the previous round) while adjusting
their level of trust according to their partner’s cooperation–
defection choice in the previous round. They start with entrusting
only a few coins, and increase the number of coins to entrust
when their partner returns them, and decrease it when the partner
does not return them. They call these type of subjects ‘cautious
unconditional cooperators’. They are the highest earning players
as well. They (n ¼ 25, 30% of all the subjects in the PD/D condition)
earn 650.8 yen on average and outperform another major type,
double TFTers (n ¼ 11, 13%), who adjust their cooperation–
defection choices as well as their level of trust to their partner’s beha-
vior on the previous round and earn an average of 477.0 yen.
Figures 2 and 3 dynamically represent the strategy of cautious
unconditional cooperation in a PD/Dc across all subjects (Matsuda
and Yamagishi 2001). They show that as soon as the subjects are
paired in repeated games (at the beginning of Trial Block 5), their
cooperation rates immediately soar to a very high level (around
80%); in other words, their cooperation immediately becomes
almost unconditional. However, their trust, measured by the number

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

302 RATIONALITY AND SOCIETY 17(3)

of entrusted coins, does not accompany the rapid increase in co-

operation. In the first trial block after pairing (Trial Block 5), the
trust level is still almost as low as in the previous trial blocks during
one-shot games; even while their cooperation is nearly unconditional,
the players are cautious. Their trust goes up only after a full trial
block of six trials, during which their cautious trust was reciprocated
by the other player’s cooperation. Only then does their trust go up as
high as their cooperation. Figure 3 shows the lagged nature in the
increase in trust during the fifth and the sixth trial blocks. Further
analysis indicates that the gradual increase in trust during the 5th
and the 6th trial blocks shown in Figure 3 occurs only among un-
conditional cooperators who never defect in these two blocks
(n ¼ 45). Their trust levels increase from 5.36, 6.76, and 7.62 in
trials 25, 26, and 27, respectively, to 9.73, 9.78, and 9.96 in trials
34, 35 and36, respectively. Among those who defect at least once
in the two trial blocks (n ¼ 53), practically no increase in trust
occurs; their trust levels in trials 25, 26 and 27 are 4.06, 4.94,
and 5.09, respectively, and in trials 34, 35, and 36, they are 5.62,

Figure 3. Trends over trials in the fourth through sixth trial blocks of trust (number of
entrusted coins) and cooperation (proportion of coins returned by the trusted player)
in the PD/D condition.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 303

5.66, and 5.89, respectively. These results clearly demonstrate that

cautious unconditional cooperation precedes (and, if reciprocated,
eventually leads to) trustful cooperation. It is cooperation that
leads to trust, not the other way around.
The success of cautious unconditional cooperators demonstrates
the importance of cooperation without trust as the first stage of
eventual success in formation of trustful cooperation, through the
process known as GRIT (Graduated Reciprocation In Tension
reduction; Osgood 1962). Lindskold’s (1978) review of empirical
studies suggests that following the prescriptions set out in the 10
principles of the GRIT strategy leads to building trust and then to
cooperative relationships between exchange partners of various
kinds. Cautious unconditional cooperation in PD/Ds can accom-
plish two goals that PDs do not allow. First, unconditional coopera-
tion allows the players to signal their trustworthiness to their
exchange partner. Second, unlike cooperation in PDs, cooperation
without trust in PD/Ds allows the players to do so without much
risk. The only thing that cautious unconditional cooperators lose
by cooperating is the opportunity cost of exploiting their exchange
partner by defecting on them; they do not risk much of their own
resources because the PD/D allows the players to cooperate without
voluntarily transferring control of their welfare to their partner.
Once their low risk trust (e.g. entrustment of one coin) is recipro-
cated, then players can gradually increase their level of trust in the
other player, by entrusting more and more coins, while at the
same time unconditionally cooperating by returning all the coins
entrusted to them. Mutual trust, and fully trustful cooperation, will
eventually develop out of initial cooperation without trust. Lawler
and Yoon’s (1993, 1996) work on relational cohesion demonstrates
that individuals often experience euphoria (‘emotional buzz’) when
they complete successful exchange via mutual cooperation. Such
emotional buzz after successful exchange might be the proximate
mechanism that leads to greater trust to the extent that individuals
are more likely to trust those with whom they are happy.

General Discussion

Individual welfare, as we noted earlier, is a multiplicative function of

trust and cooperation: welfare ¼ trust
cooperation. The level of
trust multiplies the beneficial effect of cooperation; the more one

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

304 RATIONALITY AND SOCIETY 17(3)

trusts the other, the more beneficial the other’s cooperation. The
same is true at the macrosocial level; social welfare in society is a
multiplicative function of the average level of trust among people
and the mean rate of cooperation. The society as a whole benefits
when individuals engage in trustful cooperation; trust in and of
itself does not produce social welfare if unaccompanied by coopera-
tion, and cooperation in and of itself does not produce social welfare
if unaccompanied by trust. In PD/Dc, for example, 100% trust does
not increase the number of coins in the society (i.e. the group of sub-
jects) unless they are returned, and 100% cooperation does not
increase the number of coins in the society either if nobody entrusts
any; in fact, no one even gets a chance to cooperate if nobody
entrusts.
There is a paradox here. While maximal social welfare is possible
only with mutually trustful cooperation, trust emerges from initial
cooperation without trust. Actors must first signal their trustworthi-
ness by unconditional cooperation, but must also begin slowly by
gradually and cautiously increasing their level of trust in the others.
When and only when their initial low trust is rewarded with coopera-
tion can they begin to trust more. Mutual trust emerges through a
series of risk-taking, by the act of trusting more today than yester-
day. Once mutually trustful cooperative relation emerges, however,
it is self-sustaining because it is difficult to replicate it in other rela-
tionships. Given uncertainty inherent in any new exchange relation-
ship, it would be utterly irrational for self-interested actors to defect
in a self-sustaining cooperative relationship.
The findings from the series of experiments reported here, all
using some forms of PD/D methodology, shed light on the intricate
relationship between trust and cooperation. Do we cooperate when
we are trusted more than when we are not trusted? Or, alternatively,
do we trust when our partner has behaved in a cooperative manner?
The success of the GRIT strategy adopted by the cautious uncondi-
tional cooperators – those who unconditionally cooperate, but adjust
their levels of trust to their partner’s level of cooperation – implies
the latter causality. It is important to be unconditionally cooperative
for the success of the GRIT strategy (Osgood 1962; Lindskold 1978).
Unconditional cooperation by a player invites the partner to trust
them. Once the process starts, the relationship between trust and
cooperation takes a continuous spiral. As the partner increases the
level of his trust, the relation becomes more valuable to him and, as
a result, he is less tempted to defect. The GRIT strategy and the

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 305

unconditional cooperation is needed, however, to increase the level

of trust gradually to the self-sustaining level through the spiral of
mutual trust and cooperation.
At least in the initial stage, cooperation gives rise to trust, not the
other way around. Separating trust from cooperation is especially
useful in this early stage of trust building, since it allows players to
be consistently cooperative while minimizing the risk of exploitation
by their partners. The double-TFT strategy that reciprocates in kind
– defection and reduced trust for defection, and cooperation and
increased trust for cooperation – is not a very useful strategy in
this early stage, since it destroys the partner’s fragile trust, which
is needed to get the whole process going. Turning the other cheek
can be the best strategy when your partner has no trust in you, pro-
vided that you protect your cheek well so that your enemy’s slap on
your cheek won’t hurt you much.
We end with two implications of our research with PD/D. First,
our findings suggest the need for society to encourage actors to
cooperate without trust (risk-taking). How can we as a society build
institutions to encourage initial cooperation without trust, so that
mutually trustful cooperative relationships can eventually emerge,
which the society will then not have to monitor or police? Second,
our findings suggest that, to the extent that individuals in natural
settings can separate trust from cooperation in order to eventually
achieve high levels of trustful cooperation, social psychologists’
heavy use of and reliance on PDs as a model of mixed-motive
game might not be appropriate and might actually underestimate
the extent and possibility of cooperation in society.

Acknowledgements

The experiments reported in this article were supported by grants to

the first author from the Japan Society for the Promotion of Science.
He is grateful to the Center for Advanced Study in the Behavioral
Sciences for providing an opportunity to prepare this manuscript.

NOTES

1. At each subsequent shift, the payoffs change by 10% of the original matrix – the
central matrix in Figure 1 – not 10% of the current matrix.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

306 RATIONALITY AND SOCIETY 17(3)

2. In this matrix, each player’s dominant choice is not to return (because T > R and
P > S). The total welfare (i.e. the sum of the two players’ payoffs) is greater when
both return than when neither returns (Pareto-superiority of mutual cooperation
to mutual defection). These features correspond to the defining elements of a PD
game.
3. The average cooperation rate and the F-value reported here are slightly different
from those reported in the original study since we include all subjects in our
analysis here whereas some subjects who indicated lack of understanding of the
experiment were excluded from the original analysis. The values reported here
are more conservative regarding the effect of the game. Figure 2 is based on Mat-
suda and Yamagishi’s figure.
4. Only the Random PD/Dc condition was newly added to the Japanese data.

REFERENCES

Arneson, R. J. 1982. ‘The Principle of Fairness and Free-rider Problem.’ Ethics 92:
616–33.
Axelrod, R. 1984. The Evolution of Cooperation. New York: Basic Books.
Berg, J., J. Dickhaut and K. McCabe. 1995. ‘Trust, Reciprocity, and Social History.’
Games & Economic Behavior 10: 122–42.
Bolton, G. E., E. Katok and A. Ockenfels. 2003. ‘How Effective Are Electronic Repu-
tation Mechanisms? An Experimental Investigation.’ Paper presented at the First
International Symposium on Online Reputation, Sloan School of Management,
Massachussets Institute of Technology, Boston, April 26–37.
Braithwaite, V. and M. Levi, eds. 1998. Trust and Governance. New York: Russell
Sage Foundation.
Cook, K. S., ed. 2001. Trust in Society. New York: Russell Sage Foundation.
Cook, K. S., T. Yamagishi, C. Cheshire, R. Cooper, M. Matsuda and R. Mashima.
2005. ‘Trust Building via Risk Taking: A Cross-Societal Experiment.’ Social
Psychology Quarterly 68: 121–42.
Dasgupta, P. 1988. ‘Trust as a Commodity.’ In Trust: Making and Breaking Coopera-
tive Relations, ed. D. Gambetta, pp. 49–72. Oxford: Blackwell.
Fukuyama, F. 1995. Trust: The Social Virtues and the Creation of Prosperity. New
York: Free Press.
Gambetta, D., ed. 1988. Trust: Making and Breaking Cooperative Relations. Oxford:
Basil Blackwell.
Hardin, R. 2002. Trust and Trustworthiness. New York: Russell Sage Foundation.
Hayashi, N., E. Ostrom, J. Walker and T. Yamagishi. 1999. ‘Reciprocity, Trust, and
the Sense of Control: A Cross-societal Study.’ Rationality and Society 11: 27–46.
Hofstede, G. H. 1980. Culture’s Consequences: International Differences in Work-
related Values. Beverly Hills, CA: Sage Publications.
Kakiuchi, R. and T. Yamagishi. 1997. ‘General Trust and the Dilemma of Variable
Interdependency.’ Research in Social Psychology 12: 212–21. (In Japanese with
an English abstract.)
Kiyonari, T., S. Tanida and T. Yamagishi. 2000. ‘Social Exchange and Reciprocity:
Confusion or Heuristic?’ Evolution and Human Behavior 21: 411–27.
Knack, S. and P. Keefer. 1997. ‘Does Social Capital Have an Economic Payoff?
A Cross-country Investigation.’ Quarterly Journal of Economics 112: 1251–88.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

YAMAGISHI ET AL.: SEPARATING TRUST FROM COOPERATION 307

Kollock, P. 1993. ‘Cooperation in an Uncertain World: An Experimental Study.’

Sociological Theory and Methods 8: 3–18.
Kramer, R. M. 1999. ‘Trust and Distrust in Organizations: Emerging Perspectives,
Enduring Questions.’ Annual Review of Psychology 50: 569–98.
Kreps, D. 1990. ‘Corporate Structure and Economic Theory.’ In Perspective on
Positive Political Economy, eds. J. Alt and K. Shepsle, pp. 90–143. Cambridge:
Cambridge University Press.
Lawler, E. J. 1986. ‘Bilateral Deterrence and Conflict Spiral: A Theoretical Analysis.’
In Advances in Group Processes, Vol. 3, ed. E. J. Lawler, pp. 107–30. Greenwich,
CT: JAI Press.
Lawler, E. J. and J. Yoon. 1993. ‘Power and the Emergence of Commitment Behavior
in Negotiated Exchange.’ American Sociological Review 58: 465–81.
Lawler, E. J. and J. Yoon. 1996. ‘Commitment in Exchange Relations: Test of a
Theory of Relational Cohesion.’ American Sociological Review 61: 89–108.
Lindskold, S. 1978. ‘Trust Development, the GRIT Proposal, and the Effects of
Conciliatory Acts on Conflict and Cooperation.’Psychological Bulletin 85: 772–93.
McKelvey, R. D. and Palfrey, T. R. 1992. ‘An experimental study of the centipede
game.’ Econometrica 60: 803–36.
Macy, M. W. 2002. ‘Review of Trust in Society.’ Contemporary Sociology 31: 473–5.
Mashima, R., T. Yamagishi and M. Macy. 2004. ‘Trust and Cooperation: A Compar-
ison of Ingroup Preference and Trust Behavior between American and Japanese
Students.’ The Japanese Journal of Psychology 75: 308–15. (In Japanese with an
English abstract.)
Matsuda, M. and T. Yamagishi. 2001. ‘Trust and Cooperation: An Experimental
Study of PD with Choice of Dependence.’ Japanese Journal of Psychology 72:
413–21. (In Japanese with an English abstract.)
Osgood, C. E. 1962. An Alternative to War or Surrender. Urbana: University of
Illinois Press.
Pruitt, D. G. and M. J. Kimmel. 1977. ‘Twenty Years of Experimental Gaming:
Critique, Synthesis, and Suggestions for the Future.’ Annual Review of Psychology
28: 363–92.
Putnam, R. D. 1993. Making Democracy Work: Civil Traditions in Modern Italy.
Princeton, NJ: Princeton University Press.
Rosenthal, R. 1981. ‘Games of Perfect Information, Predatory Pricing, and the Chain
Store Paradox.’ Journal of Economic Theory 25: 92–100.
Terai, S., Y. Morita and T. Yamagishi. 2003a. ‘Trust and Assurance in Ongoing Rela-
tions: An Experimental Study Using the Prisoner’s Dilemma with Variable Depen-
dence.’ Research in Social Psychology 18: 172–179. (In Japanese with an English
abstract.)
Terai, S., Y. Morita and T. Yamagishi. 2003b. ‘Trusting Behavior and Cooperative
Relation in the Selective Play Situation: An Experimental Study Using the
Prisoner’s Dilemma with Variable Dependence.’ Center for the Study of Cultural
and Ecological Foundations of Mind Working Paper Series, No. 14. (In Japanese
with an English abstract.)
Thibaut, J. W. and H. H. Kelley. 1959. The Social Psychology of Groups. New York:
Wiley.
Van Lange, P. A. M. and K. Visser. 1999. ‘Locomotion in Social Dilemma: How
People Adapt to Cooperative, Tit-for-tat, and Noncooperative Partners.’ Journal
of Personality and Social Psychology 77: 762–773.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015

308 RATIONALITY AND SOCIETY 17(3)

Yamagishi, T. 1998. The Structure of Trust: The Evolutionary Game of Mind and
Society. Tokyo: University of Tokyo Press. (In Japanese, English translation avail-
able at [Link]
Yamagishi, T., K. S. Cook and M. Watabe. 1998. ‘Uncertainty, Trust and Commit-
ment Formation in the United States and Japan.’ American Journal of Sociology
104: 165–194.
Yamagishi, T. and R. Kakiuchi. 2000. ‘It Takes Venturing into a Tiger’s Cave to Steal
a Baby Tiger: Experiments on the Development of Trust Relationships.’ In The
Management of Durable Relations, ed. W. Raub and J. Weesie, pp. 121–3. Amster-
dam: Thela Thesis Publishers.
Yamagishi, T. and K. Sato. 1986. ‘Motivational Bases of the Public Goods Problem.’
Journal of Personality and Social Psychology 50: 67–73.
Yamagishi, T. and M. Yamagishi. 1994. ‘Trust and Commitment in the United States
and Japan.’ Motivation and Emotion 18: 129–166.

TOSHIO YAMAGISHI is Professor of Social Psychology in the

Graduate School of Letters, Hokkaido University, Japan. His
current research interests include co-evolution of social institutions
and adaptive psychological mechanisms, and adaptive bases of
psychological mechanisms that sustain a system of generalized
exchange such as first-order and second-order punishment and in-
group favoring behavior.

ADDRESS: Graduate School of Letters, Hokkaido University,

N10 W7 Kita-ku, Sapporo, Japan 060–0810
[email: toshio@[Link]]

SATOSHI KANAZAWA is Lecturer in Management and Research

Methodology at the London School of Economics and Political
Science. His work on evolutionary psychology has appeared in
peer-reviewed journals in all social sciences (psychology, sociology,
economics, political science and anthropology) and in biology.

RIE MASHIMA is a doctoral student in the Graduate School of

Letters, Hokkaido University. She is conducting both theoretical
and experimental studies of strategies that can generate and sustain
a system of generalized exchange.

SHIGERU TERAI is a post-doctoral research fellow in the

Graduate School of Letters, Hokkaido University. He has just com-
pleted a PhD dissertation on cooperation and punishment between
American and Japanese groups.

Downloaded from [Link] at DEAKIN UNIV LIBRARY on March 13, 2015