ArticlePDF AvailableLiterature Review


Findings in the social psychology literatures on attitudes, social perception, and emotion demonstrate that social information processing involves embodiment, where embodiment refers both to actual bodily states and to simulations of experience in the brain's modality-specific systems for perception, action, and introspection. We show that embodiment underlies social information processing when the perceiver interacts with actual social objects (online cognition) and when the perceiver represents social objects in their absence (offline cognition). Although many empirical demonstrations of social embodiment exist, no particularly compelling account of them has been offered. We propose that theories of embodied cognition, such as the Perceptual Symbol Systems (PSS) account (Barsalou, 1999), explain and integrate these findings, and that they also suggest exciting new directions for research. We compare the PSS account to a variety of related proposals and show how it addresses criticisms that have previously posed problems for the general embodiment approach.
Embodiment in Attitudes, Social Perception, and Emotion
Paula M. Niedenthal
Laboratory in Social and Cognitive Psychology
CNRS and University of Clermont-Ferrand, France
Lawrence W. Barsalou
Department of Psychology
Emory University
Piotr Winkielman
Department of Psychology
University of California, San Diego
Silvia Krauth-Gruber and François Ric
Laboratory in Social Psychology
Université René Descartes, Paris
Findings in the social psychology literatures on attitudes, social perception, and
emotion demonstrate that social information processing involves embodiment, where
embodiment refers both to actual bodily states and to simulations of experience in the
brain’s modality-specific systems for perception, action, and introspection. We show
that embodiment underlies social information processing when the perceiver inter-
acts with actual social objects (online cognition) and when the perceiver represents
social objects in their absence (offline cognition). Although many empirical demon-
strations of social embodiment exist, no particularly compelling account of them has
been offered. We propose that theories of embodied cognition, such as the Perceptual
Symbol Systems (PSS) account (Barsalou, 1999), explain and integrate these find-
ings, and that they also suggest exciting new directions for research. We compare the
PSS account to a variety of related proposals and show how it addresses criticisms
that have previously posed problems for the general embodiment approach.
Consider the following findings. Wells and Petty
(1980) reported that nodding the head (as in agree
ment) while listening to persuasive messages led to
more positive attitudes toward the message content
than shaking the head (as in disagreement). Caciop
po, Priester, and Berntson (1993) observed that novel
Chinese ideographs presented during arm flexion (an
action associated with approach) were subsequently
evaluated more favorably than ideographs presented
during arm extension (an action associated with avoid
ance). Duclos et al. (1989) led participants to adopt
various bodily positions associated nonobviously with
fear, anger, and sadness and found that these postural
states modulated experienced affect. Strack, Martin,
and Stepper (1988) unobtrusively facilitated or inhib
ited the contraction of the zygomaticus (smiling) mus
cle by asking participants to hold a pen in their mouth
while they evaluated cartoons. Participants judged car
toons to be funnier when smiling was facilitated rather
than inhibited (see Stepper & Strack, 1993, for related
findings). Bargh, Chen, and Burrows (1996) showed
that participants in whom the elderly stereotype had
been primed subsequently walked down a hallway
more slowly than did participants in whom the stereo
type had not been primed. And Schubert (2004) showed
that making a fist influenced men’s and women’s auto
matic processing of words related to the concept of
All such findings suggest that the body is closely
tied to the processing of social and emotional informa
tion. No single theory, however, has integrated the
findings or explained them in a unified manner. Recent
theories of embodied cognition, which view knowl
Personality and Social Psychology Review
2005, Vol. 9, No. 3, 184–211
Copyright © 2005 by
Lawrence Erlbaum Associates, Inc.
The authors thank Vic Ferreira, Art Glenberg, Danny McIntosh,
Randy O’Reilly, and Cathy Reed for their helpful comments on vari
ous drafts of this article. We also thank the Society for Personality
and Social Psychology for awarding this article the SPSP 2003 The
oretical Innovation Prize. Preparation of this article was supported
by National Science Foundation grants BCS-0217294 to Piotr Winkiel
man and BCS-0350687 to Piotr Winkielman and Paula Niedenthal.
Requests for reprints should be sent to Paula M. Nieden
thal, LAPSCO, Université Blaise Pascal, 34, avenue Carnot, 63037
Clermont-Ferrand Cedex, FRANCE. E-mail: niedenthal@srvpsy. or to Piotr Winkielman, e-mail: pwinkiel@
edge acquisition and knowledge use as processes
grounded in the brain’s modality-specific systems,
hold promise of accounting for such findings and, per
haps most important, predicting the effects explicitly
and a priori (Barsalou, Niedenthal, Barbey, & Ruppert,
2003; Smith & Semin, 2004). Further, these recent the
ories are able to successfully address conceptual issues
that doomed previous embodiment proposals, making
them attractive alternatives to widely accepted amodal
theories of cognition. The aim of this article is to show
how that is so and to propose new ideas for the study of
information processing in social psychology.
The Notion of Embodied Mind
The nature of knowledge—the basic representa
tional elements of cognitive operations—lies at the
core of psychology and cognitive science. Our view of
what knowledge is determines how we conceptualize
perception, memory, judgment, reasoning, and even
emotion. It is generally agreed that the processing of
any mental content, including social and emotional
content, involves internal symbols of some sort—men-
tal representations. But this really just begs the ques-
tion. What are mental representations? Further, how do
they derive their meaning?—an issue known as the
symbol grounding problem (Harnad, 2003; Searle,
1980). If we can make progress on these questions, we
can put psychology in general and social psychology in
particular on firmer theoretical footing.
Amodal Architectures
Most models guiding current cognitive and social
psychology are based on the traditional computer met
aphor. This popular metaphor makes two major claims
about the mind. The first is that the software of the
mind is independent of the hardware of the body and
the brain (Block, 1995; Dennett, 1969). Thus, cogni
tive operations are arbitrarily related to their physical
instantiations so that any sufficiently complex physical
system could have human intelligence. In principle,
the software that constitutes the mind (including the
“social mind”) could run on anything—neurons, sili
con, or even wooden gears—as long as the elements
were arranged in proper functional relations. The sec
ond claim of the computer metaphor is that high-level
cognition, such as inference, categorization, and mem
ory, is performed using abstract, amodal symbols that
bear arbitrary relations to the perceptual states that pro
duce them (Newell & Simon, 1972; Pylyshyn, 1984).
Mental operations on these amodal representations are
performed by a central processing unit that is informa
tionally encapsulated from the input (sensory) and out
put (motor) subsystems (Fodor, 1983). The only func
tion of sensory systems is to deliver detailed
representations of the external world to the central unit.
The only function of the motor system is to dutifully
execute the central executive’s commands.
Recent years have witnessed a crumbling of the first
claim of the computer metaphor—that of software–
hardware independence. Research in cognitive and so
cial neuroscience has led to a growing appreciation
that most phenomena are best understood by jointly
considering neural, psychological, and situational con
straints (e.g., Brooks, 1991; Kosslyn, 1994; Winkiel
man, Berntson, & Cacioppo, 2001). Nevertheless, the
second claim of the computer metaphor lives on, and
many theories continue to assume that higher-order
cognition operates on amodal symbols. Noncontro
versially, these theories assume that the actual experi
ence of a current situation is initially represented in the
brain’s modality-specific systems. More controver
sially, standard theories of cognition assume that the
modality-specific states experienced during an actual
situation are redescribed and preserved in an abstract,
amodal, language-like form, which we will refer to as
amodal symbols (Fodor, 1975). For example, on inter
acting with a particular individual, amodal symbols
redescribe the experienced perceptions, actions, and
introspections to establish a conceptual representation
of the interaction in long-term memory.
As a person’s knowledge about such interactions
grows, the underlying amodal symbols become orga-
nized into structures that represent concepts extracted
across experience (e.g., Collins & Quillian, 1969).
These abstracted concepts constitute the person’s
knowledge and allow the person to engage in infer-
ence, categorization, memory, and other forms of
higher cognition. Nearly all accounts of social cogni
tion represent knowledge this way, using feature lists,
semantic networks, schemata, propositions, produc
tions, frames, statistical vectors, and so forth, to re
describe people’s perceptual, motor, and introspective
states (for discussions of such models, see Kunda,
1999; Smith, 1998; Wyer & Srull, 1984). According to
all such views, amodal redescriptions of social experi
ence constitute social knowledge.
The amodal architecture, although widely used, has
recently been criticized on several grounds. One set of
problems concerns the redescription process that pro
duces amodal symbols from modality-specific states in
the first place. No direct empirical evidence exists for
such a process in the brain. Indeed, surprisingly few
theoretical accounts of this redescription process exist
in the literature. More basically, there is no strong em
pirical case that the brain contains amodal symbols. In
fact, arguments for amodal architectures are mostly
theoretical, based on assumptions about how cognition
should work, rather than on empirical evidence that it
actually works this way. Further, as we discuss shortly,
empirical findings increasingly challenge the basic as
sumptions of the amodal architecture.
Given the lack of empirical evidence, why is the
amodal architecture so widely accepted in both cog
nitive and social psychology? There are a number of
important reasons. First, representations that employ
amodal symbols, such as semantic networks, feature
lists, schemata, and propositions, provide powerful
ways of expressing the content of knowledge across
various domains of knowledge, from perceptual im
ages to abstract concepts. Second, amodal symbols
provide a simple way to account for important func
tions of knowledge, such as categorization, categori
cal inference, memory, comprehension, language,
and thought (e.g., Anderson, 1983; Chomsky, 1959;
Newell, 1990; Newell & Simon, 1972). Third,
amodal symbols have allowed computers to imple
ment knowledge. Because frames, semantic net
works, and property lists have many similarities to
programming languages, these representations can be
implemented easily on computers, not only for theo
retical purposes, but also for applications (e.g., intel
ligent systems in industry, education, and medicine).
Fourth, until recently there were no compelling alter
natives that could account for the representation and
function of knowledge. For all these (good) reasons,
amodal approaches have dominated theories of repre-
sentation for decades, even though little positive em-
pirical evidence has accrued in their favor. Indeed, the
theoretical virtues of amodal approaches have been so
compelling that it has not occurred to most research-
ers that seeking empirical support might be necessary.
Instead, researchers typically assume that the amodal
architecture is roughly correct and then go on from
there to pursue their specific questions.
Amodal models of knowledge are widespread in
social psychology. This is true despite the fact that
social psychology has provided some of the most
compelling evidence for what we present here as an
alternative view, namely, theories of embodied cog
nition. As we summarize later, experimental findings
consistent with embodiment theories abound in re
search on attitudes, empathy, and emotion. Thus, we
contend that continuing with the common assumption
of amodal representation will lead us down a false
path and that advances in the understanding of
knowledge, in general, and social knowledge, in par
ticular, can be made if social psychology starts to
question the amodal architecture or at least looks
elsewhere for further inspiration.
Embodied Architectures
In recent years, researchers in psychology (Bar
salou, 1999; Glenberg & Robertson, 2000; Parsons et
al., 1995), philosophy (Churchland, Ramachandran,
& Sejnowski, 1994; Clark, 1997; Prinz, 2002; Varela,
Thompson, & Rosch, 1991), robotics (Brooks, 1991),
and linguistics (Lakoff & Johnson, 1999) have started
to take seriously the notion that knowledge is “em
bodied” or grounded in bodily states and in the
brain’s modality-specific systems.
It is important to
note that the term “embodiment” has been used in
multiple ways across the literature (Wilson, 2002).
Many earlier embodiment theories emphasized the
role of actual bodily states in cognition. Examples of
such theories include Piaget’s (1972) sensory–motor
account of infant memory, and Zajonc and Markus’s
(1984) hard-interface account of the interaction be
tween affect and cognition. In contrast, more contem
porary embodiment theories emphasize simulations
of experience in modality-specific systems. Examples
include Damasio’s (1994) theory of emotion,
Glenberg’s (1997) theory of memory, Barsalou’s
(1999) theory of perceptual symbol systems, and
Gallese’s (2003) theory of intersubjectivity. In the re
mainder of this article, we offer many specific exam
ples and evidence for embodiment processes in both
peripheral (body-based) and central (modality-based)
senses of the term embodiment. However, as will be
described in detail, our own theoretical perspective
primarily focuses on the central sense of embodi
ment, or the brain’s modality-specific systems. Those
systems include the sensory systems that underlie
perception of a current situation, the motor systems
that underlie action, and the introspective systems
that underlie conscious experiences of emotion, moti-
vation, and cognitive operations.
The main idea underlying all theories of embodied
cognition is that cognitive representations and opera-
tions are fundamentally grounded in their physical
context. Rather than relying solely on amodal ab-
stractions that exist independently of their physical
instantiation, cognition relies heavily on the brain’s
modality-specific systems and on actual bodily states.
One intuitive example is that empathy, or understand
ing of another person’s emotional state, comes from
mentally “re-creating” this person’s feelings in our
selves. The claim made by modern embodiment theo
ries is that all cognition, including high-level concep
tual processes, relies heavily on such grounding in
either the modalities or the body (Wilson, 2002). This
claim is significant given that embodiment theories
have traditionally been viewed as having little to say
about higher cognitive functions, not just empathy,
but also abstract concepts, categorical inference, and
the ability to combine internal symbols in novel, pro
ductive ways. As we will see, theories of embodied
cognition are increasingly able to explain how such
phenomena can be based in modality-specific sys
tems and bodily states.
Some of the more recent philosophical predecessors of embodi
ment theories can be found in writings of Ryle (1949), Merleau-
Ponty (1963), and Heidegger (1962). For further discussion, see
Prinz (2002).
Embodiment Effects
in Cognitive Psychology
Recent studies in cognitive psychology have dem
onstrated that conceptual knowledge is embodied (we
address the social literature shortly). As we review
some of these studies, we ask the reader to note two
things. First, note that these effects cannot be easily
predicted a priori by amodal theories, although as Bar
salou (1999) notes those theories can explain any ef
fect post hoc by adding increasingly complex assump
tions about representation and processing. Second,
note that whereas some studies we cite only show a
correlation between conceptual operations and modal
ity-specific systems, others provide direct causal evi
dence. The correlational results are useful because they
confirm a priori predictions derived from the embodi
ment account. But, it is increasingly essential to dem
onstrate the causal roles of embodiment in higher cog
nition. Fortunately, as we will see, many studies
experimentally manipulate embodiments across ran
domly assigned groups of participants, thereby demon
strating causal effects.
Online Embodiment
and Offline Embodiment
Wilson (2002) distinguished between online and
offline embodiment. The term online embodiment, and
the related term, situated cognition, refer to the idea
that much cognitive activity operates directly on real-
world environments. Accordingly, cognitive activity is
intimately tied to the relevant modality-specific pro-
cesses required to interact with the environment effec
tively. For example, when meeting a new individual
(e.g., a tall and imposing person), a perceiver sponta
neously produces in vivo sensory and somatic re
sponses (e.g., looking up and feeling apprehensive) as
well as motor responses (e.g., stepping back to keep
distance). The embodiment account views these sen
sory, somatic, and motor responses as necessary for the
encoding and interpretation of the new individual, not
simply as a by-product of a purely amodal analysis.
Another useful way to conceptualize online embodi
ment is as knowledge acquisition, with the perceiver
acquiring and modifying a repertoire of modality-spe
cific responses to stimuli as he or she interacts actively
with the social environment (also see Gallese, 2003). A
central tenet of recent theories is that the establishment
of this repertoire plays a central role in higher cogni
The term offline embodiment refers to the idea that
when cognitive activity is decoupled from the real-
world environment, cognitive operations continue to
be supported by processing in modality-specific sys
tems and bodily states. Just thinking about an object
produces embodied states as if the object were actually
there. Thus, perceiving a symbol, for example the name
of the previously met tall and imposing individual, can
produce embodied responses in the perceiver that un
derlie representation of the symbol’s meaning. For ex
ample, upward head orientation and defensive bodily
responding might implicitly contribute to the infer
ences that the individual is tall and imposing. A strong
embodiment view argues that the modality-specific
states engaged in during online cognition constitute the
knowledge that is acquired and later used in offline
cognition. According to this view, stored embodiments
constitute the basic elements of knowledge. To es
tablish the meaning of symbols during offline process
ing, people rely on repertoires of modality-specific re
sponses acquired previously during online processing
of these symbols’ referents.
Online Effects
The idea that modality-specific processes partici
pate in the conceptual processing of real world objects
can be illustrated with research on the compatibility
between motor actions and conceptual tasks. Tucker
and Ellis (1998) asked participants to detect whether a
cup was right side up or upside down. Although the
handle of the cup was irrelevant to the judgment, par-
ticipants responded faster when the cup’s handle was
on the same side of the display as the response hand
than when the handle was on the opposite side. This re-
sult indicates that representations of possible actions
(e.g., reaching for the cup) influence a perceptual judg-
ment even when these actions are not relevant to the
judgment. Reed and Farah (1995) asked participants to
judge whether two human figures depicted the same
posture. Participants asked to move their own arms
performed relatively better at detecting changes in the
arm position of a visually presented figure, whereas
participants asked to move their own legs did relatively
better at detecting changes in the figure’s legs. Again,
this finding suggests that representations of partici
pants’ own bodies contribute to the performance on the
visual task.
The just described behavioral studies are consistent
with neuroimaging research that found activation of the
grasping circuit when participants viewed manipulable
objects while lying passively in an fMRI scanner (Chao
& Martin, 2000). Related research with monkeys shows
that motor neurons involved in controlling tool use fire
when the tools are merely perceived and no motor re
sponse is possible (Rizzolatti & Arbib, 1998).
Offline Effects
Numerous studies have also documented offline
embodiment or the involvement of modality-specific
states when processing is decoupled from the environ
ment (as when the object is absent or represented
solely by a symbol, such as a word or a picture). For
example, in a study by Rauscher, Krauss, and Chen
(1996), participants first watched an animated action
cartoon. After a break, with the cartoon no longer pres
ent, participants were then asked to describe the car
toon to a listener. When participants were prevented
from gesturing (under the guise of recording skin con
ductance from their palms), they were significantly
slower to describe spatial elements of the cartoon. Pre
sumably, blocking the embodiment impaired access to
the conceptual elements of the representation. In an
other example, Spivey and his colleagues report that
participants who listen to vignettes including spatial
descriptions, such as “the top of a skyscraper” or “the
bottom of a canyon,” perform appropriate eye move
ments up or down, respectively, as if actually present
in the situation (Spivey, Tyler, Richardson, & Young,
2000). Finally, Glenberg and Kaschak (2002) found
that participants were faster at judging the sensibility
of a sentence when its meaning was compatible with
the hand movement required for the response (e.g.,
“Close the drawer”—forward movement; “Open the
drawer”—backward movement). Remarkably, this ac
tion–sentence compatibility effect occurred even when
the sentences referred to abstract actions that involved
directional communication (i.e., participants were fast-
est in judging the sensibility of the sentence “You
told Liz the story” with a forward movement and the
sentence “Liz told you the story” with a backward
movement). Richardson, Spivey, Barsalou, and
McRae (2003) report an analogous set of findings.
Again, these behavioral studies are consistent with
neuroscientific data. In the brain lesion literature,
many studies have found high-level cognitive impair
ments as a result of neurological damage to modal
ity-specific systems. Lesions in these systems produce
systematic deficits in category knowledge (e.g., Cree
& McRae, 2003; Damasio & Damasio, 1994; Farah,
1994; Humphries & Forde, 2001; Simmons & Bar
salou, 2003; Warrington & Shallice, 1984). Lesions in
modality-specific areas also produce deficits in the
representation of episodic memories (e.g., Rubin &
Greenberg, 1998). Recently, activation of modality-
specific areas has been observed when people perform
abstract conceptual tasks, such as concept property
verification, that require deep, nonassociative process
ing of target stimuli (Kan, Barsalou, Solomon, Minor,
& Thompson-Schill, in press).
In sum, accumulating evidence from cognitive psy
chology and cognitive neuroscience supports embodi
ment theories of knowledge. For more extensive re
views of such findings, see Martin (2001), Barsalou
(2003b), and Hegerty (2004). Importantly, such find
ings are not predicted a priori by amodal accounts. Ac
cordingly, these results are increasingly shaping theo
rizing in cognitive psychology and cognitive science,
as well as in philosophy and linguistics. For examples
of these theoretical accounts, see Barsalou (1999, 2003a),
Churchland et al. (1994), Clark (1997), Glenberg
(1997), Lakoff and Johnson (1999), Prinz (2002), and
Varela, Thompson, and Rosch (1991).
Embodiment Effects
in Social Psychology
Our primary argument here is that social psychol
ogy could profit from theories of embodied cognition.
In particular, such theories can help integrate and ac
count for the findings listed willy-nilly at the begin
ning of this article, as well as those just described. Fur
thermore, these theories can help us generate
interesting predictions that cannot be derived a priori
from the amodal accounts of knowledge representa
tions that currently dominate social psychology. To
support our argument, we next summarize findings
that illustrate embodiment in three traditional areas of
social psychology: attitudes, social perception, and
emotion. Then, we present a specific theory of embodi
ment, Barsalou’s (1999) PSS account of conceptual
processing. We compare PSS to some views already
present in the social psychology and related literatures
and explain how this account deals with prior criti-
cisms of the embodiment approach. Finally, we show
how PSS sheds new light on classic phenomena in so-
cial psychology.
In the following sections, we organize embodiment
findings around the themes of attitude, social percep-
tion, and emotion. Within each group of findings, we
distinguish again between online embodiment that oc-
curs in the presence of real external stimuli and offline
embodiment that occurs during the use of symbols that
refer to real stimuli not actually present. For example,
imitation of another person’s happy facial expression
is an example of online embodiment. On the other
hand, understanding the word “happiness” or recalling
a happy experience by recruiting modality-specific
systems is an example of offline embodiment. We find
the online/offline distinction useful in organizing the
social psychology literature, in part because it can
serve as a way to conceptualize knowledge acquisition
and knowledge use and to see similarities in their un
derlying mechanisms. Before we start, we hasten to
add that our summaries of these empirical literatures
should by no means be viewed as exhaustive. We sim
ply try to highlight findings that are representative for
each category (see Barsalou et al. 2003, for additional
discussion of these literatures).
Embodiment of Attitudes
Charles Darwin (1904) defined attitude as a col
lection of motor behaviors—especially posture—that
convey an organism’s affective response toward an ob
ject. Thus, it would not have come as any surprise to
him that the body is involved in the acquisition and use
of attitudes. Subsequent accounts similarly stressed the
importance of motor behavior in attitudes (e.g., Sher
rington, 1906; Washburn, 1926). Francis Galton
(1884), for example, also defined attitude in terms of
posture (literally, as a bodily inclination). He believed
that the way to quantify a person’s attitude toward an
other individual was to sit the two individuals in adja
cent chairs and then measure the weight that they ap
plied to the edge of the chair nearest to the other
individual (vs. the back of the chair). Individuals who
like one another should put more weight on the edge of
the chair facing the other person, he argued, thus mani
festing their positive attitude toward each other. Iron
ically, because a focus on posture no longer figures
into the definition of attitude, the following studies
may have more novelty value now than 100 years ago.
Online embodiment in the acquisition and pro
cessing of attitudes. The studies we summarize in
this section suggest that bodily responses during inter
action with novel objects influence later-reported atti
tudes and impressions. In an early demonstration of
such an effect, Wells and Petty (1980) instructed par-
ticipants to nod their heads vertically or to shake their
heads horizontally while wearing headphones, under
the pretext that the research was designed to investi-
gate whether the headphones slipped off as listeners
moved to the music. While nodding or shaking their
heads, participants then heard either a disagreeable or
an agreeable message about a university-related topic.
Later, they rated how much they agreed with the mes-
sage. Wells and Petty found that the earlier head move
ments later modulated participants’ judgments. Spe
cifically, participants who had nodded while hearing
the message were more favorable than participants
who had shaken their heads.
Tom, Pettersen, Law, Burton, and Cook (1991) ex
tended this study and induced participants to nod their
heads (in agreement) or to shake their heads (in dis
agreement) while placing a pen on the table in front of
participants. After the purported testing of the head
phones, a naive experimenter offered to give the partic
ipant the “old” pen that had been placed on the table
during the experiment or a “new” pen that the partici
pant had never seen. Individuals who had nodded dur
ing the testing of the headphones preferred the old pen,
whereas participants who had shaken their heads pre
ferred the new one. Presumably, whether participants
nodded or shook their heads during initial exposure to
the pen influenced the attitude that they developed to
ward it, as revealed in their later preference.
Cacioppo et al. (1993) explored the relation be
tween a different attitude-relevant motor behavior and
the development of attitudes toward completely novel
stimuli—Chinese ideographs. Participants were in
duced to push upward on a table from underneath (an
action typically associated with approach, reflective
of positive attitudes) or to push downward on the table
top (an action typically associated with avoidance, re
flective of negative attitudes) while they were exposed
to the ideographs. Consistent with an embodiment
hypothesis, ideographs seen during the approach be
havior were later rated more positively than were ideo
graphs seen during the avoidance behavior. A sub
sequent study in which a control group (that engaged in
no behavior) was added to the experimental design
demonstrated that both approach and avoidance behav
iors had a significant influence on attitudes.
As all these studies illustrate, bodily postures and
motor behavior are associated with positive and nega
tive inclinations and action tendencies toward objects.
Furthermore, these inclinations and tendencies influ
ence attitudes toward those objects as expressed by
self-report and attitude ratings. Thus, attitudes appear
to be determined, at least in part, by embodied re
sponses. We next look at the role of embodiment when
attitude objects are not present.
Offline embodiment in attitude processing. As
mentioned previously, the notion of offline embodi-
ment is that modality-specific systems are engaged
even when people process symbolic entities, such as
words. This is because conceptual processing draws on
the modality-specific patterns established earlier dur-
ing online acquisition processing. Furthermore, the
embodiment view proposes that conceptual processing
is maximally efficient when relevant conceptual infor-
mation is consistent with current embodiments.
This hypothesis is supported by the findings of
Chen and Bargh’s (1999) study in which participants
were exposed to words with positive or negative va
lence (e.g., love, hate) and had to report the valence by
pulling a lever toward them or by pushing it away.
Consistent with an embodiment prediction, partici
pants made pulling responses faster when responding
to positive words, compared to negative, and made
pushing responses faster to negative words compared
to positive ones. In a second study, Chen and Bargh
had participants indicate when a word merely appeared
on the computer screen—participants made the same
response to all words regardless of their affective va
lence. Subjects who indicated a word’s appearance by
pulling the lever toward them responded faster to posi
tive words than to negative ones. Participants who in
dicated a word’s appearance by pushing the lever away
responded faster to negative words. Thus, there was a
systematic relationship between the processing of the
word and the compatibility between the valence of the
word and the behavior used in response to it (see also
Neumann & Strack, 2000; Wentura, Rothermund, &
Bak, 2000).
Förster and Strack (1997, 1998) demonstrated a
similar effect in the retrieval of information from long-
term memory. Participants in their study generated the
names of famous people and later classified the people
according to whether they liked, disliked, or were neu
tral about them. During the name generation task, par
ticipants either pulled up on the table in front of them
from underneath its bottom surface (an approach be
havior, as described earlier) or pushed down on its
top surface (an avoidance behavior). Participants who
performed the approach behavior during name genera
tion retrieved more names of people they liked. Con
versely, participants who performed the avoidance ac
tion retrieved more names of people they disliked.
Thus, participants’ motor behavior influenced the re
trieval of attitude objects from long-term memory in an
attitude-congruent manner.
In sum, the studies described in this section on atti
tudes demonstrate the two embodiment effects of in
terest. First, during online exposure to objects, the pro
duction of motor movements associated with positive
attitudes leads to the later expression of positive at
titudes, and the production of motor movements as
sociated with negative attitudes leads to the later ex-
pression of negative attitudes. Second, during offline
cognition, processing symbols that stand for absent at-
titude objects are most efficient when a congruent mo-
tor behavior is maintained, suggesting that represent-
ing the conceptual knowledge involves the relevant
motor behavior.
Social Perception
It may feel intuitively correct to learn that individu
als embody the behaviors of others online or when
those others are physically present. Researchers have
long argued for the role of mimicry and imitation in so
cial modeling, social coordination, and empathy (e.g.,
Bandura, 1977; Lipps, 1907). It is perhaps more coun
terintuitive to imagine how embodiment enters into so
cial perception offline, or when other people are pres
ent only symbolically. This notion has some
interesting historical precursors. Freud, of course,
thought that the sensory–motor symptoms of hysterics
were in fact unconscious enactments of thoughts and
memories involving other significant people (Breuer
& Freud, 1983-95/1955). And in a more general em
bodiment account, psychoanalyst Felix Deutsch
(1952) proposed that “all automatic movements
represent in some way the search for a desired [per
son] from the past” (p. 210). He argued, based on
clinical cases, that parts of the body actually personify
members of one’s family (even, or especially, when
they are not physically present) and that sensations and
movements in those body parts express feelings to
ward and memories of those people. These unique per
spectives have rarely been integrated explicitly with
empirical research on social perception and impression
formation. Still, the social psychology literature is re
plete with examples that are consistent with such a
Online embodiment in social information pro
cessing: Mimicry and imitation. Research has con
sistently shown that perceivers imitate the facial ges-
tures of perceived others. O’Toole and Dubin (1968)
demonstrated that mothers open their mouths in re-
sponse to the open mouth of their infant who is about to
feed. These imitative behaviors occur very early in de-
velopment. In classic studies, Meltzoff and Moore
(1977, 1989) showed that neonates imitate basic facial
gestures such as tongue protrusion and mouth opening,
suggesting a biological basis of basic imitation skills
(for a review, see Meltzoff & Prinz, 2002). The biolog-
ical argument is strengthened further by observations
of basic imitative behaviors in other primates (Preston
& de Waal, 2002) and by the impairments of basic imi
tation as a result of developmental disorders such as
autism (McIntosh, Reichmann-Decker, Winkielman,
& Wilbarger, 2004; Rogers, 1999; Sigman, Kasari,
Kwon, & Yirmiya, 1992).
Importantly, imitation extends beyond facial behav
ior. Individuals engaged in conversation tend to syn
chronize their latency and rate of speech, the duration
of their utterances, and other speech characteristics
(e.g., Capella & Planalp, 1981; Matarazzo & Wiens,
1972; Webb, 1972). Listeners also tend to mimic talk
ers’ emotional prosody (e.g., Neumann & Strack,
2000), manual gestures (e.g., Bavelas, Black, Chovil,
Lemery, & Mullett, 1988; Maxwell, Cook, & Burr,
1985), and even their syntactic constructions (e.g.,
Bock, 1986). Much research has focused on postural
synchrony. For example, in one study Bernieri (1988)
had judges code the postures of two individuals filmed
while they were actually interacting with each other
and the same two individuals who appeared to be inter
acting with each other but who were actually interact
ing with different people. Supporting the idea of imita
As we discuss later, the embodiment view allows for substantial
flexibility in terms of what specific modality states and physical
movements are associated with specific concepts. For example, un
der different task settings, positive versus negative concepts might
be associated with different bodily movements (push or pull, up or
down, toward or away). The main message of the embodiment view
is that conceptual representations are supported by simulations in
modality systems, not that they are rigidly tied to specific bodily
states. This view is consistent with recent findings on flexibility of
the link between valence and specific bodily movement (Markman
& Brendl, in press).
Niedenthal and Halberstadt (2004) have argued that the term
social perception as used in social psychology is a misnomer, be
cause rarely have perceptual processes actually been examined.
Here, similar to most other researchers, we use this term broadly in
referring to impression formation and social information processing.
tion, the results revealed greater postural synchrony for
two individuals engaged in actual interaction than for
two individuals in a contrived interaction (for related
results, see Bernieri, Reznick, & Rosenthal, 1988; Ber
nieri & Rosenthal, 1991). In a more recent study by
Chartrand and Bargh (1999), a trained experimenter
rubbed her nose or shook her foot while she interacted
with participants. When the experimenter scratched
her nose, participants were more likely to scratch their
nose (than to shake their foot). And when the experi
menter shook her foot, participants were more likely to
shake their foot (than to scratch their nose). This result
further suggests that the mere observation of another
person engaging in a particular action facilitates the
same action in the perceiver.
It is widely believed that synchrony facilitates co
operation and empathy among interaction partners
(e.g., Hatfield, Cacioppo, & Rapson, 1993; LaFrance,
1985; LaFrance & Ickes, 1981; Neumann & Strack,
2000; Semin, 2000). Consistent with this belief, en
hancing mimicry increases smoothness of interaction
and liking between partners (Chartrand & Bargh,
1999). It is also believed that interpersonal closeness
facilitates mimicry (Bush, McHugo, & Lanzetta, 1989;
McIntosh, Druckman, & Zajonc, 1994; van Baaren,
Maddux, Chartrand, de Bouter, & Knippenberg,
2003). Consistent with this assumption, participants
show more spontaneous mimicry of a model’s behav-
ior when their liking for the model is experimentally
enhanced or when the model is the participant’s friend
(McIntosh, in press).
A specific neural mechanism appears to underlie
imitation. Rizzolatti and his colleagues have observed
that certain circuits involved in the production of an or
ganism’s motor behavior also become active in re
sponse to perceived intentional motor behavior (e.g.,
Rizzolatti, Fadiga, Fogassi, & Gallese, 2002; also see
Chao & Martin, 2000). Such mirror neuron circuits
could play two important adaptive roles in social life.
First, they may support fast learning, such that an or
ganism learns new actions through imitation (Gallese,
2003). This is consistent with our idea of online em
bodiment as knowledge acquisition. Second, these cir
cuits may be responsible for social contagion, such as
the induction of congruent emotional states in others,
especially if those others are psychologically close
(Decéty & Chaminade, 2003). If so, then these circuits
should be centrally implicated in empathy and social
cooperation (Hatfield, Cacioppo, & Rapson, 1992).
More generally, the growing evidence for the close
involvement of mirror neurons in empathy, imitation,
and attribution of mental state supports the embodi
ment view. Reflecting on this evidence, Gallese (2003)
noted that according to amodal accounts of social
when faced with the problem of understanding the
meaning of others’ behaviors, adults humans must
necessarily translate the sensory information about the
observed behavior into a series of mental representa
tions that share, with language, the propositional for
mat. This enables one to ascribe others’ intentions,
desires and beliefs, and therefore to understand the
mental antecedents of their overt behavior [but this
is a] disembodied view I think that there is now
enough empirical evidence to reject a disembodied
theory of the mind as biologically implausible.
Offline embodiment in social information pro
cessing: Category priming and motor responding.
In a recent and already classic series of studies, Bargh,
Chen, and Burrows (1996) demonstrated embodiment
in social information processing when the actual social
stimuli were not present. Participants in one study were
instructed to form sentences from groups of words pre
sented in random order. In the critical conditions,
many of the sentences contained words related to the
stereotype of the elderly (e.g., gray, Florida, and
bingo). Importantly, those words were not specifically
related to motor movement. In the control conditions,
the sentences that participants constructed contained
neutral words unrelated to the elderly stereotype. Re-
sults showed that when participants had been primed
with the elderly stereotype, they actually took longer to
walk from the experimental room to the elevator than
did control participants. Presumably, this occurred be-
cause the priming task activated the elderly stereotype
which contains knowledge that old people tend to
move slowly. In turn, this knowledge activated action
schemas, which caused the embodiment effect of
slow walking. Studies in the same research project
have demonstrated embodiment effects of processing
other social categories on other kinds of behavior, in
cluding rudeness and aggressiveness (for a review, see
Dijksterhuis & Bargh, 2001).
Another class of embodiment effects in offline so
cial perception occurs during activation of evaluative
knowledge. In one demonstration, Vanman, Paul, Ito,
and Miller (1997) instructed participants to form im
ages of people with whom they might later work on a
problem-solving task. A number of different variables
moderated participants’ facial responses, as measured
by EMG. For example, participants were most likely to
display positive facial reactions when their imagined
partners, who were competent (vs. incompetent), ex
erted high (vs. low) effort, or belonged to the same (vs.
different) racial category. Thus embodiment occurred
when individuals activated representations of people
who were not actually present.
Finally, Andersen, Reznik, and Manzella (1996)
obtained personality descriptions about significant
others in their participants’ lives, and then developed
descriptions of fictional characters who partially re
sembled them. In a later experimental session, partici
pants read the descriptions of these fictional charac
ters, not realizing that they were in any way related to
their significant others. In line with an embodiment ac
count, participants’ facial expressions, as coded by a
naive judge, were influenced by what they read. When
participants read about characters based on significant
others they liked, they tended to produce positive facial
expressions. Conversely, when participants read about
characters based on significant others who they dis
liked, they tended to produce negative facial expres
sions. Thus, again, simply reading about abstract social
stimuli influenced facial responding, suggesting an
embodied representation of social knowledge.
To conclude this section, we have described illus
trative studies showing, first, that individuals tend to
mimic the motor behavior of other individuals when
those others are actually present and, second, that em
bodied responses are engaged when individuals ma
nipulate information offline about other people stored
in long-term memory. Recent findings in social neur
oscience also provide strong evidence for this inter
pretation. All these findings are closely related to the
research findings on emotion and empathy that we ad-
dress next.
Students of emotion most often associate the idea of
embodiment with William James (1890). According to
James, the basis of emotion is the bodily activity that
occurs in response to an emotional stimulus. Thus,
James was claiming that emotions are embodiments.
Our goal here is not to assess whether everything about
James’s theory is correct. In particular, we are not en
dorsing James’s claim about the necessity of the auto
nomic nervous system for emotion. We are also not en
dorsing a strong mapping between specific emotions
and specific embodiments. We do suggest, however,
that embodiment is critically involved in information
processing about emotion—not only “online,” when
people respond to real emotion objects, but also “off
line,” when people represent the meanings of emo
tional symbols, such as words. Furthermore, as de
scribed shortly, we propose a concrete embodiment
account grounded in current research in psycholo
gy and neuroscience that allows us to make specific
predictions about the role of embodiment in emotion
Online embodiment in emotion processing.
We have already discussed the ubiquity of embodied
responses to nonemotional actions and gestures. In
this section, we focus on embodied responses to emo
tion stimuli. Evidence is accumulating that people
mimic others’ emotional facial expressions (Bush,
Barr, McHugo, & Lanzetta, 1989; Dimberg, 1982). In
Bavelas, Black, Lemery, and Mullett (1986), for exam
ple, a confederate faked an injury and then grimaced
in pain. When participants observed the grimace, they
grimaced themselves. Furthermore, the magnitude of
participants’ grimaces increased with how clearly they
could see the confederate’s grimace. Emotion imi
tation appears to be relatively automatic and to even
be elicited outside awareness, as when participants
react with slight smiles and frowns to subliminal
happy and angry expressions (Dimberg, Thunberg, &
Elmehed, 2000). Further evidence suggests that em
bodied consequences of subliminal facial expressions
extend beyond facial mimicry. In one study, for exam
ple, participants were first subliminally exposed to
happy or angry faces and were then asked to try a novel
beverage. The results showed that participants exposed
to subliminal happy faces later behaved more in an
approach-oriented fashion (by pouring and drinking
more beverage) than subjects who were exposed to
subliminal angry faces (Winkielman, Berridge, & Wil-
barger, in press).
According to embodiment views, bodily responses
should facilitate cognitive processing of emotion stim-
uli. In one demonstration of this effect, Wallbott
(1991) had participants categorize the emotional facial
expressions displayed in photographs of other people.
As participants categorized the photographed expres
sions, their own faces were surreptitiously videotaped.
Results showed that the participants tended to mimic
the facial expressions as they categorized them. When
they categorized happy faces, for example, they smiled
themselves. Furthermore, participants’ accuracy in
classifying the facial expressions was positively corre
lated with the extent of mimicry. The more participants
mimicked the faces, the better they were at discerning
what expression the face was displaying.
Niedenthal, Brauer, Halberstadt, and Innes-Ker
(2001) demonstrated that this mimicry plays a causal
role in the processing of emotional expression. Partici
pants watched one facial expression morph into an
other and had to detect when the expression changed.
Some participants were free to mimic, whereas others
were prevented from mimicking by holding a pencil
laterally between their lips and teeth. Consistent with
the embodiment hypothesis, participants free to mimic
the expressions detected the change in emotional ex
pression earlier (more efficiently) for any facial ex
pression than did participants who were prevented
from mimicking the expressions (for further discus
sion see Niedenthal, Ric, & Krauth-Gruber, 2002).
More specifically, James argued that the conscious experience
of emotion (the subjective feeling component of emotion) derives
from the conscious perception of embodiments. Note, however, that
emotion can be embodied without these embodiments being con
sciously represented as feelings (for discussion, see Berridge &
Winkielman, 2003).
Adolphs, Damasio, Tranel, Cooper, and Damasio
(2000) report further evidence for the causal involve
ment of somatosensory processes in recognition of
facial expressions. Clinical patients with lesions in
somatosensory cortex showed poorer performance in
classifying facial expressions than individuals without
such lesions. Presumably, simulating emotional ex
pressions on one’s own face, and experiencing the re
sulting somatosensory feedback, is necessary for the
process of recognition.
In short, the results reported
by Wallbott (1991), Niedenthal et al. (2001), and
Adolphs and his colleagues all converge on the conclu
sion that feedback from facial mimicry is importantly
involved in a perceiver’s ability to process emotional
expressions. Carr, Iacoboni, Dubeau, Mazziotta, and
Lenzi (2003) have begun to explore the neural circuit
that underlies this process.
We have noted several times that the mimicry of
emotional gestures has been proposed as a mechanism
that supports empathy (Lanzetta & Englis, 1989; McIn
tosh et al., 1994; Vaughan & Lanzetta, 1980; see
Levenson, 1996 for discussion). Zajonc, Adelmann,
Murphy, and Niedenthal (1987) further demonstrate
this relationship. These researchers compared the facial
similarity of couples at the time of their marriage to their
facial similarity after 25 or more years of marriage.
Zajonc et al. reasoned that married partners should fre-
quently mimic each other’s facial expressions, presum-
ably because they are particularly motivated to empa-
thize with each other. As a consequence of this frequent
mimicry, the couples’ faces should grow more similar
over time. Consistent with this reasoning, Zajonc et al.
found that after 25 or more years of marriage, facialsim-
ilarity between couples was greater than at the time of
their marriage and also greater than between random
people ofthe same age. Furthermore, this effect was cor
related with the quality of the marriage and therefore
presumably success in empathizing.
The rather indirect finding of Zajonc and his col
leagues (1987) is supported by recent studies on the
neural basis of mirroring effects. One fMRI study ob
served very similar changes in brain activity of a fe
male participant while painful stimulation was applied
to her own hand or to her partner’s hand (Singer et al.,
2004). A related study used single cell recording and
found activation of pain-related neurons when a pain
ful stimulus was applied to the participant’s own hand
and when the patient watched the painful stimulus ap
plied to the experimenter’s hand (Hutchison, Davis,
Lozano, Tasker, & Dostrovsky, 1999). Yet another
study found activation in the insula (a brain area re
sponsible for processing somatosensory information)
when the participant was exposed to disgusting odors
and when the participant simply watched a movie of
other people’s expressions of disgust (Wicker et al.,
2003). All these findings were interpreted as evidence
of an embodied simulation in the perceiver of what was
happening to the perceived person (for summaries of
related research, see Gallese, 2003; Iacoboni, in press).
In short, there is now converging evidence that em
bodiment, in the sense of actual motor and soma
tosensory responses, is involved in empathy.
Offline embodiment in emotion. As discussed
earlier, offline embodiment is the use of modality-spe
cific representations to represent the meaning of sym
bols whose referents are absent. Offline embodiment
often appears central to the representation of emotion
knowledge. In a study by Laird, Wagener, Halal, and
Szegda (1982), participants studied both anger-pro
voking and happiness-provoking material. Later, par
ticipants were covertly induced to smile or frown and
then were instructed to recall the earlier learned mate
rial. Results showed that the induced expression mod
erated recall. Participants induced to smile recalled the
happy material better than those induced to frown,
whereas participants induced to frown recalled the an-
gry material better than those induced to smile. Impor-
tantly, this effect was found only when participants’ fa-
cial expressions were accompanied by a congruent
emotional state. That is, participants’ memory perfor-
mance was maximized when the motor behavior, the
emotional state, and the emotional meaning of the
learned material were all compatible (see related stud-
ies by Strack et al., 1988; Zajonc, Pietromonaco, &
Bargh, 1982).
In a similar study by Riskind (1984), participants
were instructed to retrieve pleasant or unpleasant auto
biographical memories while they adopted different
postures and facial expressions. The embodiment manip
ulation was expected to influence the emotional nature
of the memories recalled. As predicted, postural and fa
cial manipulations modulated the latencies to retrieve
positive versus negative life experiences. Adopting an
erect posture and smiling speeded the retrieval of pleas
ant autobiographical memories, relative to the speed of
retrieving unpleasant autobiographical memories.
Research by Stepper and Strack (1993) generalized
these effects to how people respond emotionally to
evaluations of their performance. Participants were led
to sit in an upright or slumped position under the pre
text that the experimenters were concerned with task
performance under varying ergonomic conditions.
While upright or slumped, participants performed an
achievement test and received bogus feedback that
they had done well. Later, participants rated their feel
ing of pride at the time. Participants who had sat up
right while receiving task feedback reported experi
Somatosensory mechanisms are not only involved in recogniz
ing facial expressions. Similar effects of damage to the somatosen
sory cortex have also been obtained with tasks requiring emotion
recognition from prosody and body movement (Adolphs, Damasio,
& Tranel, 2002; Heberlein, Adolphs, Tranel, & Damasio, in press).
encing more pride than participants who had sat in a
slumped position (see also Riskind & Gotay, 1982).
To summarize, these findings provide strong evi
dence for the embodiment of emotion processing. This
evidence also suggests that embodiment supports im
portant cognitive and social functions, such as recogni
tion, memory, empathy, and understanding. As such,
the research on emotion embodiment is consistent with
earlier reviewed research on embodiment in social per
ception and attitudes and points to the critical impor
tance of modality-specific states in the representation
and processing of social knowledge. But just how does
this work?
An Integrative Theory of Embodiment
For over 30 years, evidence has accumulated that
implicates embodiment centrally in the acquisition and
expression of attitudes, in social perception, and in the
learning and use of emotion knowledge. Despite all
this evidence, however, no major theory has explained
it (Smith & Semin, 2004). Furthermore, the common
interpretation of such findings is that embodiment ef-
fects arise “simply” as peripheral by-products of con-
ceptual knowledge, which is typically viewed as the
critical cause of social cognition. Important accounts
of the relationship between perception and action have
been proposed in recent years (e.g., Bargh & Char-
trand, 1999; Neumann, Förster, & Strack, 2003). These
accounts, however, have emphasized the direct and au-
tomatic nature of the relation between perceptual and
motor processes. In contrast, we explicitly focus on the
dynamical role of modality-specific systems in repre
senting and manipulating conceptual knowledge. In
what follows, we describe the PSS theory, which puts
embodiment at the core of information processing,
including attitudes, social perception, and emotion.
Comprehensive presentation of PSS can be found else
where (e.g., Barsalou, 1999, 2003a, 2003b). Our pur
pose here is to describe the account in enough detail to
show how it explains embodiment effects in social
psychology, how it predicts novel phenomena, and
how it compares to other accounts of social informa
tion processing (also see Barsalou et al., 2003).
Overview of the PSS Theory
According to PSS, the modality-specific states that
represent perception, action, and introspection in on
line situations are also used to represent these situa
tions in the offline processing that underlies memory,
language, and thought. Rather than using amodal
redescriptions of online modality-specific states to
represent these situations, the cognitive system uses
reenactments (simulations) of them instead. Thus, the
key notion in PSS is that simulations of perceptual,
motor, and introspective experience underlie the repre
sentation and processing of knowledge. In the
following sections, we describe what we mean by mo
dality-specific representations in further detail and
then address how such representations are used in the
simulations that underlie conceptual processing.
Feature Maps and Convergence Zones
The PSS account takes as a starting point Damasio’s
(1989) theory of convergence zones (CZ) proposed by
Damasio and his colleagues (see Simmons & Barsalou,
2003, for an elaborated account). CZ theory assumes
that the perception of an object activates relevant fea
ture detectors in the brain’s modality-specific systems.
The populations of neurons that code featural informa
tion in a particular modality are organized in hierarchi
cal and distributed systems of feature maps (Palmer,
1999; Zeki, 1993). When a stimulus is perceived on a
given modality, populations of neurons in relevant
maps code the stimulus’ features on that modality in a
hierarchical manner. For example, visual processing of
a happy face activates feature detectors that respond to
the color, orientation, and planar surfaces of the face.
Whereas feature detectors early in the processing
stream code detailed perspective-based properties of
the face, higher-order detectors code its more abstract
and invariant properties. The pattern of activation
across relevant features maps represents the face in vi-
sual processing.
Analogously, CZ theory assumes that systems of
feature maps reside in the other sensory–motor modal-
ities and in the limbic system for emotion. All these
maps operate in parallel, so that while a face is being
represented in visual feature maps, sounds produced
by the face are being coded in auditory feature maps,
affective responses to the face are being coded in
limbic feature maps, bodily responses to it are being
coded in motor feature maps, and so forth.
CZ theory further proposes that conjunctive neu
rons in the brain’s association areas capture and store
the patterns of activation in feature maps for later rep
resentational purposes in language, memory, and
thought. Damasio (1989) referred to these association
areas as convergence zones. Like feature maps, CZs
are organized hierarchically such that the CZs located
in a particular modality-specific system (e.g., vision)
capture patterns of activation in that modality. In turn,
higher-level CZs conjoin patterns of activation across
modalities. What this means is that when we hear a
sound (e.g., a fire cracker exploding), conjunctive neu
rons in auditory CZs capture the pattern of activation in
auditory feature maps. Other conjunctive neurons in
motor CZs capture the pattern of activation caused by
jumping away from the location of the sudden sound.
And at a higher level of associative processing, con
junctive neurons in modality-specific CZs conjoin the
two sets of modality-specific conjunctive neurons for
the combined processing of sound and movement.
It is worth highlighting how the CZ architecture dif
fers from traditional ways of conceptualizing knowl
edge acquisition and use. First, during knowledge ac
quisition (perception and learning), all relevant
processing regions participate in knowledge represen
tation—there is no single “final” region where all ex
perience is abstracted and integrated together. Higher
level CZs capture only conjunctions of lower-level
zones (so that CZs can later coordinate their feature-
level reactivation)—they do not constitute some form
of “grand” representation that independently repre
sents all lower levels of the representational hierarchy.
Second, during knowledge use (e.g., conceptual pro
cessing and recall), the cognizer activates the multiple
modality-specific regions that encoded the experience,
rather than, as traditionally assumed, only the “final”
abstract regions at the end of the processing streams.
Reenactments of Modality-
Specific States
What is important about the CZ architecture is the
idea that conjunctive neurons can later reactivate the
states of processing in each modality and across mo-
dalities, without any input from the original stimulus.
This mechanism provides a powerful way to imple-
ment offline embodiment. The modality-specific pro-
cessing that occurred in reaction to a previously en-
countered stimulus can be reenacted without the
original stimulus being present. For example, when re-
trieving the memory of a person’s face, conjunctive
neurons partially reactivate the visual states active
while perceiving it. Similarly, when retrieving an ac
tion, conjunctive neurons partially activate the motor
states that produced it. Indeed, this reentrant mecha
nism is now widely viewed as underlying mental imag
ery in working memory (e.g., Farah, 2000; Grezes &
Decéty, 2001; Kosslyn, 1994).
Importantly—and different from typical assump
tions about imagery—the CZ architecture and PSS do
not require the reenactment process to be conscious.
Explicit construction of mental imagery can certainly
produce compelling reenactments, which are often
viewed as the process that underlies, for example,
counterfactual simulations. Nevertheless, memory, con
ceptualization, comprehension, and reasoning pro
cesses may rely heavily on unconscious reenactments
(e.g., Barsalou, 1999, 2003b). Many of the demonstra
tions illustrated in the previous sections of this article
may primarily reflect spontaneous and unconscious
reenactments (e.g., several phenomena reported by
Bargh and his colleagues).
Simulators and Simulations
The CZ architecture describes how knowledge is
distributed across the brain’s feature and associa
tion areas. To explain how the cognitive system ac
tually uses that knowledge, PSS relies on two cen
tral constructs: simulators and simulations (Barsalou,
1999). Simulators integrate modality-specific infor
mation across a category’s instances. Simulations im
plement specific conceptualizations of a category. We
describe these constructs in turn.
Simulators. A sizable literature on concepts has
demonstrated that categories possess statistically cor
related features (e.g., Chin & Ross, 2002; Rosch &
Mervis, 1975). Thus, when different instances of the
same category are encountered over time and space,
they activate similar neural patterns in feature maps
(cf., Cree & McRae, 2003; Farah & McClelland,
1991). One result of this repeated firing of similar neu
ral patterns is that similar populations of conjunctive
neurons in CZs respond to these regular patterns
(Damasio, 1989; Simmons & Barsalou, 2003). Similar
to the notion of abstraction, over time, these groups of
conjunctive neurons integrate modality-specific fea-
tures of specific categories across their instances and
across the situations in which they are encountered.
This repetition establishes a multimodal representation
of the category: a concept.
PSS refers to these multimodal representations of
categories as simulators (Barsalou, 1999, 2003a). A
simulator integrates the modality-specific content of
a category across instances and provides the ability to
identify items encountered subsequently as instances
of the same category. Consider a simulator for the so
cial category, politician. Following exposure to differ
ent politicians, visual information about how typical
politicians look (i.e., based on their typical age, sex,
and role constraints on their dress and their facial
expressions) becomes integrated in the simulator,
along with auditory information for how they typically
sound when they talk (or scream or grovel), motor pro
grams for interacting with them, typical emotional re
sponses induced in interactions or exposures to them,
and so forth. The consequence is a system distributed
throughout the brain’s feature and association areas
that essentially represents knowledge of the social cat
egory, politician.
According to PSS, a simulator develops for any as
pect of experience attended to repeatedly. Because at
tention is highly flexible, it can focus on diverse com
ponents of experience, including objects (e.g., chairs),
properties (e.g., red), people (e.g., politicians), mental
states (e.g., disgust), motivational states (e.g., hunger),
actions (e.g., walking), events (e.g., dinners), settings
(e.g., restaurants), relations (e.g., above), and so forth.
Across development, a huge number of simulators de
velop in long-term memory, each drawing on the rele
vant set of feature and association areas needed to rep
resent it. Once this system is in place, it can be used to
simulate those aspects of experience for which simula
tors exist. Furthermore, as discussed later, simulators
can combine to construct complex representations that
are componential, relational, and hierarchical. Thus
PSS is not a theory of holistic images. In contrast to
how theories like PSS are often mistakenly viewed,
photo-like images of external scenes do not underlie
knowledge. Instead, componential bodies of accumu
lated information about the modality-specific compo
nents of experience underlie knowledge, where these
components can represent either the external environ
ment or the internal states of the agent.
Computational implementations of simulators have
not yet been developed. Clearly such development is
important for many reasons. Nevertheless a variety of
computational architectures exist that could potentially
be used to implement them. For example, object-ori
ented (as opposed to bit-mapped) drawing programs
contain much of the componential, relational, and hier
archical structure that underlies PSS’s productive use
of simulations. Although this architecture has not been
developed as a psychological theory, its functionality
closely resembles many of PSS’s conceptual opera-
tions. Another relevant computational approach is the
interactive neural network architecture, which repre-
sents categories as distributed feature profiles across
modality-specific areas (Farah & McClelland, 1991;
O’Reilly & Munakata, 2000). In these networks, mo-
dality-specific representations are used not only for es-
tablishing conceptual knowledge, but also for perform-
ing high-level conceptual operations. The success of
these existing architectures offers evidence for the
mechanistic plausibility of the simulators and simula
tions in PSS.
Simulations. The use of simulators in conceptual
processing is called simulation. A given simulator can
produce an infinite number of simulations, namely,
specific representations of the category that the simu
lator represents. On a given occasion, a subset of the
modality-specific knowledge in the simulator becomes
active to represent the category, with this subset vary
ing widely across simulations. For example, a simula
tor that represents the social category, my significant
other, might be used to simulate love making with a
significant other on one occasion, to simulate fights on
another, to simulate quiet togetherness on another, and
so forth. A simulation can be viewed as the reverse
process of storing modality-specific information in a
simulator. Whereas learning involves feature map in
formation becoming linked together by conjunctive
units in CZs, simulation involves later using these con
junctive units to trigger feature map information. Thus,
a simulation, too, is a distributed representation.
According to PSS, the simulation process is highly
dynamic and context dependent. It is dynamic in that a
given simulator can, in principle, produce an infinite
number of simulations. Depending on the current state
of the simulator, the current state of associated simula
tors, the current state of broader cognitive processing,
and so forth, a unique simulation results (Barsalou,
1987, 1989, 1993, 2003b). The simulation process is
context dependent in that the simulation constructed
on a given occasion is tailored to support situated ac
tion (Barsalou, 2002, 2003b). Ideally, the current sim
ulation of a category should provide useful inferences
about specific category members currently being expe
rienced (or likely to be experienced), actions that could
be performed on them, mental states that might result,
and so forth. Thus, PSS does not view the simulation
process as producing a static, generic category repre
sentation. Instead, PSS views simulation as a skill or
competency for representing a particular category flex
ibly in myriad ways that support successful interac
tions with its members.
Notably, simulations do not implement a single rep
resentational type, such as exemplars. Instead, sim
ulations can implement a variety of representational
types, including exemplars, prototypes, and rules (Bar-
salou, 1999, 2003a). To the extent that a simulation re-
enacts the modality-specific states of a particular ex-
perience with a category member, it represents an
If a simulation, however, draws on multiple
exemplar memories to produce a simulation that is an
average of them, then it functions more as a prototype.
Similarly, various types of rules can be implemented
when relational structures construe particular regions
of simulations as required for category membership.
Depending on the information simulated and how it is
interpreted, a wide variety of representational types
can in principle be implemented. Although computa
tional accounts of these different representational
types remain to be developed, this approach, in princi
ple, is capable of implementing them.
Using simulators and simulations to implement
basic conceptual functions. Once a collection of
simulators exists in long-term memory, it can imple
ment basic functions that are central to a conceptual
system: types versus tokens, categorical inference,
productivity, propositions, and abstract concepts. We
address each briefly in turn (for further detail, see
Barsalou, 1999, 2003b).
In the type–token distinction, type representations
stand for categories (e.g., politicians), whereas token
representations stand for individual category members
When a simulator represents an exemplar, it only produces a
partial simulation that is typically relatively sketchy and that may be
distorted by a variety of factors. We do not assume that complete
veridical records of perception are reproduced.
(e.g., Napoleon), and more specifically, for individuals
on particular occasions (e.g., Napoleon at Waterloo).
In PSS, simulators represent types because they ag
gregate modality-specific information across category
members. Conversely, simulations represent tokens,
namely, specific category members, along with spe
cific category members on particular occasions. Thus,
the simulator–simulation distinction in PSS naturally
implements the classic type–token distinction essential
for a conceptual system.
In categorical inference a category member is per
ceived, which activates the category representation.
In turn, knowledge likely to be true for the category
is extended to the category member. On seeing a par
ticular dog, for example, category knowledge about
dogs becomes active, which might then produce in
ferences about the dog being likely to bark, wag its
tail, and so forth. In PSS, such inferences arise as
simulations drawn from the modality-specific content
of a simulator. Once the perception of the dog acti
vates the dog simulator (via similarity between the
content of the perception and the content of the simu
lator), the simulator runs a simulation of likely per
ceptual content that has not yet been experienced. A
major issue for PSS (and for any theory of knowl-
edge) is how the correct inferences are generated. In
general though, PSS can produce categorical infer-
ences via the simulation process, simulating likely
modality-specific content that has not yet been expe-
rienced for the perceived individual.
In productivity, concepts are combined systemati-
cally to construct complex conceptual representations
(e.g., productively combining striped and purple with
waterfall and river yields striped waterfall, purple wa
terfall, striped river, and purple river). Notably, the
conceptual combination process is capable of repre
senting an infinite number of concepts whose referents
have never been experienced, such as purple waterfall.
Because PSS establishes simulators for individual
components of experience, it has the necessary build
ing blocks for implementing productivity. Once simu
lators for striped, purple, waterfall, and river exist, they
can be combined to form more complex concepts. For
example, a person could run the waterfall simulator to
produce a particular waterfall simulation. Once this
simulation is in place, then the color of the waterfall
can be systematically varied, using simulators for
color, such as purple, orange, and gold, to differen
tially simulate the waterfall’s color.
Central to productivity is having relational knowl
edge about how various components combine to form
more complex structures. Just as PSS establishes simu
lators for objects and properties, however, it also estab
lishes simulators for a wide variety of relations, such as
the aforementioned. To productively simulate differ
ent above relations, PSS first uses the simulator to con
struct a configuration of two regions that were previ
ously established as one instance of this relation (e.g.,
two spherical regions equal in horizontal position, dif
ferent in vertical position, nearly touching each other).
Once this relational simulation is in place, its regions
can be systemically filled with simulations of different
objects, such as birds, planes, barns, and trees, to repre
sent above relations such as above (bird, barn), above
(bird, tree), and so forth. Thus, by combining simula
tions hierarchically in simulated relational structures,
productivity results.
The most basic form of propositional interpretation
results from applying the type–token distinction to the
process of categorization. Essentially, the categoriza
tion process binds a type for a category to one of its to
kens, thereby establishing a type–token proposition.
On seeing a particular overhead projector, for exam
ple, categorizing it as an overhead projector binds cate
gory knowledge about overhead projectors to the ob
ject. This binding represents the proposition that this
particular object is a member of this particular cate
gory. Note that the proposition could be false. The ob
ject could actually be a piece of abstract art that some
one has mistakenly categorized as an overhead
projector. In this case, the proposition that represents
the agent’s belief is false, not true.
Most notably, the type–token proposition estab-
lished constitutes one possible interpretation of the ob-
ject. To see this, consider the infinite number of inter-
pretations of an actual overhead projector that can be
implemented with type–tokenpropositions. The projec-
tor could be interpreted correctly as an overhead projec-
tor, as an office tool, as an artifact, as an increasingly
dated piece of technology, and so forth. Similarly, the
projector could be interpreted incorrectly as a piece of
art, as a mammal, or as a space alien. In each case, differ
ent categorical knowledge is applied to the same object
to create a different type–token proposition. In each
case, a different interpretation results, accompanied by
a different family of categorical inferences.
It has been widely argued that modality-specific ap
proaches to knowledge cannot implement proposi
tions, because these approaches implement holistic im
ages that record experience, rather than implementing
concepts that interpret it (e.g., Pylyshyn, 1973). In
PSS, however, the binding of simulators to perceived
or simulated individuals naturally implements propo
sitions. Because PSS implements the type–token dis
tinction using simulators and simulations, type–token
propositions are a natural consequence.
Finally, it is often argued that modality-specific ap
proaches fail because they cannot represent abstract
concepts such as truth. Furthermore, it is often as
sumed that these approaches fail because representa
tions of the external world cannot be used to represent
abstract ideas. As we have seen, however, PSS estab
lishes simulators, not only for components of the ex
ternal world, but also for components of introspection,
including emotions, motivational states, cognitive op
erations, and so forth. Barsalou (1999) proposes that
abstract concepts are abstract because they focus
heavily on introspections and complex situational
events. In contrast, concrete concepts are concrete be
cause they focus on physical entities, settings, and sim
ple behaviors in the external world. Because simula
tors can be established for introspections and events
(not just for concrete objects), they can in principle
represent the conceptual content of abstract concepts
(not just the content of concrete concepts).
To assess this hypothesis, Barsalou and Wiemer-
Hastings (in press) used the property listing task to as
sess the content of abstract concepts (truth, freedom,
invention) and of concrete concepts (car, sofa, bird).
After participants listed the properties of these con
cepts, detailed coding schemes were applied to assess
the content produced. Most notably, the general types
of content for abstract and concrete concepts were
highly similar. For all concepts, participants tended to
describe situations that included objects, people, set
tings, behaviors, events, mental states, and relations.
For both types of concepts, participants situated their
conceptualizations of them, not just representing the
focal category content, but also representing extensive
background situational content relevant to understand-
ing and using the category.
The two types of concepts differed in their focus on
this content. Whereas concrete concepts focused on
entities, settings, and simple behaviors, abstract con-
cepts focused on introspections, social entities, and
complex events. Furthermore, the abstract concepts
were more complex, including greater relational struc-
tures, organized in greater hierarchical depth.
This exploratory study did not assess how partici
pants represented this content. In principle, though, it
seems possible that all of this content—for both con
crete and abstract concepts—could be simulated. Ev
erything that participants mentioned is something ex
perienced either in the external or internal world. As a
result, the process that establishes simulators could act
on this content and establish simulators for it, thereby
making it possible for a theory like PSS to explain it.
Notably, amodal theories face the same problem of
specifying the content of abstract concepts and of ex
plaining how their representations implement this con
tent, something that is far from having been accom
plished satisfactorily.
Situated Conceptualizations
As we just saw, people situate their representations
of categories. Thus, situated conceptualizations consti
tute another central construct in the PSS framework. A
situated conceptualization is one particular simulation
of a category in a background situation, where the spe
cific content of the category and situation prepare the
conceptualizer for action in it (Barsalou, 2002, 2003b).
In representing the category of chairs, for example, a
conceptualizer does not just simulate a generic repre
sentation of chairs in a vacuum. Instead, the concept
ualizer simulates one particular kind of chair in a par
ticular setting, along with actions and mental states
likely to occur while interacting with it. For example, if
the conceptualizer were on a jet, a chair simulation
would take the form of a jet chair, along with the rele
vant actions and mental states for interacting with it ef
fectively. Conversely, if the conceptualizer were in a
movie theater, a chair simulation would take the form
of a theater chair, along with the appropriate actions
and mental states. Thus, a situated conceptualization
simulates the focal category entity, along with simula
tions of the likely setting, actions, and introspections.
Because the simulation includes the conceptualizer’s
actions and introspections, the simulation creates the
experience of “being there” with the category member.
As a result, the conceptualizer is well prepared to inter
act with the entity in the anticipated situation.
Barsalou et al. (2003) propose that the construct
of situated conceptualization is useful in explaining
the wide variety of embodiment effects reported in
social psychology. Specifically, they propose that sit-
uated conceptualizations for repeated social situations
become entrenched in memory. Consider, for example,
the repeated situation of parents dealing with an upset
child. Across these repeated situations, a situated con-
ceptualization becomes entrenched that represents
how the child typically appears, how the parents feel,
what the parents do, and so forth. Because these sit-
uations contain many embodiments (e.g., perceptual
simulations of events, bodily states associated with
emotion and action), these embodiments become rep
resented in the situated conceptualization. Later, these
embodiments, when experienced, can trigger the situ
ated conceptualization via the inference process of pat
tern completion. Specifically, the experienced embodi
ment activates a larger pattern that contains it, with
nonperceived aspects of the pattern constituting infer
ences about the situation. Conversely, if through lin
guistic conversation, the situated conceptualization be
comes active, it can, in turn, produce corresponding
embodiments via the same inference process.
The PSS framework, with its construct of mul
timodal, situated conceptualization, accounts for the
diverse collection of social embodiment effects re
ported in the literature. Consider priming effects on be
havior, as when exposure to words associated with the
elderly stereotype produces slower walking (Bargh et
al., 1996). On the PSS account, stereotypes are situated
multimodal conceptualizations of social categories. In
the case of the elderly stereotype, its content includes
embodiments of slow motor movements. Activating
these embodiments during conceptualization of the
stereotype influences action, via top-down processing,
in related modality-specific systems, as when walking
slowly toward the elevator. In another example, con
sider how bodily movements influence conceptual
processing, as when head nodding during a persuasive
message leads to more positive attitudes (Wells &
Petty, 1980). On the PSS account, understanding a
message produces a multimodal situated conceptual
ization to represent the messages’ meaning. Nodding
while representing the message’s meaning activates
multimodal conceptualizations for prior situations in
which positive affect occurred. Once positive affect
becomes active, it influences both how the message is
represented and how it is evaluated, producing greater
positivity than if no action or a negative action were
performed instead.
Shallow Versus Deep Processing
Another important construct in the PSS framework
is the distinction between shallow versus deep process
ing. PSS does not require that all cognitive tasks utilize
simulation. Following Paivio (1986) and Glaser
(1992), PSS assumes that people can also use word-
level representations to perform superficial “concep-
tual” processing when task conditions permit. Con-
versely, when task conditions block superficial word
strategies and force people to perform conceptual pro-
cessing, simulations come into play. Not surprisingly,
participants adapt flexibly to task conditions. When
they can use shallow processing, they do, but when
they cannot, they perform deep processing.
To see how participants adapt flexibly to task condi-
tions, consider an experiment by Solomon and Bar-
salou (2004; also see Niedenthal et al., 2002). On each
trial, participants read the word for a concept (e.g.,
“CHAIR”), and then verified whether a subsequently
presented property (e.g., “seat”) was true of the con
cept (where the correct response on true trials, such
as this one, was “yes”). The key manipulation was
whether the properties tested on false trials (where the
correct response was “no”) were completely unrelated
to the target concept (e.g., CHAIR–feathers), or were
associatively related to it (e.g., CHAIR–table; note that
a true property had to be a part of the target concept).
Whereas some participants only received unrelated
false properties, others received only related false
properties, with both groups verifying the same true
Solomon and Barsalou (2004) predicted that the re
latedness of the false properties would determine
whether participants used a shallow or deep strategy
for verifying the true properties. When the false prop
erties were completely unrelated (e.g., CHAIR–feath
ers), participants could employ a shallow processing
strategy. Because all the true properties were related to
their respective concepts, whereas all the false proper
ties were unrelated, simply detecting an association be
tween the concept and property words was adequate
for determining the correct response. Whenever an as
sociation was detected, participants could correctly re
spond true; whenever they did not detect one, they
could correctly respond false. Participants did not need
to access conceptual knowledge about the concept and
the property to assess whether the property actually
belonged to the concept. Consistent with this predic
tion, Solomon and Barsalou (2004) found that, under
these task conditions, the associative strength between
concept and property words best predicted verification
times and errors.
When, however, the false properties were always
related to their concepts, participants could not rely on
superficial associations between the concept and prop
erty words, given that the false trials possessed them
too, not just the true trials. As a result, participants had
to access conceptual knowledge to verify that the con
cepts indeed possessed the properties. Solomon and
Barsalou (2004) predicted that if simulators represent
conceptual knowledge, then perceptual variables
should become the best predictors of verification time
and errors under these task conditions—not the asso
ciative strength between concept and property words.
In support of this prediction, perceptual variables, such
as property size, became the best predictors of perfor-
mance. When participants were forced to abandon the
superficial word association strategy, the conceptual
knowledge they used appeared to take the form of per-
ceptual simulation.
To corroborate this conclusion, Kan et al. (in press)
ran the same experiment in an fMRI scanner. When
participants received only related false properties, vi-
sual processing areas in the fusiform gyrus became
active to simulate the properties conceptually. When
participants received only unrelated false properties,
however, these areas were no longer active, supporting
the conclusion that participants were using the superfi
cial word association strategy instead, given that task
conditions allowed. As all these experiments illustrate,
conceptual processing is flexible, and participants
need not always use simulation. When conditions al
low, participants adopt more superficial strategies.
Support for the Simulation Hypothesis
Increasing evidence supports the central assump
tion of PSS that simulation underlies conceptual pro
cessing. Here we focus on several lines of supporting
evidence reported by Barsalou and his colleagues, al
though much evidence has been accumulating in other
laboratories as well (for reviews, see Barsalou, 2003b;
Hegerty, 2004; Martin, 2001). In particular, we focus
on three types of evidence for simulation: modality
switching costs, instructional equivalence, and percep
tual effort.
Modality Switching Costs
PSS predicts that switching costs should occur
when people verify properties on different modalities.
Pecher, Zeelenberg, and Barsalou (2003, 2004) tested
this prediction in a series of studies using the property
verification task (also see Marques, 2004). Participants
were first asked to verify a property of a concept that
requires simulation in one modality, such as BLEND
ER–loud (which requires an auditory simulation).
Next, participants were asked to verify a property of a
second concept, which either required simulation in
the same modality, such as LEAVES–rustling (again
requiring an auditory simulation), or in different mo
dality, such as LEMON–sour (requiring a gustatory
As PSS predicts, switching costs occurred. When
the modality of the property changed from the first trial
to the second, participants were slower to verify the
second property than when the modality stayed the
same. This finding suggests that when people represent
properties, they simulate them in the respective modal
ity-specific systems. Amodal theories of knowledge do
not predict switching costs a priori. Instead, these theo-
ries assume that properties are represented amodally
and can be verified without activating modality-spe-
cific systems. For a review of modality switching ef-
fects and for further discussion of their theoretical in-
terpretation, see Barsalou, Pecher, Zeelenberg,
Simmons, and Hamann (in press).
Instructional Equivalence
PSS predicts that people, by default, engage in per
ceptual simulation. One test of this prediction involves
examining instructional equivalence, or whether par
ticipants perform similarly under instruction condi
tions that emphasize or do not emphasize simulation
(Barsalou, Solomon, & Wu, 1999). Recent research
has recently used the property verification task to as
sess instructional equivalence (Krauth-Gruber, Ric, &
Niedenthal, 2004; Wu & Barsalou, 2004). Across stud
ies, participants in the imagery condition were explic
itly asked to image a referent of a concept before listing
its properties (e.g., construct an image of a chair and
then describe properties in the image). Much previous
work indicates that imagery instructions typically in
duce participants to construct images, which are typi
cally heavily visual. In contrast, participants in the
concept condition received no special instructions and
were simply asked to produce properties that are typi
cally true of the concept. The primary prediction from
PSS is that concept participants will spontaneously, by
default, construct simulations, much like those con
structed by imagery participants, although perhaps less
vivid and less conscious. Finally, participants in the
word association condition were asked to generate
words associated with the concept name. In each in
structional condition, participants’ protocols were
coded using a detailed coding scheme that established
a profile of the conceptual content produced.
PSS predicts that participants in the imagery and
concept conditions should produce similar profiles of
properties. If concept participants produce simulations
by default, the representations that they use to produce
properties should be highly similar to those that imag
ery participants produce. Furthermore, the profiles in
the concept and imagery conditions should differ from
those in the word association condition. Because word
association participants rely primarily on the associa
tive lexical system, they minimize use of the concep
tual system, adopting a shallow processing strategy
that produces the requested responses.
Amodal theories most naturally make a different a
priori prediction (e.g., feature list and semantic net
work models). According to these theories, concept
participants should, by default, use amodal represen
tations, not simulations. As a result, their property pro
files should differ significantly from the property pro
files of imagery participants. Furthermore, some
amodal approaches, at least, predict that the profiles of
concept participants should be similar to those of word
association participants. Because amodal theories of-
ten assume that symbols for concepts bear a systematic
relation to the units of language, they lend themselves
to the prediction that the default strategy of concept
participants is to activate symbols that correspond to
the words produced in the word association condition.
Findings from Wu and Barsalou (2004) and Krauth-
Gruber et al. (2004) support the PSS predictions. Across
multiple experiments, the property profiles for the con
cept and imagery conditions were more highly corre
lated with each other than with the profiles for the word
association condition. Furthermore, the correlations
between the concept and imagery condition were high,
suggesting that concept participants adopted the same
representational strategy used in the imagery condition
by default (i.e., simulation).
Perceptual Effort
Perceptual effort constitutes another indicator of
simulation by default (Barsalou et al., 1999). If people
spontaneously use simulations to represent concepts,
then these representations should have perceptual
qualities. As a result, manipulating perceptual vari
ables should affect the ease of conceptual processing,
much like manipulating these variables affects the ease
of processing of mental images in working memory
(e.g., Finke, 1989; Kosslyn, 1980; Shepard & Cooper,
1982). Manipulating perceptual variables such as size,
orientation, and occlusion should influence the ease of
processing concepts and their properties. As simula
tions of a property become increasingly large, for ex
ample, greater time and effort are needed to construct
it. Similarly, the more an object simulation must be ro
tated to achieve an upright position, the greater the
time and effort needed to complete this operation.
Wu and Barsalou (2004) reasoned that if people
are representing object properties with simulations, the
perceptual variable of occlusion should affect the ease
generating properties in the property generation task.
To see this, imagine participants being asked to pro
duce the properties of, say, watermelons. If perceptual
effort affects this task, then participants should pro
duce the properties that require the least effort to per
ceive in their simulations, namely, those that are
unoccluded. Producing occluded properties should re
quire more perceptual effort, because simulation must
be transformed to reveal them. Thus, for watermelons,
participants should produce outer unoccluded proper
ties, such as green and stripes more often than occluded
properties such as red and seeds. Across several exper
iments, Wu and Barsalou found evidence for this pre
diction. Unoccluded properties were produced more
often than occluded properties, and also earlier and in
larger clusters.
Wu and Barsalou (2004) also tested a second pre-
diction. If the normally occluded properties of an ob-
ject become unoccluded, they should require less per-
ceptual effort to produce and therefore be produced
more often. Thus, in some conditions of these experi-
ments, participants produced properties of the same
objects but where the name of each object included a
revealing modifier. Rather than producing the proper-
ties of watermelons, for example, these participants
were asked to produce the properties of a half water-
melon. Because the normally occluded inner parts of a
watermelon become unoccluded in a half watermelon,
less effort is required to perceive them in a simulation.
Thus, normally occluded properties should increase in
production rates, becoming more comparable to nor
mally unoccluded properties. Across several studies,
Wu and Barsalou (2004) observed this pattern, consis
tent with the prediction that participants were using
simulations to represent the concepts.
Krauth et al. (in preparation) found that producing
the properties of an emotional state produced a similar
occlusion effect. When participants were asked to de
scribe properties of “his anger,” they produced oc
cluded properties of the other’s mental state less often
than when they produced properties of “my anger,”
when these mental state properties are unoccluded.
Again, the amount of effort required to simulate prop
erties determined their rates of production.
To further explore perceptual effort, Solomon and
Barsalou (2004) again employed the property verifi
cation task (also see Solomon & Barsalou, 2001). If
people simulate properties to verify them in their re
spective objects (e.g., PONY–mane), then perceptual
variables should explain the time to verify properties,
and also the accompanying error rates. To assess this
issue, Solomon and Barsalou regressed verification
RTs and error rates onto potential factors that could
predict performance, including linguistic, perceptual,
and expectancy variables. Under conditions that re
quired conceptual processing (i.e., related false prop
erties as described earlier), the perceptual variables
predicted performance better than the linguistic and
expectancy variables. In particular, the size of proper
ties was important in predicting performance. As prop
erties became larger, they took increasingly longer to
verify and led to higher rates, presumably because
large properties take more time and effort to represent
than small ones. Under conditions that did not require
conceptual processing (i.e., unrelated false properties),
the linguistic variables best predicted performance,
suggesting that participants used the superficial word
association strategy instead of deeper conceptual pro
cessing, as described earlier.
Related Views
In this article, we propose that the PSS account is
productive for understanding embodiment phenomena
in social psychology. Previous theories in social psy-
chology, however, have made related proposals. Thus,
it is useful to compare and contrast PSS with these re-
lated accounts. The two accounts most relevant are the
Hard Interface Theory (Zajonc & Markus, 1984) and
the Associated-Systems Theory (Carlston, 1994).
Hard Interface Theory (HIT)
In 1984, Zajonc and Markus published an influen
tial chapter titled “Affect and cognition: The hard in
terface.” In this chapter, Zajonc and Markus encour
aged social psychologists to pay more attention to
embodiment, and in this spirit they introduced the HIT.
Although, Zajonc and Markus focused primarily on the
interaction between emotion and cognition, they also
made general proposals about the representation of so
cial knowledge, which can be viewed as precursors to
more recent developments. In particular, Zajonc and
Markus (1984) declared that purely propositional,
“disembodied” theories of social knowledge are unsat
isfactory for several reasons.
For one, propositional
theories rarely address how abstract representations
are connected to actual behavior. Thus, in the emotion
domain, there is no clear account of how cognition
(e.g., “hearing an insult”) elicits bodily components of
anger (clenched fists, red faces, bulging veins, etc). A
The term propositional is used widely throughout cognitive and
social psychology when referring to amodal theories of knowledge,
even though, as we have seen, embodied theories can also be propo
sitional (Barsalou, 1999, 2003a).
related problem arises when one tries to understand the
influence of emotion on cognition. In propositional
theories, the problem is reduced to the influence of one
associative structure on another by assuming that the
bodily (muscular, hormonal) components of emotion
somehow get transduced into a propositional form.
Finally, propositional theories have a hard time ac
counting for embodiment phenomena in cognition.
Why, for example, Zajonc and Markus asked, “do peo
ple engaged in an arithmetic problem often gnash their
teeth, bite their pencils, scratch their heads, furrow
their brows or lick their lips?” (p. 74), “Why do people
scratch their heads and rub their chins when they try to
remember something?” (p. 84).
Zajonc and Markus’s (1984) response is that bodily
states constitute “hard representations.” According to
this view, bodily states have representational content.
When a dog hears a bell, for example, and withdraws a
leg to avoid shock, this leg flexion can be said to repre
sent the dog’s knowledge of the bell–shock relation.
Importantly, HIT proposes that the bodily movement
itself has representational content and does not require
a more cognitive “soft representation” to have repre
sentational value. A purely hard representation, such
as the dog’s leg flexion, can guide behavior in a dan-
gerous situation as effectively as an abstract soft repre-
sentation of the event. Individual hard representations
can also interact with one another, as when, for exam-
ple, muscles associated with a forward movement
(hard representation of positivity) interfere with mus-
cles associated with a backward movement (hard rep-
resentation of negativity).
Probably the best known empirical illustration of
HIT is the “chewing gum” experiment (Zajonc et al.,
1982; cf. Graziano, Smith, Tassinary, Sun, & Pilking
ton, 1996). Participants were first asked to study 78
photographs of faces in four between-participant con
ditions. One group imitated the targets’ head and gaze
orientations and their facial expressions (instructed
mimicry). A second group chewed gum (blocked mim
icry). A third group squeezed a sponge with their
nonpreferred hand (motor control). A fourth group
judged the head orientations and facial expressions in
the photographs (judgment control). After studying the
pictures, participants received a recognition test. As
the embodiment view predicts, memory was best when
participants’ embodiments were compatible with the
pictures—the instructed mimicry group scored highest
(73% correct). The worst performance occurred for
participants who performed the most competitive mo
tor response—chewing gum (59%). Participants who
squeezed a sponge fell in between (65%) as did partici
pants who judged the faces (64%). This pattern of re
sults suggests that participants represented properties
of the perceived faces in their musculature, such that
when their musculature was consistent with a face, it
enhanced memory via consistent elaboration.
Another important aspect of HIT is that it provides
intriguing accounts of classic psychological phenom
ena that differ considerably from standard accounts.
For example, Zajonc and Markus (1984) proposed that
people’s particularly good memory for faces reflects
their ability to imitate perceived faces and to cre
ate hard muscular representations that complement
soft representations. In a similar fashion, Zajonc and
Markus proposed that mood congruence effects occur
because affective states create bodily configurations
that facilitate the production of compatible responses.
The state of happiness, for example, creates a specific
posture and muscle tone, which facilitates “happy”
muscular responses to objects. In yet another example,
Zajonc and Markus suggested that the mere-exposure
effect results from the gradual relaxation of bodily re
sponses to a stimulus as a result of habituation, thereby
making positive evaluation more compatible with the
bodily state.
Despite the shared affinity for embodiment, differ
ences exist between HIT and PSS. Perhaps most sig
nificantly, HIT assigns representational functions to
actual bodily states. This includes not only muscles,
but also gastrointestinal, glandular, and cardiovascu-
lar systems. In the original PSS account (Barsalou,
1999), the critical mechanism is simulation in the
brain, where the body plays a role only as represented
in neural systems. A related difference is that HIT al-
lows representational functions of bodily states to
work “by themselves,” without translation into “soft”
representations. In contrast, the original PSS account
required that bodily states are first encoded into neu-
ral representations in modality-specific systems,
which then allow bodily states to interact with cogni
tion. In subsequent developments of PSS, however,
bodily states have been central to the situated concep
tualizations that underlie higher cognition—not just
simulations of these states (e.g., Barsalou, 2003b;
Barsalou et al., 2003).
Another difference in relative emphasis between
the two theories concerns associative versus dynamic
processing. HIT is fundamentally associative, estab
lishing relatively stable associations between particu
lar cognitions and their accompanying embodiments.
PSS contains many associative mechanisms as well,
but focuses more on the dynamic construction of situa
tional conceptualizations and on the productive con
struction of simulations into complex relational struc
tures that function propositionally. Analogously, HIT
has focused more on explaining social phenomena,
whereas PSS has focused more on explaining a variety
of higher cognitive functions, such as type–token pro
cessing in categorization and the productive construc
tion of complex simulations in language comprehen
Associated-Systems Theory (AST)
AST provides a multimodal model of how the cog
nitive system represents other people in person percep
tion (Carlston, 1994). According to AST, representa
tions of other people develop through the use of four
primary mental systems: (a) the visual system, (b) the
verbal/semantic system, (c) the affective system, and
(d) the action system. Each system captures the rele
vant features of external stimuli and produces repre
sentations that are specific to it. Thus, the visual sys
tem represents a target person’s appearance, the action
system represents an action sequence directed at the
target person, the affective system represents affect to
the target, and the verbal system represents the target’s
personality traits.
Each system in AST is hierarchically organized.
Whereas the lowest levels contain highly specialized
structures used to perceive stimuli and produce re
sponses, the highest levels contain abstract concepts
related associatively to perceptual or response pro
cesses. At the lowest levels, the four primary systems
are independent of each other. At higher levels, how
ever, they become progressively intertwined, interact-
ing to produce hybrid forms of representation. When
categorizing individuals into social groups, for exam-
ple, a perceiver accesses both an image of the group (in
the visual system) and descriptive labels for their traits
(from the verbal system). Similarly, when evaluating a
target person, a perceiver uses both the verbal and the
affective systems. Categorizations and evaluations, to-
gether with orientations (a hybrid of affect and action
system) and behavioral observations (a hybrid of vi-
sual and action systems), constitute the four products
of AST’s secondary systems.
According to AST, the representations that one es
tablishes for a target person typically utilize one distri
bution of the four primary systems more than other
possible distributions. For instance, passive observa
tion of a target person produces appearance representa
tions grounded mainly in the visual system. In contrast,
actually interacting with a person activates not only the
visual system, but also the action system, thereby pro
ducing a different distribution. In a related manner, re
sponses to a target person depend on the distribution of
system(s) solicited in the response situation.
Furthermore, because AST’s systems can function
independently of each other, the representations in dif
ferent systems can potentially conflict when processing
a given individual. When such contradictions occur,
judgments of the target depend on which of the conflict
ing systems dominates the judgment situation. If the vi
sual system dominates (e.g., because the target is pre
sented via a picture), responses to the target will be
based on appearance, even when contradictory judg
ments arise in other systems. If, however, the verbal sys
tem dominates (e.g., because the target is presented via
text description), responses to the target will be based on
personality traits. Todemonstrate the relative independ
ence of the four systems and the potential for one to
dominate others, Claypool and Carlston (2002) showed
that interference in the visual system at encoding de
creased the influence of visual information in liking
judgment but did not alter the influence of verbal infor
mation. Conversely, Claypool and Carlston found that
interference in the verbal system decreased the influ
ence of verbal information but not the influence of ap
pearance information. Depending on the information
presented and the availability of relevant systems, dif
ferent profiles of processing for a target ensued.
AST shares many assumptions with PSS, in particu
lar, the central assumption that multiple modality spe
cific systems underlie cognitive processing. Both theo
ries also view CZ theory as an appropriate description
of knowledge representation in the brain (Damasio,
1989). Finally, AST and PSS both assume that pro
cessing is dynamic. Depending on current task condi
tions, different representations and different forms of
processing control task performance.
AST and PSS differ, however, on other points. First,
AST gives primacy to the abstract verbal system, with
its representational elements of words and personality
traits. According to PSS, however, even abstract con-
cepts like personality traits are grounded in modal-
ity-specific systems, as are the representation of words
in the respective modalities (e.g., orthographic repre-
sentations of the word in the visual system and acoustic
representations of words in the auditory system; Bars-
alou, 1999; Simmons & Barsalou, 2003). Thus, PSS is
somewhat more radical than AST in its reliance on the
modalities as the basis of representation, whatever its
degree of abstractness.
Second, AST, as its name implies, attempts to ex
plain the role of modal information in terms of associa
tive processes. Accordingly, the previously mentioned
limitations of HIT also apply to AST. In particular, it is
not clear how AST accounts for classic conceptual
functions such as the type–token distinction, produc
tivity, propositional interpretation, and the representa
tion of abstract concepts.
Criticisms of Embodiment Theories
Theories of embodied cognition have appeared reg
ularly in the cognitive and social literatures. So far,
however, these theories have not become a major force
in explaining cognition, in general, and social cogni
tion, in particular. The major reason is previous ac
counts have had trouble dealing with certain classes of
criticism. In the preceding sections, we have already
mentioned how PSS goes beyond earlier proposals in
social psychology. Here we list criticisms that have
been raised against embodiment theories and show
how PSS addresses them.
Selective Embodiment
Embodiment theories need to solve the problem of
how some cognitive activity can proceed without in
volvement of bodily states and modality-specific sim
ulations. PSS addresses this issue with the distinction
between shallow versus deep processing. A perceiver
simulates primarily when needed. When conceptual
tasks can be solved using shallow strategies, such as
word association, simulations of conceptual content
are not recruited or play only peripheral roles.
Dynamic Use of Embodiment
A related challenge to embodiment theories arises
from findings suggesting that perceivers are quite flexi
ble in their use of embodied information. For example,
when empathizing with others, people typically rely on
their own modality-specific reactions, suggesting use of
a simulation strategy (Decéty & Chaminade, 2003;
Gallese, 2003). However, the exact nature of the stimu-
lation depends on task goals. For example, in some
cases, people empathize by simulating emotional states,
whereas in other cases people empathize by simulating
cognitive states. In some situations, people may decide
not to simulate at all (e.g., when dealing with a mass
murderer). More generally, according to modern em-
bodiment theories such as the PSS, whether or not peo-
ple decide to simulate, what specific modality people
decide to simulate in, and how they are going to use the
products of the simulation is open to task goals.
The notion of goal-driven stimulation allows mod
ern embodiment theories to account for social psy
chological findings showing people’s flexibility in us
ing experiential sources of information in judgment.
Thus, people typically draw on experiential feedback
(mood, feelings of processing difficulty, etc.) when
solving a variety of cognitive tasks, including fre
quency judgments, self-assessment, and value judg
ments (see Schwarz, Bless, Waenke, & Winkielman,
2003). When people realize that their experiential reac
tions are not diagnostic, however, they switch to alter
native sources of judgments, making less use of current
experience. Critics could interpret this finding as sug
gesting that people use simulation only under default
conditions, and that under conditions that produce
misattributions, people fall back on amodal process
ing. Instead, we propose that such findings reveal peo
ple’s flexibility in their use of the simulation strategy.
When an external factor compromises validity of a
simulation in one modality, people can switch to a sim
ulation in an alternative modality.
Representational Limitations
of the Body
Embodiment theories must address the problem of
the body’s representational capacity. Starting with
Cannon (1927, 1929), critics have argued that bodily
feedback is too undifferentiated and too slow to serve
as the basis of experience. Furthermore, there is the
problem that the same bodily state may be associated
with different cognitive and emotional representations
(Zajonc & McIntosh, 1992). These issues are actually
quite old and have been effectively used by Cannon
and many others to argue against the James–Lange the
ory of emotion (James, 1896/1994).
Several responses to this criticism are possible.
Zajonc and Markus (1984) note that the motor system
can support extremely subtle distinctions, as the so
phistication of sensory-motor processing in spoken
language illustrates. Further, even a limited number of
bodily states can support a very large number of repre
sentational distinctions. (Consider how many melodies
can be played with 88 piano keys—a number much
lower than the number of muscles in the body.)
More important, recent embodiment approaches,
such as PSS, CZ theory, and Somatic Marker Theory
(Damasio, 1999), avoid the “body-is-too-crude-too-
slow-and-too-varied” criticisms by focusing on the
brain’s modality-specific systems, instead of on actual
muscles and viscera. The circuits in modality-specific
brain areas are as fast and refined as any other form of
cortical representation and are thus able to flexibly
process a large number of modal states at the same time
(Damasio, 2003).
Higher Cognitive Functions
Embodiment theories must account for the basic
functions of higher cognition, such as the type–token
distinction, categorical inference, productivity, propo
sitional interpretation, and abstract concepts. As dis
cussed earlier, PSS uses the constructs of simulators
and simulations to explain these phenomena. Because
simulators develop for components of experience, cat
egorical knowledge results that represents types, pro
duces categorical inference, and combines produc
tively. Furthermore, simulators become bound to both
perceived and simulated individuals to implement
type–token propositions. Because PSS uses selective
attention to break experience into components and
then establish categorical knowledge about them, it de
velops the classic capabilities of a conceptual system.
Perhaps the most common criticism of theories of
embodied cognition concerns their ability to represent
abstract concepts. An implicit assumption of these crit
icisms, however, is that only simulations of external
experience can be used for representational purposes
in embodiment theories. As we have seen throughout
this article, however, simulations of introspective ex
perience are also available for representational pur
poses. Furthermore, as we saw earlier, simulations of
introspections are central to the representation of ab
stract concepts (Barsalou & Wiemer-Hastings, in
press). More generally, the simulation of situations ap
pears to provide a natural approach to representing a
wide variety of concepts, including abstract ones.
Regressing to Behaviorism
or Mere Associationism?
Closely related to the issue of higher cognitive func
tions is the worry that embodiment theories advocate a
return to the view that simple associations between
sensory and motor processes underlie all cognition.
Long ago, behaviorist theories were conclusively
shown to be incapable of explaining cognitive achieve
ments such as memory, language, and thought (Chom
sky, 1959). We endorse these conclusions strongly and
agree that simple associations between sensory–motor
states, drawn only from experience, cannot explain
cognitive processes. What we propose here shows how
the sophisticated symbolic processing characteristic of
human cognition is grounded in embodied representa-
tions. An associationist account is as inadequate now
as it was 50 years ago.
Are Observed Mind-Body
Relationships Epiphenomenal?
Embodiment theories make clear claims of causal-
ity (e.g., “modality-specific systems areas are critical
for conceptual processing”). Nevertheless, supporting
evidence has sometimes been viewed as only correla
tional in nature, consisting mainly of observations
of co-occurrence between embodiment and cognition.
Thus, critics have argued that perceptual representa
tions are not constitutive of concepts but only become
active epiphenomenally during the processing of
amodal symbols, which themselves do the work of
high-order cognitive processing. As reviewed through
out this article, however, theories of embodied cogni
tion can now draw on many experimental findings that
show that manipulations of embodiment causally mod
ulate cognitive and social performance. Demonstrating
that the inhibition or facilitation of a specific