Social cognition is the
study of the acquisition, representation, and use of social
knowledge -- in general terms, it is the study of social
A comprehensive theory of social cognition must contain several elements (Hastie & Carlston, 1980; Kihlstrom & Hastie, 1987):
The first two of these facets have to do with perception and attention -- the processes by which knowledge is acquired.
Or are they? Actually, a prior question is: Where does social knowledge come from?
When asked about knowledge in general, psychology and cognitive science offer two broad answers -- and a third, correct answer.
And so it
is with social cognition: it, too, begins with social
We can begin by distinguishing between sensation and perception.
has to do with the detection of stimuli in the environment -- in
the world outside the mind, including the world beneath the
skin. To make a long story short: Sensory processes:
For example, the rods and cones in the retina of the eye convert light waves emitted by an object into neural impulses that flow over the optic nerve to the occipital lobe.
gives us knowledge of the sources of our sensations -- of the
objects in the environment, and of the states of our bodies:
What objects are in the world outside the mind, where they are
located, what they are doing, and what we can do with
them. Put another way, perception assigns meaning to
respects, sensation is not an intelligent act.
By contrast, perception is the quintessential act of the intelligent mind. Perception goes beyond the mere pickup of sensory information, and involves the creation of a mental representation of the object or event that gives rise to sensory experience. In order to form these mental representations, the perceiver (in the lovely phrase of Jerome Bruner, a pioneering cognitive psychologist) "goes beyond the information given" by the stimulus, combining information extracted from the current stimulus with pre-existing knowledge stored in memory, employing processes of judgment and inference.
In cognitive psychology, there are basically two views of perception.
By any standard, the constructivist view dominates research and theory on perception. But, as we will see, the ecological view also finds its proponents.
The study of social perception begins with an analogy between social and nonsocial objects. The study of social perception assumes that any person is an object who has an existence independent of the mind of the perceiver. Accordingly, the perceiver's job is to extract information from the stimulus array to form an internal, mental representation of the external object of regard.
The term person perception was introduced by Bruner and Tagiuri (Handbook of Social Psychology, 1954) to reflect the status of persons as objects of knowledge. As with any other aspect of perception, they argued that a number of factors influence perceptual organization:
Jerome Bruner and the "New Look" in Perception
Bruner was a pioneering cognitive psychologist and cognitive scientist. Among his notable accomplishments was the introduction of what he called a "New Look" in perception, which sought to redirect perception research from an analysis of stimulus features to an analysis of the perceiver's internal mental states. Although perception is obviously a field of cognitive psychology, to some extent Bruner's New Look was influenced by psychoanalysis, as when he argued that emotional and motivational processes interacted with cognitive processes -- so that, in some sense, our feelings and desires affect what we saw.
Our impressions of other people are typically represented linguistically, often as trait adjectives. Consider a survey by the Washington Post, which asked respondents to describe in three words the various candidates for the Democratic and Republican presidential nominations. The most frequent responses, arranged as "tag clouds" in which the font size represents the frequency with which the word was used, looked like these:
similar survey, conducted over Facebook by the Daily Beast, a journalism
website (with, perhaps, a somewhat liberal bent), following the February 2012 Republican
presidential debates. During the debate, CNN
correspondent John King had asked each of the candidates to
describe themselves in one word. The Daily
Beast polled subscribers to its Facebook page with
the same question, resulting in the following word
For good measure, they also asked their Facebook
subscribers to describe Barack Obama, who was
unopposed for the Democratic nomination.
In 2015, in the run-up to the 2016 election, YouGov.com, a global online community that promotes citizen participation in government, conducted a similar survey in which visitors to the organization's website were asked to characterize some of the leading presidential candidates in one word (at the time, there was some indication that Mitt Romney would join the race). Separate word clouds were constructed from the responses of people who liked and disliked each candidate.
Often, the stimulus information for person perception also comes in verbal form, as a list of traits and other descriptors. This is certainly the case with the self-descriptions that appear in "personals" ads in newspapers and magazines. But it is also true when we describe other people. Consider this passage from the Autobiography of Mark Twain, the author describes the countess who owned the villa in Florence where he and his family stayed in 1904:
"excitable, malicious, malignant, vengeful, unforgiving, selfish, stingy, avaricious, coarse, vulgar, profane, obscene, a furious blusterer on the outside and at heart a coward."
Or, this description of Osama Bin Laden, which the former chief of the CIA's "Bin Laden Issues Station" has endorsed as a "reasonable biographical sketch" of the man.
When Fiske and Cox (1979) coded peoples' open-ended descriptions of other people, they identified six major categories:
Because verbal lists of traits are easy to compose and present to subjects, many studies of person perception begin with traits, and proceed from there. This is reasonable, because so much of our social knowledge is encoded and transmitted via language.
Actually, the study of person perception began before 1954, with the work of Solomon Asch (1946). Like Lewin (about whom you've already heard), and Fritz Heider (about whom you'll hear a lot in the future), Asch was a German refugee from Hitler's Europe. And like them, he was heavily influenced by European Gestalt psychology. Much of Asch's early work was on aspects of nonsocial perception, but he brought the Gestalt perspective to bear on problems of social psychology in his classic textbook, Social Psychology (1952), which was the first social psychology text to be written with a unifying cognitive theme running throughout.
Asch (1946) set out the problem of social perception as follows:
[O]rdinarily our view of a person is highly unified. Experience confronts us with a host of actions in others, following each other in relatively unordered succession. In contrast to this unceasing movement and change in our observations we emerge with a product of considerable order and stability.
Although he possesses many tendencies, capacities, and interests, we form a view of one person, a view that embraces his entire being or as much of it as is accessible to us. We bring his many-sided, complex aspects into some definite relations....
How do we organize the various data of observation into a single, relatively unified impression?
How do our impressions change with time and further experiences with the person?
What effects in impressions do other psychological processes, such as needs, expectations, and established interpersonal relations, have?
addressing these questions, Asch set out two competing theories:
Obviously as a Gestalt psychologist, Asch had a pre-theoretical preference for the latter theory.
In order to study the process of person perception, Asch (1946) invented the impression-formation paradigm. He presented subjects with a trait ensemble, or a list of traits ostensibly describing a person (the target) -- varying the content of the ensemble, the order in which traits were listed, and other factors. The subjects were asked to study the trait ensemble, and then to report their impression of the target in free descriptions, adjective checklists, or rating scales.
Asch's first experiment compared the impressions engendered by two slightly different trait ensembles. Subjects were presented with one of two trait lists, which were identical except that target A was described as warm while target B was described as cold.
After studying the trait ensemble, the subjects reported their impressions in terms of a list of 18 traits, presented as bipolar pairs such as generous-ungenerous.
The two ensembles generated two quite different impressions, with A perceived in much more positive terms than B. There were significant differences between the two impressions on 10 of the 18 traits in subjects' response sets. A later experiment, varying only intelligent-unintelligent, yielded similar results.
But when the experiment was repeated (Experiment 3), with the words polite and blunt substituted for warm and cold, there were relatively few differences between the two impressions.
From these and related results, Asch concluded that traits like warm-cold and intelligent-unintelligent were central to impression formation, while traits like polite-blunt were not. In Asch's view, central traits are qualities that, when changed, affect the entire impression of the person. Other traits are more peripheral, in that they make little difference
Although being described as warm rather than cold led the target to be described in highly positive terms, Asch distinguished the effect of central traits from the halo effect described by Thurstone, by which targets described with one positive trait tend to be ascribed other positive traits as well. Being described as warm rather than cold does not lead to an undifferentiated positive impression; the warm-cold effect is more differentiated than that.
As proof, Asch pointed out that warm-cold is not always central to an impression. In his Experiment 2, where warm and cold were embedded in a different trait ensemble, there were few differences between the resulting impressions. If anything, the person was perceived as somewhat dependent, rather than the glowingly positive terms that emerged from Experiment 1.
Consistent with Gestalt views of perception, the effect of one piece of information (whether the person is warm or cold) depends on the entire field in which that information is embedded. To explain why traits are sometimes central and other times peripheral, Asch offered the change of meaning hypothesis, which holds that the total environmental surround changes the meaning of the individual elements that comprise it. Remember, for Gestalt psychologists, the distinction between figure and ground is blurry, because both figure and ground are integrated into a single unified perception. Perception of the figure affects perception of the background, and perception of the background affects perception of the figure.
In addition to studying the semantic relations among stimulus elements, Asch also studied their temporal relations. After all, he argued, impression-formation is extended over time: as we gradually accumulate knowledge about a person, our impression of that person may change. In his Experiment 6, Asch presented subjects with two identical trait ensembles, except that intelligent was the first trait listed for target A, and the last trait listed for target B. The two impressions differed markedly, revealing an order effect in impression formation.
In order to explain order effects, Asch held that the initial terms in the trait ensemble set up a "direction" that influences the interpretation of the later ones. The first term sets up a vague but directed impression, to which later characteristics are related, resulting in a stable view of the person -- just as our perception of a moving object remains stable, even though our perspective on it may change over time.
Taken together, Asch's studies illustrate principles of person perception that are familiar from the Gestalt view of perception in general. The whole percept is greater than the sum of its stimulus parts, because the elements interact with each other; just as the perception of the individual stimulus elements influences perception of the entire stimulus array, so the perception of the entire stimulus array influences the perception of the individual stimulus elements.
1946 experiments set the agenda for the next 20 to 30 years of
research on person perception and impression formation, which
basically sought clarification on questions originally posed by
Interestingly, however, a recent large-scale study failed to replicate one of Asch's findings: the "primacy of warmth" effect, by which warm-cold serves is not only a central trait, but more important to impression-formation than the other big central trait, intelligent-unintelligent. Nauts et al. (2014) carefully repeated Asch's (1946)procedures (for his Studies I, II, and IV), in a sample of 1140 subjects run online via Mechanical Turk.
So, just to be clear, Nauts et al. confirmed that arm-cold and intelligent-unintelligent are central to impressions of personality; but they failed to find, as he claimed, that warm-cold was more important than intelligent-unintelligent.
- Despite the impression given by their title, Nauts et al. actually replicated the primary findings of Asch's study.
- Content analysis of the subjects' open-ended descriptions of the targets (a quantitative procedure that was not available to Asch in 1946) showed that subjects used the terms warm, cold, and intelligent more often than any other trait term. This is evidence for the importance of these dimensions to impression-formation.
- In addition, subjects who saw warm (cold) in the trait ensemble described the targets in more (less) positive terms, compared to those who saw polite (blunt).
- So, how could Nauts et al. claim a failure to replicate Asch?
- It turned out that warm-cold was not used more often than intelligent-unintelligent in subjects impressions, contradicting the "primacy of warmth" hypothesis. In fact, intelligent-unintelligent appeared more often in the impressions than warm-cold.
- When subjects were asked to rank how important the various items in the trait ensemble were to their overall impressions, they ranked intelligent-unintelligent higher than warm-cold.
Asch's distinction between a central and a peripheral trait was made on a purely empirical basis: He discovered that some traits, such as warm-cold and intelligent-unintelligent, exerted a disproportionate effect on impressions of personality, while others, such as polite-blunt, did not. But although he could predict the effects of central traits on impressions, he had no theory that would enable him to predict which traits would be central, and which peripheral.
So what makes a trait central as opposed to peripheral? Julius Wishner (1960) offered a plausible answer. He administered a 53-item adjective checklist, derived from the checklists that Asch had used, asking his subjects to describe their acquaintances (often, their teacher in introductory psychology). Using the power of high-speed computers that simply were not available to Asch in 1946 (and which, frankly, are dwarfed by the computational power of the simplest laptop or even palmtop computer today), Wishner calculated the correlations between each trait and every other trait in the list. Examining the matrix of trait intercorrelations, Wishner observed that traits such as warm-cold and intelligent-unintelligent, which Asch had identified as central, had significant correlations with many other traits (e.g., mean rs = .62 and .56, respectively); by contrast, peripheral traits such as polite-blunt had relatively few correlations (e.g., mean r = .43).
The upshot of Wishner's study is that central traits carry more information than peripheral traits, in that they have more implications for unobserved features of the person. By virtue of their high intercorrelations with other traits, knowing that a person is warm or intelligent tells us a great deal about the person, while knowing that a person is polite does not. In the same way, a change in one central trait, from warm to cold or from intelligent to unintelligent, implies changes in many other traits as well, while a change from polite to blunt does not.
Wishner's findings also explained why a trait like warm-cold was not always central: it depends on the precise list of traits on which subjects make their ratings. Any trait from the stimulus ensemble will function as a central trait, so long as it is highly correlated with many of the traits on the response list.
Wishner's solution to the problem of central traits made his paper a classic in the person-perception literature, but it is not completely satisfactory. For example, it might be nice if "centrality" was a property of the trait itself, and did not depend on the context provided by the response set (though, frankly, Asch, as a Gestalt psychologist, might not think this was so desirable!). Are there any traits that are inherently central?
Rosenberg (1968), making use of even more computational power
than had been available to Wishner, factor-analyzed the
intercorrelations among a large number of trait terms, yielding
a hierarchical structure consisting of subordinate
traits, primary traits, secondary and even tertiary
traits. He discovered that Asch's central traits tended to
load highly on two very broad superordinate factors of
personality ratings representing two dimensions:
Interestingly, these two "superfactors" are not entirely independent of each other: people who tend to be described in positive social terms also tend to be described in positive intellectual terms. There is a "super-duper" factor of evaluation which runs through the entire matrix of personality traits, and gives rise to Thurstone's halo effect.
all of this material together, we can conclude that central
traits have two properties:
Talking with Strangers
Gladwell also discusses other, more prosaic examples of egregious misundestanding, which he explains with two basic principles, both drawn from soial-science research.
Gladwell probably makes too much of two relatively small principles (he made similar mistakes in Blink and The Tipping Point), but his book only underscores the importance of understanding both how we perceive other people, and how accurate, or inaccurate, those perceptions can be.
While a great deal of personality research has been devoted to determining the hierarchical structure of personality traits (e.g., the Big Five structure of neuroticism, extraversion, agreeableness, conscientiousness, and openness to experience), it seems likely that laypeople possess some intuitive knowledge of the structure of personality as well. In fact, Asch's concept of the central trait assumes that laypeople possess some intuitive knowledge about the relations among personality traits. If they did not, all traits would be created equal, none more central, or more peripheral, to impression formation than any other.
The term implicit personality theory (IPT) was coined by Bruner and Tagiuri (1954 -- they were busy that year) to refer to "the naive, implicit theories of personality that people work with when they form impressions of others". Bruner and Tagiuri understood that person perception entailed "going beyond the information given" by combining information extracted from the stimulus with information supplied by pre-existing knowledge. In other words, in the course of person perception the person must make use of knowledge that he or she possesses about the relations among various aspects of personality.
People's implicit theories of personality may be quite different from the formal theories of personality researchers. In fact, compared against formal theories resulting from methodologically rigorous research, they may even be wrong. But right or wrong, they are used in the course of person perception.
The domain of IPT was further explicated by L.J. Cronbach in his contribution to a 1954 book on person perception edited by Bruner and Tagiuri (it was a very good year for person perception). Cronbach discussed IPT in the context of personality ratings made by judges in traditional trait-oriented personality research, and suggested that in addition to information derived from the judge's observations of the target, the ratings will be influenced by the "Judge's description of the generalized Other" -- that is, by the judge's beliefs about what people are like in general. In Cronbach's view, IPT consists of several elements:
Cronbach believed that IPT was widely shared within a culture, but he acknowledged that there might also be individual differences in IPT. For example, some people might assume that most people are friendly and well-meaning, and that becomes the "default option" when they make judgments about some specific person; but other people might assume that most people are hostile and aggressive. In addition, he suggested that there may be cultural differences in IPT. For example, within Western culture, IPT seems to be centered on clusters of traits, or stable individual differences in behavioral dispositions; but other cultures might have more "situationist" or "interactionist" views of personality.
Cronbach, an expanded concept of implicit personality theory
might look like this:
For example, people might think that Rosenberg's traits of social and intellectual "good-bad" are normally distributed, with populations right about the midpoint.
Osgood's "Pollyanna Principle" reflects the assumption that the distribution of positive traits in the population is skewed toward the positive end of the continuum.
Thorndike's "halo effect" reflects the assumption that socially desirable traits, whether social or intellectual, are positively correlated with each other, as in the two-dimensional structure uncovered by Rosenberg.
Note for statistics mavens: the correlations between variables can be represented graphically by vectors, with the angles between vectors reflecting the correlation between them such that r is equal to the cosine of the angle at which the vectors meet (this is how factor analysis can be done geometrically). Thus, two variables that are uncorrelated with each other (r = 0.0) are represented by two vectors that meet at right angles (cos 90o = 0); two variables that are perfectly correlated (r = 1.0) are represented by two vectors that overlap completely (cos 0o = 1.0); and two variables that are highly but not perfectly correlated, say .60 < r < .65, are represented by vectors that meet at an angle of about 45o (cos 45o = 0.635).
theories of personality have been studied through the
application of multivariate statistical methods, such as factor
analysis, multidimensional scaling, and cluster analysis to
various types of data:
But because these techniques are very time-consuming, developments in implicit personality theory had to await the proper technology, particularly the availability of cheap, high-speed computational power. In the 1960s, as appropriate computational facilities became widely available, two competing models of implicit personality theory began to emerge.
The first of these was a semantic differential model of IPT based on Charles Osgood's tridimensional theory of meaning (e.g., The Measurement of Meaning by Osgood, Suci, & Tannenbaum, 1957). According to Osgood, the meaning of any word can be represented as a point in a multi-dimensional space defined by three vectors:
In this EPA scheme, closely related words are represented by points that lie very close to each other in this space. Osgood's method was to have subjects rate objects and words on a set of bipolar adjective dimensions. When these ratings were factor analyzed, the three dimensions of evaluation, potency, and activity came out regardless of the domain from which the objects and words were sampled -- people, animals, inanimate objects, even abstract concepts. If these three dimensions are the fundamental dimensions of meaning, they are likely candidates for implicit personality theory -- the cognitive framework for giving meaning to people and their behaviors -- as well.
The principal problem with the semantic differential model is that the best evidence for the three factors came from studies employing adjective checklists, where the adjectives on the list were deliberately chosen to represent evaluation, potency, and activity. Accordingly, it seems possible that evaluation, potency, and activity came out of analyses of adjective ratings because, however unintentionally, they were built into these ratings to begin with. Interestingly, Osgood's three dimensions were not clearly obtained from free-response data which was not constrained by the experimenter's choices. Left to their own devices, without the experimenter's constraints, subjects' cognitive structures look somewhat different from Osgood's scheme.
For example, Rosenberg and Sedlak (1972) asked subjects to provide free descriptions of 10 people each. These investigators then selected the 80 traits that occurred most frequently in these descriptions, and then submitted the 80x80 matrix of trait co-occurrences to a technique of multivariate analysis called multidimensional scaling (why only 80 traits? An 80x80 matrix, generating more than 3,000 unique correlation coefficients, exhausted the computational power available at the time). They found that Osgood's evaluation and potency factors were highly correlated (r = .97): people who were perceived as good were also perceived as strong. Osgood's activity factor was quite weak in the data, but was also positively correlated with the evaluation and potency factors (r = .57). Accordingly, Rosenberg & Sedlak concluded that the evaluation dimension dominated people's implicit theories of personality.
free-description data such as this, Rosenberg proposed an
alternative evaluation model of IPT. He argued that
evaluation was the only perceptual dimension common to all
individuals, and that any additional dimensionality came from
correlated content areas such as social and intellectual
evaluation. The figure graphically represents the loadings
on the two dimensions of a representative set of trait
adjectives. Note that Asch's central traits, warm-cold and intelligent-unintelligent,
lie fairly close to the axes that define the two-dimensional
Kim and Rosenberg (1980) offered a direct test of the two models. In the Rosenberg and Sedlak (1972) study, and other studies of implicit personality theory, individual subjects rated only a single person, and then the subjects' responses were aggregated, so that the resulting IPT structure reflected the average "Judge's description of the generalized Other". But averaging may obscure the structures that exist in individual Judges' minds. It is entirely possible that individual judges have something like the Osgood structure in their heads, but when their responses are aggregated, only evaluation remains. Accordingly, Kim and Rosenberg decided to compare the adequacy of the two models at the individual level (again, this is the kind of analysis that can only be done when computing resources are cheap). Because multivariate analysis requires multiple responses, they had subjects describe themselves and 35 other people that they knew well; and they collected both free descriptions and ratings on an adjective checklist. Multidimensional scaling of the individual subject data revealed that something resembling Osgood's three-dimensional EPA structure appeared in only 8 of the 20 subjects studied, and in 3 of these 8, the potency and activity dimensions were not independent of evaluation. More important, the evaluation dimension emerged from every individual subject's data set.
Rosenberg concluded that that Osgood's EPA structure was an
artifact of aggregation across subjects. All subjects use
evaluation as a dimension for person perception; some use
potency, some use activity, and some use both, and these
dimensions are strong enough to create the appearance of a major
dimension when data is aggregated across subjects. Potency
and activity are positively correlated for some subjects, and
negatively correlated for others; these tendencies balance out,
and give the appearance that potency and activity are
independent of each other and of evaluation. But this is
an illusion produced by data aggregation. In their view,
only the evaluation dimension is genuine; the potency and
activity dimensions are largely artifacts of method.
Fiske et al. (Trends in
Cognitive Sciences, 2007) has drawn new attention to Rosenberg's
work by claiming that warmth and competence are
universals in social cognition, which exert a powerful influence
on how we interact with other people. In particular, she
has argued that various combinations of warmth and judgment
characterize various out-group stereotypes. This is true
in both "individualist" and "collectivist" (or 'independent" and
"interdependent") societies. Fiske has even argued that
there is a specific module in the brain, located in the medial
prefrontal cortex (MPFC), which constitutes a "social evaluation
area". In any event, she and her colleagues have argued
that assessments of warmth and competence are made automatically
and unconsciously -- even though they may not necessarily be
Implicit theories of personality are like formal scientific theories, except that they are "naive" and "implicit". Recently, scientific research on personality has focused on a five-factor model of personality structure originally proposed by Norman (1963). In his research, Norman examined subjects' ratings of other people on a representative set of trait adjectives. Factor analysis reliably revealed five reliable dimensions of personality:
(1981) proposed that the Big Five comprised a universally
applicable structure of personality. By universally
applicable Goldberg meant that it could be used to assess
individual differences in personality under any circumstances:
Goldberg noted that the Big Five are so ubiquitous that they have been encoded in language, as familiar trait adjectives like extraverted and cultured. Of course, if ordinary "laypeople" (not just trained scientists) notice these dimensions enough to evolve words for them, the Big Five structure may exist in people's minds as well as their behavior. That is, the Big Five may well serve as the structural basis for people's implicit theories of personality, as well as a formal theory of personality structure.
these lines, I have often thought of the Big Five as The Big
Five Blind Date Questions -- representing the kind of
information that we want to know about someone that we're
meeting for the first time, and will be spending some
significant time with:
And in fact, there is some empirical evidence that The Big Five -- whatever its status as a scientific theory of personality -- serves as an implicit theory of personality as well.
The evidence comes from a provocative study by Passini and Norman (1966), who asked subjects to use Norman's adjective rating scales to rate total strangers -- people they had never met before, and with whom they were not permitted to interact during the ratings session. The subjects were simply asked to rate others as they "imagined" them to me. Nevertheless, factor analysis yielded The Big Five, just as had earlier factor analyses of ratings of people the subjects had known well. Note that the Passini and Norman study violates the traditional assumption of personality assessment: that there is some degree of isomorphism between personality ratings and the targets' actual behavior. In this case, the judges had no knowledge of the targets' behavior. The Big Five structure that emerged from their ratings was not in their targets' behavior -- simply because they had no knowledge of their targets' behavior; but it certainly existed in the judges' heads, as a "description of the generalized Other".
Based on this evidence, it may be that the Big Five provides a somewhat more differentiated implicit theory of personality than the two-dimensional evaluation model promoted by Rosenberg and Sedlak. If so, we would have another answer to the question of what makes a central trait central. Just as R&S argued that central traits loaded highly on the two dimensions of evaluation, perhaps central traits load highly on one or the other of the Big Five dimensions of personality. Certainly that's true for warm-cold, which loads highly on extraversion, and intelligent-unintelligent, which loads highly on openness.
By now there have been many studies similar to that of Passini and Norman, all with similar results: every factor structure derived from empirical observations has been replicated by judgments of conceptual similarity. Thus, we do seem to carry around in our heads an intuitive notion concerning the structure of personality -- the co-occurrences among certain behaviors, the covariances among certain traits, the notion that certain things go together, and other things contradict each other. This conceptual structure -- this implicit personality theory -- is thus cognitively available to influence people's experience, thought, and action in the social world.
existence of implicit personality theory is interesting, but in
some sense it is also troublesome, because it raises a difficult
question that has long bedeviled theorists of perception in
general -- the question of realism vs. idealism:
But this kind of evidence is problematic. In principle, factor analysis should be applied to objective observations. But for pragmatic reasons, this is generally impossible in the domain of personality research, simply because it is very difficult to perform the systematic observations of behavior that are required for this purpose. Because we have no direct measurements of personality traits, factor analysis is generally applied to rating data -- subjective impressions of behavior and traits; judgments that rely heavily on memory.
The problem is the reconstructive nature of memory retrieval. Memory for the past is contaminated by expectations and inferences. When factor analysis is applied to memory-based ratings, therefore, we cannot be sure what the factor matrix represents: the structure residing in the personalities of the targets, or the structure residing in the minds of the raters.
The fact is, we know from studies like Passini and Norman's that the structure of personality -- and, specifically, the Big Five structure that is so popular -- resides in the minds of raters. Therefore, it is possible that the structure of personality is to some degree illusory -- in a manner somewhat resembling the Moon illusion familiar to perception researchers. The moon looks larger on the horizon than at zenith, even though it isn't, because of "unconscious inferences" made by perceivers that take account of distance cues in estimating size. Perhaps personality raters make similar sorts of unconscious inferences in rating other people's personalities (or their own). Note that the existence of a moon illusion doesn't imply that there is no moon. It simply means that the moon isn't as big as it looks.
Similarly, in the realm of person perception, our expectations and beliefs can distort our person perceptions, and thus our person memories; in particular, our expectations and beliefs about the coherence of personality can magnify our perception of that coherence.
Research has already established that the structure of personality exists in the mind of the observer. The important question is whether it also has an independent existence in the world outside the mind. As in the moon illusion, we usually take a modified realistic view of perception -- that our perceptions are fairly isomorphic with the world. Accordingly, we may assume that our beliefs about personality are to some extent isomorphic with the actual structure of personality. But are there really?
controversy about the nature of implicit personality theory is
reflected in two competing hypotheses:
can test the accurate reflection hypothesis against the
systematic distortion hypothesis by comparing the structures
derived from three types of data:
experiment, by Shweder and D'Andrade (1980), employed 11
categories of interpersonal behavior as target items: these were
behaviors such as advising, informing, and suggesting.
For example, examining the correlations between parallel cells of the matrices, they observed the following pattern of correlations:
the results across the 7 different tests, they observed the
following pattern of correlations:
Systematic Distortion or Accurate Reflection?
Although the Shweder & D'Andrade study seems quite compelling, it has come under criticism from advocates of the accurate reflection hypothesis. In particular, UCB's own Prof. Jack Block has been an ardent defender of the notion that memory-based ratings, and implicit theories of personality, are accurate reflections of external reality. See, in particular, an exchange between Shweder and D'Andrade and Block, Weiss, and Thorne that appeared in the Journal of Personality & Social Psychology for 1979.
To be honest, the systematic distortion hypothesis is somewhat paradoxical, because it seems to refute the realist assumption that there is a high degree of isomorphism between the structure of external reality and our internal mental representations of it. Where does implicit personality theory come from, if not from the world outside? According to the ecological perspective on semantics, "the meanings of words are in the world" ( a quote from Ulric Neisser): our cognitive apparatus picks up the structure of the world, and so our mental representations are faithful to that structure. But they apparently aren't, at least in the case of person perception, where it is very clear that our cognitive structures depart radically from the real world that they attempt to represent.
So where else might implicit personality theory come from? How do our behaviors and traits become schematized, organized, and clustered into coherent knowledge structures?
and Shweder have suggested a number of possibilities:
Regardless of the ontological status of implicit personality theory, Asch's initial question remains on point: How do we integrate information acquired in the course of person perception into a unitary impression of the person along some dimension? Asch (1946) considered two possibilities: either we simply sum up a list of a person's individual features to create a unitary impression, or the unitary impression is some kind of configural gestalt. Asch clearly preferred the gestalt view to the additive view, a preference that integrated social with nonsocial perception, but his impression-formation paradigm has permitted later investigators to consider simpler alternatives.
Chief among these investigators
has been Norman Anderson (1974), who has promoted cognitive
algebra as a framework for impression formation and for
cognitive processing in general. According to Anderson,
perceptual information is integrated according to simple
algebraic rules, which take information (about, say, primary
traits) and performs a linear (algebraic) combination that
yields a summary of the trait information in terms of a
superordinate dimension (say, a superordinate trait).
particular, Anderson has considered two very simple algebraic
models (where S = stimulus information and R = the impression
Anderson's experiments include critical
comparisons that afford a test of the adding and averaging
models of impression formation. Assume that the trait
ensemble includes a mix of traits:
Thus, the adding and averaging functions give the following values to various trait ensembles:
Thus, the adding rule predicts that both the HHHH and the MMHH ensemble will be preferred to the HH ensemble; but the averaging rule predicts no difference between HH and HHHH, and that both will be preferred to MMHH.
(1965) actually performed the comparison, the empirical results
were a little surprising:
resolved the conflict by adding three new assumptions:
In a revised test, Anderson set aside the matter of stimulus weightings. Instead of asking subjects whether they were biased positively or negatively, he simply assumed that positive and negative biases would average themselves out, so that the average subject could be considered to be neutral at the outset (in fact, there is probably an average positive bias, but the essential point remains intact). Accordingly, a value of 0 was entered into the adding and averaging equations, along with the values of the stimulus information. Of course, adding 0 does nothing to sums; but it can have a marked effect on averages. Compare, for example, the following table to the table just above:
(1965) actually performed the comparison, the empirical results
were less confusing:
experiment seemed to offer decisive evidence favoring the
weighted averaging model of impression formation. In the
weighted averaging model, the perceiver's final impression
builds up slowly, and is heavily constrained by his or her
initial bias and first impressions.
Cognitive Algebra as Mathematical Modeling
Anderson's cognitive algebra is an attempt to represent a basic cognitive function as a mathematical formula. As such, cognitive algebra is intended to be a formal mathematical model of the impression-formation process. So, if you've always been wary of mathematical modeling (perhaps because you've thought it was too dry, or perhaps because of a little math phobia), but you've followed the arguments about cognitive algebra so far, then
You've just successfully worked your way through a mathematical model of a psychological process.
Anderson's cognitive algebra, and especially the weighted-averaging rule, is an extremely powerful framework for studying social judgment. Cognitive algebra can be applied to any social judgment, so long as the stimulus attributes are quantifiable, and so long as the perceiver's judgment response can be expressed in numerical terms.
cognitive algebra also has some problems:
Despite these problems, lots of work in social cognition has been done within the framework of cognitive algebra -- so much so that we could devote an entire course to it. But we won't.
Another prominent model
of person perception is the Social Relations Model (SRM) developed by David Kenny
(1994; for earlier versions, see Kenny & LaVoie, 1984;
Malloy & Kenny, 1986; Kenny, 1988), based on the early work
of the existential psychiatrist (before he became an "anti-psychiatrist") R.D.
Liang (Liang et al., 1966).
Link to an overview of the Social Relations Model, based on Kenny's 1994 book, Interpersonal Perception: A Social Relations Analysis. For critical reviews of the SRM, see the review of Kenny's book by Ickes (1996), as well as the book review essays published in Psychological Inquiry (Vol. 7, #3, 1996).
The SRM is focused on dyadic relations -- that is, relations between two people, say Andy and Betty, and in particular how these two people perceive each other. For purposes of illustration, let's suppose that Andy perceives Betty as high in interpersonal warmth. In Kenny's s analysis, this perception -- or impression -- has a number of components, such that A's perception of B's warmth is given by the sum of four quite different perceptions (actually, five, depending on how you count):
relationship among the components is additive, you can think of
the SRM as a version of Anderson's additive model for
impression-formation. But because the addition includes a
constant, which reflects A's biased view of people in
general, it's actually a weighted additive model.
But it's not necessarily a strictly additive model (which would
go against Anderson's results, which favor averaging). The
relationship component may well be achieved by an
averaging process. We're not going to get into that
detail: it will be enough just to explore the surface features
of the SRM.
This is because the SRM is actually quite complicated, because Kenny has built into the model two features that are not found in other, simpler models of person perception:
In applying the SRM,
Kenny prefers to employ a round-robin research design, in which each person in a group
rates everyone else in the group, as well as themselves.
The particular rating scales can be selected for the
investigator's purposes, but we might image that the ratings are of likability
(Anderson), warmth and competence (Rosenberg, Fiske), or the
Big Five traits of extraversion,
neuroticism, agreeableness, conscientiousness, and openness to
experience. Of course, selection of the traits
to be rated matters a great deal: results may differ
greatly if subjects are forming impressions of a
person's masculinity, sexual orientation,
or likeability, or extraversion.
Note that in the round-robin design, the perceiver is also a target, and the target also a perceiver -- just as in the General Social Interaction Cycle, the actor is also a target and the target also an actor. The SRM treats the individual as both subject and object, stimulus and response, simultaneously.
The mass of data from the round robin design is then decomposed into three components:
With the round-robin
design in hand, Kenny can proceed to address a number of
questions about interpersonal perception. Here are these questions, and short
answers, based on some 45 studies reported in Kenny's 1994
cognitive algebra, Kenny's
Social Relations Model is
more a method than a
round-robin design, coupled
statistical tools, can be
used to partition person
perception into its various
components, and so to answer
a wide variety of questions
about a wide variety of
topics in social
Research on impression-formation, from Asch (1946) to Anderson (1974) and beyond, has largely made use of trait terms as stimulus materials. This is certainly appropriate, because -- as Fiske & Cox (1979) demonstrated, as if we needed any proof -- we often describe ourselves and others in terms of traits. Working with traits injects substantial economies into impression-formation research, because they're easy to manage. Moreover, traits may fairly closely represent the way information about people is stored in social memory. But it's also clear that an exclusive focus on traits can give a distorted view of the process of impression formation, because traits are not really -- or, at least, not the only -- stimulus information for social perception.
We don't walk around the world with our traits listed on our foreheads, to be read off by those who wish to form impressions of us. Rather, the real stimulus information for person perception consists of our physical appearance, our overt behavior, and the situational context in which they appear. Accordingly, in addition to describing how we make use of trait information to form impressions of personality, a satisfactory account of person perception needs to answer a different sort of question -- to wit:
Or, put another way,
Again, the same
question about perception
occurs in the social domain
as in the nonsocial domain.
In the social
consists of a
The use of physical features to make inferences about character and personality has its roots in physiognomy, a pseudoscience in which a person's character was judged according to stable features of the face -- much as the 19th-century phrenologists judged character from the bumps and depressions on the skull.
The word physiognomy comes from the Greek Physis (nature) and gnomon (judge), and began with the observation that some people looked like certain animals. It was only a short step, then, to infer that those individuals shared the personality traits presumed to be characteristic of those animals. More generally, physiognomy was based on the assumption that a person's external appearance revealed something about his internal personality characteristics.
References to physiognomy go back at least as far as Aristotle, who wrote in his Prior Analytics (2:27) that
Aristotle (or perhaps one of his students) actually produced a treatise on the subject, the Physiognomonica.
disrepute in the
and was revived
della Porta (De
In a study of
based on a
|Here are some physiognomic drawings by Charles LeBrun (1619-1690) a French artist who helped establish the "academic" style of painting popular in the 17th-19th centuries (from Charles LeBrun -- First Painter to King Louis XIV).|
good example of
involving descriptions of
physical stimuli is the work
of Ekman (1975, 2003; Ekman
& Friesen, 1975) and
others on facial expressions
of emotion. In his
work, Ekman has been
particularly concerned with
determining the "sign
vehicles" by which people
about their emotional states
to other people. The
fact that such communication
occurs necessarily entails
that there is a receiver
who is able to pick up on
the communications of a sender
-- and this information
pickup is exactly what we
mean by perception.
Facial Expressions in Art
Among the many formalisms taught in European painting academies in the 17th-19th centuries were standards for the depiction of emotion on the face. Among the most popular of these texts was the Methode pur apprendre a dessiner les passions proposee dans une conference sur l'expression general et particuliere (1698) by Charles LeBrun, a leader of French academic painting. Here are samples from LeBrun's book, showing how various emotions should be depicted (from Charles LeBrun -- First Painter to King Louis XIV).
One of Ekman's
most famous findings is that
people can reliably "read"
certain emotions from the
expressions on people's
faces. This is true
even when the sender and
receiver come from widely
Close analysis of these
expressions shows that each
of them is comprised of a
particular configuration of
muscle activity. These
system for coding the facial
musculature is known as the
Facial Action Coding System
(FACS). The system has
more than 60 coding
categories for various
muscle action units (like
the Inner Brow Raiser
or the Lip Corner Puller)
and other action descriptors
(such as Tongue Out
or Lip Wipe).
Each basic emotion, and
every variant on each basic
emotion, can be described as
a unique combination of
categories. And each
coding category is
associated with a specific
pattern of muscle activity.
Cross-cultural studies show that Ekman's basic emotions are highly recognizable across cultures. Nelson and Russell (2013) summarized several decades' worth of such studies, involving subjects from literate Western cultures (mostly, frankly, American college students), literate non-Western cultures (e.g., Japan, China, and South Asia), and non-literate non-Western cultures (e.g., indigenous tribal societies in Oceania, Africa, and South America). Subjects from all three cultures recognized prototypical displays of the six basic emotions at levels significantly and substantially better than chance. This is consistent with the hypothesis that the basic emotions, and the apparatus for producing and reading their displays on the face, is not a cultural artifact but something that is, indeed, biologically basic.
like this is
They are a
product of our
with at least
is a product
And, as the
The ability to
read the basic
does not depend
literacy, or stage
origins in the
also in the
mentor; but it
thesis is widely
those who have
least, it has
They note, in
emotions is not,
look at the
of the Nelson
see clearly that,
as we move to
Just as the face may be the pre-eminent social stimulus, so the smile may be the pre-eminent social behavior.
for example, Ekman distinguishes between two
kinds of smile:
According to the
Associated Press, customer-service employees
at the Keihin Electric Express Railway Company
in Japan can check their smiles against the
Okao Vision face-recognition software system,
to make sure that they are smiling properly at
customers "Japan Train Workers check Grins
with Smile" by Jay Alabaster, Contra Costa
Times 07/26/2009). It's not clear
whether they're being checked against a
Duchenne smile or a Pan American smile.
And as another example, anger involves a large number of muscles. So, as your grandmother told you, it really does take more muscles to frown than to smile. So smile and save your energy.
Ekman's analysis of facial emotion
has been used to offer a solution to a famous
question in art history: what is it about the
smile of Mona Lisa, in Leonardo da
Vinci's famous painting (c. 1503-1505)?.
It's not just Nat "King" Cole who has found
this smile mysterious. Part of the
mystery of the smile is the ambiguous way it's
painted -- which, according to a conventional
theory, reflects the "archaic smiles" in the
ancient Greek and Roman paintings and
sculpture that so inspired Leonardo and other
artists of the Renaissance.
In 2005 NIcu Sebe, a computer-vision researcher at the University of Amsterdam, scanned the Mona Lisa with an emotion-recognition program he developed with colleagues at the Beckman Institute of the University of Illinois, and based on Ekman's analysis of facial expressions of basic emotions. Using this program, he determined that the Mona Lisa's smile consisted of 83% happiness, 9% disgust, 6% fear, and 2% anger (New Scientist, 12/17/05). So that's part of the mystery. Or maybe it's just a smirk, as in this New Yorker cartoon by Emily Flake (08/30/2021).
On the other hand, Peter Schjeldahl, commenting on the sale (for almost half a billon dollars) of another Leonardo painting, Salvator Mundi, remarked on the "ambiguous Mien" of Jesus, and went on to write "Giving an ambiguous character an ambiguous mien doesn't seem a stop-the-presses innovation. The trick of it, by the way is the same as that of the "Mona Lisa": painting different expressions in the eyes and in the mouth. When you look at one, your peripheral sense of the other shifts, and vice versa. You try to reconcile the impressions, with frustration that seeks and finds relief in awe."
Since then, work on computer
recognition of emotion, based largely on
facial cues, has progressed apace, evolving
into a new sub-discipline known as affective
computing. Among the most highly
developed of these systems is Affdex,
a product of Affectiva,
an offshoot of the MIT Media Lab. Based
largely on Ekman's FACS system, Affdex scans
the environment for a face, isolates it
from its background, and identifies major
regions such as mouth, nose, eyes, and
eyebrows -- distinguishing between non
such as the tip of the nose, which
remain stationary, and deformable
points, such as the corners
of the lips, which change with different
facial expressions. It
computes various geometric
relations between these points,
and compares the current face to a
very large number of other faces, previously
analyzed, stored in
memory. It then outputs a
probabilistic judgment of
whether the face is displaying
such basic emotions as
happiness, disgust, surprise,
confusion. It can
distinguish between social
smiles and genuine "Duchenne"
smiles, and between real and
feigned pain. And it does
this in real time.
For an article on affective computing, including the story of how market forces turned Affdex from an emotional prosthetic for autistic people into a marketing tool, see "We Know How You Feel" by Raffi Khatchadourian, New Yorker, 01/19/2015.
FACS is intended for use by professional researchers and clinicians, including computer analysis of facial expressions. But the fact that people can reliably read emotions from other people's faces suggests that our perceptual systems are sensitive to changes in facial musculature. Ekman's FACS system is a formal description of the physical stimulus that gives rise to the perception of another's emotional states.
The stimulus faces used in much of Ekman's research comes from actors posing various expressions according to Ekman's instructions, but we can read emotional states from people's faces in other circumstances as well.
Consider, for example, photographs taken in April 2000, when Elian Gonzalez, a Cuban boy who had lost his mother during an attempt to escape from Cuba, was being sheltered by some of his mother's relatives in Miami. Elian's father, who was estranged from his mother, and had remained in Cuba, demanded that he be returned to Cuba. The US Department of Justice, for its part, determined that, legally, custody of the boy should be given to his closest living relative -- his father, who was in Cuba. The Miami relatives refused to turn Elian over for repatriation, and in the final analysis an armed SWAT team from the Border Patrol forced its way into the relatives' house to retrieve the boy. Little did the officers know that a newspaper reporter and photographer were already in side the house. The resulting remarkable sequence of photographs shows the surprise of one officer when he discovered the photographer in the room.
Like Ekman, Paula Niedenthal and her colleagues have cataloged a number of different types of smiles, but the differences they observe go far beyond patterns of facial activity.. It turns out (to quote the headline in the New York Times over an article by Carl Zimmer, 01/25/2011), that there's "More To a Smile Than Lips and Teeth". Some smiles are expressions of pleasure, while others are displayed strategically, in order to initiate, maintain, or strengthen a social bond; some smiles comprise a greeting, others display embarrassment -- or serve as expressions of power. Niedenthal has been especially active in examining the process of smile recognition -- recognizing the differences among smiles of pleasure, embarrassment, bonding, or power. She proposes that smiles are "embodied" in perceivers through a process of mimicry, in which different types of smiles initiate different patterns of brain activity in the perceiver -- patterns that are similar to those in the brain of the person doing the smiling.
found, not surprisingly,
that subjects could
between these different
types of smiles. But
when they held a pencil
between their lips,
essentially interfering with
the facial musculature that
would mimic the target's
smile, accuracy fell off
the judgments of subjects in
the "pencil" condition were
more influenced by the
contextual background, than
by the smiles
Apparently, mimicry plays an
important role in the
recognition of smiles (and,
probably, other facial
expressions as well).
work exemplifies a larger
movement in cognitive
psychology known as embodied
cognition or grounded
most of its history,
psychology has assumed that
the brain is the sole
physical basis of mental
life. Embodied cognition
assumes that other bodily
processes -- in this case,
the facial musculature --
are also important
determinants of mental
states. And so is the
environment -- like the
context in which a smiling
Proponents of embodied
cognition do not deny the
critical role of the brain
for the mind. They
just argue that other
factors, in the body outside
the brain, and in the world
outside the body, are also
Full disclosure: Prof. Niedenthal worked in my laboratory as an undergraduate. But even so, her work on the smile is the most thorough analysis yet. For an overview of her work on smiles, see P.M. Niedenthal et al. "The Simulations of Smiles (SIMS) Model: Embodied simulation and the Meaning of Facial Expression", Behavioral & Brain Sciences, 33(6), 2010.
Ekman and Darwin
work on facial
in his book on The
in Men and
that the facial
emotions as fear
by which other
animals, such as
apes and dogs,
emotion are part
Ekman edited the
3rd edition of
According to Ekman, Darwin made five major contributions to the study of emotional expressions (Transactions of the Royal Society, 2009):
that there are
at least six basic
with an evolved
mode of facial
may also be
Ekman's evolutionary theory of facial emotion is interesting, but we do not have to accept it to construe his work on emotional expression as an aspect of person perception. After all, the fact that people can "read" emotions in others' faces is precisely what we're interested in: how we get from physical stimulus information -- the facial expression -- the perception of the states (cognitive, affective, conative) of the person.
The face is a major channel for communicating emotional states, and probably the most important, but it is not the only one. Tone of voice, gesture, posture, and gait are also available channels -- although they have not been given as much systematic attention as the face.
importance of nonfacial
expressions of emotion is
underscored by cases of
Moebius Syndrome, a
congenital condition first
described by Paul Julius
Moebius in 1888. The
condition entails a
paralysis of the facial
illustration shows Kathleen
Bogart, who has Moebius
syndrome, and who studies
the disorder, with her
husband, Beau, from the New
with Moebius syndrome cannot
express emotions on their
faces, so they must find
other means of emotional
expression, including both
verbal and nonverbal
channels. They have no
difficulty recognizing other
people's facial expressions,
however. And they
still feel various
emotions. This is
important, because a major
theory of emotion
communication implies that
we mimic other people's
facial expressions, and
feedback from our own facial
expressions shapes both our
perception of their
emotional states, and our
experience. That can't
happen in cases of Moebius
syndrome, of course, because
the facial paralysis
prevents the feedback.
so either mimicry isn't
important, or there are
other mechanisms for emotion
problem is encountered in
Bell's palsy, a
involving the usually)
temporary paralysis of
the facial musculature
caused by inflammation
of the VII cranial
Kalb, a theater
professor at Fordham
written about his own
experience with Bell's
palsy in "Give Me a
Smile" (New Yorker,
has never completely
recovered from the
illness, with the
result that his smile
is "an incoherent
tug-of-war between a
grin on one side and a
frown on the other: an
expression of joy
spliced to an
reports that he has
positive affect to
other people, and
they have difficulty
He also suggests
that, because of
from the facial
affect, and must
some drawn from
topic for person perception
research has to do with the
perception of facial
beauty. We know
from research on
that physical attractiveness
is the most powerful
determinant of likeability
(e.g., Berscheid &
Walster, 1974). And we
also know that likeability
-- evaluation, in
Anderson's terms --
influences a host of social
judgments through the halo
effect. But exactly
what determines physical
attractiveness remains a
mystery. As Berscheid
and Walster (1974), two
social psychologists who are
probably the world's
foremost experts on
concluded, "There is no
answer to the question of
is this question
important? One reason
is Thorndike's halo
tend to believe (regardless
of whether it's actually
true) that socially
desirable features go
if someone is physically
attractive, they'll also
tend to think that they're
socially attractive -- on
the "warm" end of the social
good-bad scale, and on the
"intelligent" end of the
scale. As the English
Romantic poet John Keats
wrote (in Ode on a
Grecian Urn, 1819),
"Beauty is truth, truth
Interestingly, there may also be a reverse halo effect. Vincent Yzerbt, Kocolas Kervyn, and their colleagues have found that, when comparing people with each other, a person (or group) who receives high ratings on warmth may receive low ratings on competence, and vice-versa. Apparently, the traditional halo effect occurs when evaluating individuals separately, while the reverse halo effect occurs when comparing one individual with another.
There's no question about the bias toward facial attractiveness -- and not just in bars and bedrooms. A number of writers have commented on the pervasiveness of "lookism", a concept modeled on racism, having to do with discrimination against those who are less than a perfect "10" (to use the title of a 1979 movie on this theme starring Dudley Moore, Julie Andrews, and Bo Derek as the eponymous beauty. For more on lookism, see the following books, discussed by Rachel Shteir in "Taking Beauty's Measure" (Chronicle of Higher Education, 12/16/2011):
Actually, maybe there is. A large body of literature now strongly suggests that attractiveness is strongly related to averageness -- in other words, that we find most attractive those faces (and, for that matter, bodies) that are close to the average for the population. As counterintuitive as that may seem, there are actually good reasons to think that average faces really are highly attractive.
Referring to the Berscheid and Walster (1974) quote above, Langlois and Roggman (1990) concluded that the question of facial beauty had been solved: [A]ttractive faces... represent the central tendency or the averaged members of the category of faces".
Averageness or Symmetry?
But the question isn't entirely resolved, because evolutionary psychology has a somewhat different answer to the question of why we prefer average faces. According to evolutionary psychology, patterns of experience, thought, and action that were adaptive in our ancestral environment (the Environment of Early Adaptation -- roughly the East African savanna during the Pleistocene era) have been preserved in current members of the human species through natural selection. In this view, mate selection prefers healthy, fecund mates: facial symmetry is a marker of health and fecundity, while fluctuating asymmetries on the face (and elsewhere on the body are signs that the organism is unhealthy, and less desirable from the point of view of reproductive fitness. Averaging eliminates these fluctuating asymmetries, and produces symmetrical faces. So, according to evolutionary psychology, average faces are not attractive because prototypes seem familiar, but because average are more symmetrical.
There's certainly anecdotal evidence in favor of a connection between symmetry and attractiveness. Queen Nefertiti, wife and co-ruler of ancient Egypt with the Pharaoh Akhenaten (14th century BCE) was widely acclaimed as the most beautiful woman in the ancient world (this was before Helen of Troy): her name Nefertiti even means (in rough translation) "The perfectly beautiful woman has come". And Nefertiti is portrayed in images that survive from her time as having a perfectly symmetrical face. Of course, these are only images -- we don't know what she really looked like. She might just have had a good public-relations firm.
But we do know what the actress Elizabeth Taylor (who died in 2011 at age 79) looked like, and she really did have a perfectly symmetrical face -- and was universally acknowledged as fabulously beautiful.
On the other hand, 20th-century culture gives us lots of examples of very attractive woman who have prominent fluctuating asymmetries on the face: consider, for example, the prominent "beauty marks" on the faces of Marilyn Monroe and Cindy Crawford. Beauty marks are called "beauty marks" precisely because they enhance the person's facial beauty, but as fluctuating asymmetries they're supposed to mark a lack of reproductive fitness, and thus make the person less attractive, not more.
So something's wrong with the evolutionary argument. In fact, the evolutionary story, like many of the "just-so" stories that abound in evolutionary psychology, sounds good, but doesn't stand up to close scrutiny.
In the first place, the connection between facial attractiveness and reproductive fitness appears to be pretty weak, perhaps nonexistent. Kellick, Zebrowitz, Langlois, and Johnson (1998) analyzed data from the Intergenerational Studies conducted by the Institute for Human Development at the University of California, Berkeley, which included data from a large group of individuals who were born in the Berkeley-Oakland area between 1920 and 1929. These subjects had been photographed as adolescents, and health assessments had been made on them during adolescence (ages 11-18), middle age (30-36), and old age (56-66). There was essentially zero correlation between facial attractiveness, rated from the adolescent photographs, and health at any stage of life. So, attractiveness does not seem to serve as a marker of health -- and thus of reproductive fitness. Men aren't attracted to women because they think they'll produce lots of healthy babies. Men are attracted to women because -- well, they're attractive.
the second place, the
relation between averageness
and attractiveness does not
appear to be mediated by
symmetry. In another
experiment, Rhodes, Sumich,
and Byatt (1999) employed
of facial photographs that
varied in their averageness,
as defined in the Langois et
al. (1990) study.
Subjects then rated these
photographs on symmetry,
Rhodes et al. (1999) assert that their experimental results "settle the dispute" between averageness and symmetry. But the question remains why average faces are attractive. Their best guess (and mine) is that average faces look like lots of other faces, and so they seem familiar; and we know from the mere exposure effect that we find the familiar more attractive than the unfamiliar.
As the examples
of Marilyn Monroe and Cindy
Crawford suggest, there's
probably more to facial
beauty than averageness and
symmetry. In addition
to those "beauty marks",
there's skin tone, body-mass
index and waist-to-hip ratio
-- and a genuine smile.
facial feature that has been
studied by researchers of
person perception is babyfacedness.
such as Konrad Lorenz have
long noted that immature
organisms, whether mammals
or even birds and reptiles,
share certain features in
Using computer "morphing" programs, it is possible to take line drawings or photographs of faces and adjust their features to make them appear more or less baby-like. Subjects then rate the targets for various personality traits. Research by Zebrowitz and her colleagues generally finds that people perceive baby-faced individuals as warmer, weaker, more naive and trusting; they are also more likely to help baby-faced people, even when help isn't needed.
In one study, Friedman and Zebrowitz (1992) took schematic drawings of male and female human faces and manipulated their facial features to create or erase aspects of babyfacedness. As it happens, there is a sex difference here, with the typical female face possessing more "baby-faced" features than the typical male face. Therefore, by adding baby-faced features they made the typical male face more "babyish" in appearance, and the typical female face appear more "mature". They then had male and female subjects view the sketches, and make ratings of their impressions of the targets' personalities.
Baby-faced males and females alike were rated lower on power, compared to their mature-faced counterparts. But because the typical male face has more mature features than the typical female face, the typical male was rated as more powerful than the typical female.
Baby-faced females (but not baby-faced males) were rated higher on warmth than their mature-faced counterparts. Again, because the typical male face has more mature features than the typical female face, the typical male was rated as less warm than the typical female.
Perhaps not surprisingly, because of the sex difference in babyfacedness, baby-faced males and females alike were rated lower on masculinity (and thus higher on femininity), compared to their mature-faced counterparts. Again, because the typical male face has more mature features than the typical female face, the typical male was rated as more masculine than the typical female.
Baby-faced females (but not baby-faced males) were rated more likely to be the "child caretaker" in the family higher on warmth than their mature-faced counterparts. Again, because the typical male face has more mature features than the typical female face, the typical male was rated as less likely to be a child caretaker than the typical female.
Baby-faced females (but not baby-faced males) were rated less likely to be the "financial provider" in the family than their mature-faced counterparts. Again, because the typical male face has more mature features than the typical female face, the typical male was rated as more likely to be a financial provider than the typical female.
These are social stereotypes, of course, but that's the point: social perceivers use the physical properties of the face to make inferences about the emotions and dispositions of the person.
The baby-faced stereotype is so commonly held that it has been employed in cartoon characters (such as Elmer Fudd and Tweetybird). And in political humor as well. After Vice President Cheney was involved in a quail-hunting accident, in which he peppered one of his companions with bird shot, he was depicted as Elmer Fudd -- a clear contrast between the befuddled lovableness of the cartoon character and Cheney's own reputation as a humorless right-wing ideologue.
Of course, "baby-facedness" is a stereotype, and stereotypes can be misleading, and sometimes downright wrong. For example, Umar Farouk Abdulmutallab, the "Underpants Bomber" of Christmas Day, 2009, was commonly described in press accounts as "baby-faced".
research on social
perception has focused on
the face, which is after all
the most salient, perhaps
the quintessential, social
other nonverbal cues play a
part in person perception,
including vocal (prosodic)
cues, gestures, and other
aspects of body
language. Here, in a
famous photograph from the New
York Times (1957), the
future president Lyndon B.
Johnson (then Majority
Leader of the United States
Senate) discusses a point of
legislation with a
Edward T. Hall first drew attention to several aspects of body language in his popular book, The Hidden Dimension (1966).
Robert Rosenthal and his colleagues have developed a psychological test to assess individual differences in people's sensitivity to nonverbal cues -- including, but going beyond, facial cues (Rosenthal, Hall, DiMatteo, Rogers, & Archer, 1979). The Profile of Nonverbal Sensitivity (PONS) consists of 220 2-second audio/video clips portraying a 24-year-old woman (Judith Hall, now a Professor of Psychology at Northeastern University) acting out a set of 20 vignettes. The subjects' task is to guess which of two vignettes is being acted out.
are classified into a 2x2
scheme crossing positive-negative
with 5 scenes in each
Rosenthal and his colleagues were interested in using the test to measure individual differences in sensitivity -- i.e., in perceptual ability to various channels of nonverbal communication. In the context of this course, the PONS is a good illustration of the point that there are physical sources of social information beyond the face, including vocal and gestural cues.
Person perception is shaped by the person's physical features, but it is also influenced by aspects of his or her dress -- what one hides and reveals, what one draws attention to are also stimulus cues as to a person's internal psychological state. Even a person's office or bedroom can provide clues to his or her personality.
A Different Sort of "White Coat Syndrome"
An interesting illustration of this point occurred in 2000, in a dispute over the dress code at Duke University Medical Center (a similar dispute also arose at the Massachusetts General Hospital in Boston, one of Harvard's teaching hospitals). At Duke, physicians wore two different types of white coats: a knee-length duster was restricted to senior physicians, while a hip-length jacket was imposed on interns and residents. Thus, in the Duke (and MGH) environment, the type of white coat worn by a physician conveyed information about the person wearing it --his or her level of training, and presumed levels of knowledge and expertise.
Anyway, Duke's medical residents complained about this policy. They said that they usually wore white trousers with the short white jacket, and the combination made them feel like ice-cream vendors. More important, perhaps, patients picked up on the fact that senior physicians wore long coats, and often questioned the profession, status, and competence of those who wore short jackets. Moreover, female residents were often confused with nurses, who also wore white coats; for the same reason, male residents were often confused with the janitorial staff. The short coat-long coat distinction even affected relations among the medical staff. Senior staff were more likely to speak to residents who wore long coats, and especially more likely to address them as equals.
In the event, Duke altered the dress code as it pertained to residents, but short coats were still imposed on interns. So, class distinctions prevailed after all!
The white coat, which has been worn by physicians since the scientific revolution in medicine of the late 19th century, is part of the identity of most physicians, and part of the concept of physician held in the mind of the public. If nothing else, it creates a link between medicine and science.
But not for long. As Vinciguerra notes, it has become increasingly clear that the white coat, whether long or short, is a major carrier of bacteria and thus a major source of hospital-based infections, and a major contributor to morbidity and mortality -- not to mention increased healthcare costs. In 2007, the British national health system adopted a policy of "bare below the elbow" banning lab coats as well as long-sleeved shirts and blouses, neckties, long fingernails, and jewelry on the hands and wrists. Maybe physicians will have to settle for wearing their stethoscopes.
Much of Ekman's work on facial emotion, and much of the interest in nonverbal communication generally, has to do with the detection of deception -- or, put bluntly, with lie detection. How can we know when someone is deceiving us? Note that, in terms of person perception, the question of deception is this: how can we know that a person is deceiving us about his or her internal mental state -- about what he or she is thinking, feeling, or desiring? Ekman's work on behavioral (as opposed to physiological) lie-detection has been extremely influential. He has consulted with law-enforcement agencies at all levels of government, and has even "gone Hollywood" as a consultant to the TV show Lie to Me (Fox), about Dr. Cal Lightman (played by Tim Roth), a "human polygraph" who can read body-language "micro-expressions".
The problem, of
course, is that people
rarely tell us that they are
lying -- what would be the
point of that? (Even
Epimenides, the Cretan
philosopher of the
6th-century BC, who asserted
that "All Cretans are
liars", could not have been
lying, because if all
Cretans really were liars,
he -- a Cretan himself --
would have been telling the
truth, thus disproving his
Instead, we usually have to
infer, from their nonverbal
behavior, that their verbal
communications are not
And let's be
clear -- lying is a
serious problem of social
et al. (1996) conducted a
study of everyday lying by
means of diary study
in which subjects
were asked to keep
track of all of their
for a week,
in which they
found that lying is
a common feature of
twice a day, on
average, in 1/3
of all their
once a day, in
about 1/5 of
DePaulo et al.
point out that
most of these
So lying is an
important aspect of
social interaction, and so our ability to detect lying is an
important aspect of social perception.
out that we are surprisingly
bad at this. Our poor
lie-detection abilities were
dramatically illustrated in
a study by Ekman and
O'Sullivan (1991). For
this study, they created
1-second video clips,
showing the full
of a target's face
Then the target
described his or
her positive emotions
as s/he was
in which case s/he was
telling the truth about
state. The other half
of the targets were actually
viewing a very gruesome
scene -- in which case, s/he
was not telling the
truth. The subjects'
task was to identify which
of the targets were
the truth, and which were
and O'Sullivan tested
groups of subjects,
ranging from college
Averaged across all the groups, the subjects were only about 57% correct -- barely above chance levels. Only agents of the United States Secret Service, a branch of the Treasury Department that has responsibility for protecting the President and other high officials, were particularly good at picking out liars: 53% had 70% or greater accuracy, compared to 50% "chance" level.
A later study by Ekman, O'Sullivan and Frank (1999), focused on subjects who had special professional interests in lie-detection yielded similar, if somewhat better, results. This time, the subjects achieved about 63% accuracy -- which is better than chance, but not all that great. The top scorers (those with 70% or greater accuracy) were federal "law enforcement" officers, most of whom were actually agents of the Central Intelligence Agency.
A cautionary note: The data in this study was collected while Ekman delivered a research on behavioral lie detection to various professional audiences. After presenting the film clips, Ekman revealed to his audience which targets had been lying, and which telling the truth. He then asked the members of the audience to raise their hands if they got 10, 9, 8,etc. correct, and tallied the results. Given natural tendencies for self-enhancement, it seems likely that the audience self-reports of accuracy were inflated somewhat. By how much, however, we cannot know for sure. Presumably, the same polling procedure, also resulting in possibly inflated scores, was employed in a 1999 follow-up
So while most people are
pretty bad at behavioral
people are better than
Ekman argued that
possible when perceivers
up on the leakage
tend to display
truth, but "Pan
also tend to
show an increase
Ekman and his
able to detect
means, such as
and noticing micro-expressions
of affect that are
incongruent with the
content of the target's
can also be picked up in real
time, especially by
people -- like Secret
and CIA agents,
perhaps -- who have
had a lot of
the 1991 and
able to pick up
situation is a little more
complicated than this,
because even some of the
"top scorers" didn't perform
better than chance.
This is a little
you'd think that anything
better than 50% would count
as "greater than
chance". But as
Nickerson and Hammond (1993)
pointed out, when the
probability of a hit p
equals probability of a miss
q, even 8 hits out of
10 is not significantly
greater than chance with p
.05 (actually, it just
detection of deception can
be construed as a problem
for signal detection
contrast to traditional
analyses of accuracy, which
focus on hits (and their
focuses on hits and false
alarms -- in this
case, instances where a
target is called a liar but
is actually telling the
truth. If you call
everyone a liar, you'll
correctly identify every
actual liar, but you'll also
misidentify all the
lie-detection will maximize
hits while minimizing false
problem has to do with the
measure of "accuracy"
employed in these Ekman
studies, which takes only
correct responses into
example, in the studies
described, a subject would
"catch" 100% of the
liars simply by calling
everyone a liar.
So it's important to
take error into
account. From this
perspective, we can
responses into four
It was just to address this
problem that signal detection theory (SDT) was invented
(Green & Swets, 1966; see also Tanner & Swets,
1954). In sensory psychophysics, the observer's problem is
to discriminate between trials in which a signal is presented
against a background of noise, and other trials in which only
noise is presented, no signal. On any trial, an observer
might actually detect the signal. Alternatively, he might
miss the signal, because it's too faint. Or, the signal
might be strong enough, but he might miss it because he's not
expecting it. Or, he might miss it because the costs of
making a mistake are relatively low. There are other
possibilities. In any event, the point is that the
observer's performance must take account f both the observer's
sensory acuity and his s biases, expectations, and
motivations. SDT does this by separating performance into
For the purpose of this
course, you don't need to know how to calculate either d'
or beta (or any of the other SDT
parameters). You just need to know the concepts.
are set up so that on some trials (e.g., half), a signal is
presented against a background of noise; on other trials, the
signal is omitted, and only the noisy background is
presented. On each trial, the observer responds with a
"Yes", indicating that the signal was present, or a "No",
indicating that it was absent. The 2x2 arrangement yields
the proportion of trials representing "Hits", "Misses", "False
Alarms", and "Correct Rejections". The translation of this
framework into the lie-detection situation is obvious.
Unfortunately, all too many studies of lie-detection aren't amenable to analysis in terms of SDT, because all too many investigators fail to report false alarms as well as hits. This was the case with the Ekman & O'Connor (1991) study, but Ekman et al. (1999) did report separate values for accuracy in lie-detection and accuracy in truth-detection, which enables us to calculate the false-alarm rate (as 100% - accuracy in truth detection).
the Federal Officers, accuracy in lie-detection was 80%,
while accuracy in truth-detection was about 66% -- which
means that truth-tellers were falsely called liars about
34% of the time. Applying the formulas of SDT yields
a d' measure of sensitivity of about 1.26, and a C
measure of bias of about -.21.
All by themselves, these numbers aren't too meaningful. But if you construct tables representing the values of various combinations of hits and false alarms, you can get a sense of the subjects' performance.
all the subjects in the 1999 study, the results were much
the same, except that all of the subjects taken together,
including the Federal Officers, showed less sensitivity (d'
= .66), and only a very slight liberal bias (C =
Apparently, if most people are bad at detecting lies, most of us are pretty good at lying. In fact, maybe that's why we're so bad at lie-detection: it's not so much that we're bad at detecting lies, but that we're so good at lying undetectably!
Or, put another way, lie-detection is a problem of signal-detection, and people are typically bad lie-detectors because so often there is no signal to detect!
Ekman and O'Sullivan based their conclusions about people's lie-detection abilities on their own studies -- where, frankly, the experimental procedures are somewhat informal (the subjects are typically members of the audience to whom Ekman is giving a talk). More systematic laboratory research comes to much the same conclusion: People just aren't particularly good at it.
In their reviews of the experimental literature, Kraut (1980), Vrij (2000), and Bond and DePaulo (2006) all found that the average receiver was barely better than chance at detecting lying under natural conditions -- that is, when the senders included both leaky and non-leaky liars. Things looked a little better, though, when B&DeP looked at continuous ratings of honesty, rather than dichotomous judgments of lying. Under these circumstances, honesty ratings distinguished between liars and truth-tellers to a modest degree.
course, these results can also be cast in terms of
signal-detection theory. Applying the formulas, as
we did in the study by Ekman et al. (1999):
the problem with
lie-detection may be that
most of us may tend to
assume that people are
telling the truth.
B&DeP discovered a small
"truth bias" -- to judge
that people are telling the
truth, even when they're
not. This effectively
reduces our ability to make
correct judgments of the
examined a number of other
variables that might affect
|There is a modality effect, such that lie-detection is more accurate when receivers have access to an audio channel, as well as -- or instead of -- a video channel. This may also help explain Ekman & O'Sullivan's results: because of their emphasis on facial cues, they typically present subjects with only a visual channel.|
of deception does
not depend on
whether the sender
motivated to be
fact, there is a paradox
such that the
bias" is reduced
when the sender is
receivers are not
very accurate at
Nor does it depend on whether the sender has been given an opportunity to prepare to deceive. Nor does the receiver's expertise matter much (remember, even Ekman and O'Sullivan's experts weren't all that good, and there aren't that many of them to begin with.
However, detection of deception is better if the receiver has had some prior exposure to the sender -- that is, prior to the experimental test.
|A very provocative finding is that third parties, observing the interaction between the sender and the receiver, are better at detecting deception than the receivers are.|
is that people do not have
very accurate knowledge
about valid cues to
deception -- and many of
their beliefs about valid
cues turn out to be
wrong. When Miron
Zuckerman (1981, 1985)
reviewed research on
nonverbal cues to deception,
he discovered that there
were a number of valid cues
on which we could base a
judgment that a person was
However, there were two
important aspects of his
|Some cues are components of the deceiver's verbal behavior.|
|Other cues are paralinguistic -- that have less to do with what the deceiver says than how he or she says it.|
|Still other cues are visual rather than vocal or auditory.|
|And then there are some miscellaneous cues to deception.|
availability of such cues,
people are surprisingly poor
at reading them -- partly
because they're attending to
cues that are, in fact,
findings were obtained in
a review by DePaulo
et al (2003), who examined
than 100 studies and more
than 150 possible cues.
Presumably, though, people could be taught to read these cues properly, in the same way that Ekman's FACS system presumably teaches people to read people's emotional expressions more accurately. If so, the detection of deception from verbal and nonverbal cues, like the reading of facial expressions of emotion, is a perceptual skill that can be acquired through perceptual learning, much as people can learn to adjust to viewing the world through distorting prisms.
This is, in
fact, the premise of a
program, initiated by the
SPOT -- for Screening of
Beginning in 2007, the TSA
spent approximately $200
million per year training
personnel to spot
behavioral cues to
deception in airline
expressions and other
aspects of body
2013 evaluation by the
recommended that the SPOT
program be terminated,
on the grounds that it
large part, the
based its conclusions
of the debate,
'Tell'" by Christopher
suggest that law-enforcement
personnel -- or, at least,
some of them -- tend to have
acquired a particular
perceptual skill of
But Bond and DePaulo's
studies, among others,
suggest that most people
don't have the
knack. And other
critiques suggest that
even the skills of
personnel may be
First, let's get
one thing -- actually, two
closely related things -
detectors don't work very
Lie-Detection and Berkeley
Subsequent legal debates over the validity of polygraphic lie detection resulted in the Frye Rule concerning the admission of scientific evidence in court -- that "expert testimony deduced from a well-recognized scientific principle or discovery" requires that "the thing from which the deduction is made must be sufficiently established to have gained general acceptance in the particular field in which it belongs".
For the whole story, see The Lie Detectors: The History of an American Obsession (2010).
The only physiological technique that is really good at detecting lies is the Guilty Knowledge Test, devised by David T. Lykken, which assesses a suspect's "secret knowledge" of the details of a crime. So, for example, in the case of a stolen watch, the investigator might ask the suspect whether the stolen watch was a Bulova or a Rolex. All things being equal, an innocent person will not respond differentially to the two probes; but a guilty person will know the truth, and this knowledge will show up in his physiological response. Studies have shown that the GKT produces a hit rate of 80-90%, and a false alarm rate of less than 10%. The trick, of course, is that there has to be some aspect of the crime that only the perpetrator would know. For this reason, it is not always possible to use the GKT; but when it's possible, it's dynamite.
Traditionally, the GKT is performed with a traditional polygraph. More recently, EEG and fMRI have been touted for this purpose, but in this case, the neural signature is just another physiological response. People may be more inclined to believe a "neural signature", but it should be clear that the EEG or fMRI is nothing more than a hopped-up polygraph.
So how do law-enforcement personnel determine who is lying to them? To some extent, they rely on nonverbal cues, such as facial expressions and posture, a la Ekman. But new work focuses on what people actually say, rather than how they say it (see "Judging Honesty by Words, not Fidgets" by Benedict Carey, New York Times, 05/12/2009). Problems with false confessions have led police to focus their interrogations on gathering information about the crime and a suspect, instead of forcing a confession.
But even in the
determination of honesty and
dishonesty, there are
linguistic as well as
paralinguistic cues that can
make a difference.
But even these kinds of clues are far from infallible. In one experiment, the rate of lie detection was only about 70%
In the absence
of disciplined perceptual
learning, however, most of
us are pretty bad at
detecting deception -- we
perceive people as lying who
are telling the truth, and
we perceive people as
telling the truth who are in
Person perception is the perception of a person's internal mental states of knowledge and belief, feeling and desire. In addition to making judgments of competence, neuroticism, extraversion, and the like, we also make judgments of other people's sexuality -- both sexual orientation in general, and -- if we're interested -- sexual interest in us. The problem of judging sexual orientation is known colloquially as gaydar -- the idea that people, especially gay people, can intuitively tell whether another person is gay or not.
with a set of studies by
Rule and Ambady (2008; Rule
et al., 2009), a number of
studies have demonstrated
that people can identify, at
better than chance levels, a
target's sexual orientation
based on visual, auditory,
and even olfactory (don't
ask) cues, even when the
stimulus is severely
degraded (e.g., exposures of
only 50 milliseconds in
A representative study is one by Lyons et al. (2014a), in which women, self-identified as straight or lesbian, viewed head-shots of men and women who were self-identified as gay or straight on social media. The study was conducted over the internet, and the subjects were simply asked to classify each target as homosexual or heterosexual. Women were pretty good at this, averaging about 61% hits (i.e., classifying as gay people who really were gay, and straights as straight), and about 27% false alarms, for both male and female targets. Both values differ significantly from the chance level of 50%. Applying signal-detection theory yields substantial values for the d' measure of accuracy; it also revealed a bias toward classifying women as gay, especially by perceivers who themselves were lesbians.
Joshua Tabak and Vivian
Zayas (2012) employed more degraded
materials. They presented
(mostly female) judges with
very brief (50 msec) flashes
of faces of male and female
targets who were
heterosexual or homosexual,
and found that subjects were
accurate in judging the
targets' sexual orientation
about 60% of the time -- as
compared to the 50% accuracy
that would be expected just
by chance. Women were
more accurate than men,
Judgments of women's faces
were more accurate (64%)
than those of men's faces
researchers have not (yet)
uncovered the specific cues
that perceivers use in this
task, Tabak and Zayas found
that judgments were more
accurate when the faces were
presented right-side up, as
Presenting faces upside-down
facial recognition -- what
is known as the face
(Valentine, 1988; Farah et al.,
1995). The face
inversion effect, in turn,
is commonly attributed to
configural processing -- that
is, people recognize
faces not just by
nose, or eyes, or
mouth as individual
features, but rather by
recognizing the length of
the nose relative to the
distance between the eyes --
you get the drift (Maurer et
al., 2002). Anyway,
the superiority of
indicates that it was a
configuration of cues,
rather than individual
features that was the
relevant cue. Research
by Nicholas Rule suggests
that the mouth may be an
Perhaps, Tabak and Zayas
speculate, perceivers judged
"effeminate" male faces and
"masculine" female faces (as
indicated, for example, by
the ratio of width to
height) as more likely to
belong to homosexuals.
But they didn't actually
"gaydar" were taken to a
new level by a study
Yilun Wang and Michal
Kosinski (JPSP, in press
2017) that garnered
articles in The
Economist, the New
Yorker, and the New
Wang and Kosinski
employed 35,000 images
of the faces of white
men and women who had
reported their sexual
orientation on online
(there weren't enough
When they presented
these images to a
of human judges
of the time for
male faces and
about 54% of the
time for female
faces -- barely
chance, and in
line with the
findings of Tabak
the same faces to
for male faces
If the program
was given five
faces for each
the machine: (1)
by virtue of
it was able to
more cue information
than would be
possible for a
in making its
At the same
doing what the
male faces and
"masculine" female faces
to the "gay"
The machine is not
necessarily, a good
model of human
it's likely that
people rely on other
behavior to make
these judgments -- a
man who has an
in musical theater,
perhaps, or a woman
who's really into
these too are
all the way
Moreover, even accuracy of 74-91% shouldn't be overestimated, because of the low base rate of homosexuals in the population. Consider this example taken from an article about the W&K study which discusses other controversies surrounding this study ("Why Stanford researchers Tried to Create a 'Gaydar' Machine" by Heather Murphy, New York Times, 10/10/2017). Assume, for purposes of argument, that 5% of the population is gay. A facial-recognition algorithm that is 91% accurate would mistakenly classify 9% of straight people as gay, and 9% of gay people as straight. In a sample of 1000 individuals, that would mean that 4 or 5 of the 50 homosexuals (1000 x .05) x .09) would be mistakenly classified as straight, while as many as 85 of the 950 ((950 x .05) x .09) heterosexuals would be mistakenly classified as gay. The problem is not so much with the algorithm as with the base rates: with a low-baserate event, like homosexuality, there are going to be a lot of mistaken classifications.
Yes, stereotypes can be accurate, in the sense that they can accurately capture what a group is like on average, even if it's not accurate with respect to all the individual group members. In fact, Lee Jussim (Behavioral & Brain Sciences, 2017) has argued that even racial and gender stereotypes are more accurate than usually believed. I think his evidence is actually pretty weak, but he's right in principle that stereotypes are not necessarily inaccurate representations of groups.
The W&K study was parodied in the New Yorker in "Modern Science", by Paul Rudnick, the American playwright and humorist (12/04/2017). Excerpts follow:
On several occasions, when a photo of an especially attractive subject was scanned, the hardware would disappear from the lab for many hours and then return with a sheen of perspiration and the categorization "YES....
The presence of a single arched eyebrow and a slight contraction of the lips cannot be used as evidence of male homosexuality, except when the subject is examining furniture from West Elm....
The algorithm was able to ascertain sexual preference with 98% accuracy when using only photos of the subjects' shoes....
Photos of male, female, and nonbinary subjects currently attending progressive liberal-arts colleges refused to be categorized as "gay" or "straight," and made disgusted noises.
Facial Recognition and Artificial Intelligence
Ekman's work, and research like W&K's study of "gaydar", signaled a trend toward the use of artificial intelligence and machine learning to create algorithms for facial recognition. In an important Op-Ed article in the New York Times, Sahil Chinoy (a UCB graduate in physics and economics who worked at the Times before going on to graduate school in economics at Harvard), discusses some of the problems with the practice ("The Racist History Behind Facial Recognition", 07/14/2019). See also "Spying on Your Emotions" by John McQuaid, Scientific American 12/2021.
One of these
problems, at least from the point of view of social
policy, is the "perpetual
lineup" problem: if a photograph (from, say,
closed-circuit TV) can be matched against millions of
photographs from a database of driver's licenses, then,
in a sense, we're always under surveillance. We
are very quickly headed toward a surveillance
society in which we're always being watched,
identified, and tracked whenever we're outside the
privacy of our own homes. And even in our homes,
we're already part of a surveillance
economy in which our every Google search and
Facebook like will result in an advertisement appearing
on our computer screens.
Another problem, from the point of view of psychological research and theory, is that the idea of identifying people's internal mental (especially emotional) states from their facial expressions, bodily postures, gestures, and the like may be simply wrongheaded. Ekman's work, and other work like his, has been severely criticized by Lisa Feldman Barrett and other researchers who point out that the correlations between facial expressions and emotion are far from perfect. Chinoy cites a report from the AI Now Institute argues that, based on the current state of both scientific knowledge and computer technology, widely available AI systems for identifying race, sexuality, emotions, and personality traits are "being applied in unethical and irresponsible ways".
In his article, Chinoy traces the current enthusiasm for facial recognition technology back to its roots in the 19th-century pseudosciences of phrenology and physiognomy. In phrenology, people's traits and states are identified by virtue of bumps and depressions in their skulls which ostensibly correspond to high or low levels of benevolence or conscientiousness. In physiognomy, traits are thought to correspond to people's physical appearance -- a person who looks like a fox, for example, was thought to be sly. We laugh at such notions, perhaps, and recognize that they're based on the crudest form of stereotyping. But these ideas have staying power. Sir Francis Galton, who almost single-handedly invented psychometrics in the late 19th century, superimposed pictures of convicts one on the other, hoping that the average would reveal "the essence of the criminal face". And Cesare Lombroso, a 19th-century proponent of physiognomy, argued that intellectual inferiority could be determined from face and body measurements. More recently (2016, to be exact), a group of Chinese researchers employed essentially the same method in an attempt to reveal the "average face" corresponding to criminality.
The ability of people to detect other people's sexual orientation, even with degraded exposure, is impressive. Still, as with Ekman's studies of lie-detection, it should not be exaggerated, because, as with Ekman's famous studies, there is a subtle procedural feature that magnifies the subjects' accuracy levels. Not to pick on it, because it's a perfectly good study as far as it goes, let's take the Lyons study as an example. Like most other signal-detection studies, the "signal" (i.e., a gay target) was "on" for half the trials -- that's just how these studies are done. And when half the targets were gay, the subjects were pretty good -- though far from perfect -- at "detecting" their sexuality. But the problem is that, in the real world outside the laboratory, half the targets aren't gay. A reasonable estimate of the proportion of gays in the population is closer to 5%, and that changes everything.
The reason it changes everything
has to do with Bayes' Theorem, first proposed by Thomas Bayes,
an English clergyman who also dabbled in statistics, in the
18th century (the origin myth is that he was trying to
formulate a statistical proof of the existence of God). The
problem in Bayes' Theorem is to determine the likelihood that
some proposition (A) is true, given some observation or
evidence (B). Bayes argued that in calculating this
probability, you have to take account of the base-rates: (1)
first, the probability that A is true, regardless of B; (2)
and second, the probability that B is true, regardless of
restatement and expansion of Bayes' Theorem, so you
can see how the calculations work out in what
follows. For a nice introduction to Bayes'
Theorem, see the fabulous book by Reid Hastie and
Robyn Dawes, Rational Choice in an Uncertain World
(2001; 2nd Ed., 2010).
In a friendly critique of the
Lyons study, Ploderl (2014) applied Bayes' theorem, which takes
account of base rates, to the calculation of detection
accuracy. Given a base rate of 5%, a hit rate of 70% and a
false-alarm rate of 20% (both figures are reasonably close to
what Lyons found) would yield "gaydar" accuracy of only
15%. Even a more liberal base-rate estimate of 10%
increases gaydar accuracy only to about 22%. That's still
not bad: as Dr. Johnson once said about a dog who could walk on
its hind legs, "It is not done well; but you are surprised to find
it done at all". Still, accuracy of 15-22% is a lot lower
As Lyons et al.
(2014b) pointed out in
reply, Ploderl's analysis
undercuts the ecological
validity of their findings
to some extent, but the
point of the study was
simply to demonstrate that
gaydar can be accurate at
all. This finding then
will motivate future
laboratory research intended
to identify the valid (and
invalid) cues to sexual
orientation. For that
purpose, the problem of baserate
neglect (so named by
Kahneman & Tversky,
1974) isn't really a
problem. These studies
did not identify the precise
visual cues that the
subjects employed to make
their judgments, though
other research has shown
that judgments of sexual
orientation are based
largely on stereotypes: men
who have "feminine"
features, and women who have
"masculine" features, are
more likely to be classified
problem crops up in the
Again In the
current context, the accuracy of
can be reformulated as
So now let's
return to the question of
lie detection, and apply
Bayes' Theorem to that
situation. To begin
with, here's an example from
Tools website developed by Neil
Cotter, a professor of
electrical engineering at
the University of
Utah. In his discussion
he considers a
detects lies with
about 90% accuracy.
That is, the
the lie detector
lied" when you
really did lie
is .89, and the
the machine says
"You told the
truth" when you
really did tell
the truth is
|DL = Detector says "You Lied"
||p(DL | L = .89
|DT = Detector says "You told the Truth"
||p(DT | L = .11
|L = You actually Lied
||p(DL | T) = .10
|T = You actually told the Truth
||p(DT | L) = .90