Spontaneous Inference of Personality Traits and Effects on Memory for
Kristin Stecher & Scott Counts
University of Washington & Microsoft Research
Department of Psychology, 351525, Seattle, WA, 98195
One Microsoft Way, Redmond, WA, 98052
As users navigate online social spaces, they encounter
numerous personal profiles, each displaying a unique
constellation of attributes. How do users make sense of this
information? In our first study, we provide evidence that
users spontaneously make personality trait inferences about
people from profiles they encounter online, and for certain
profiles, preferentially remember this inferred trait content
over actual profile content. Study 2 uses several measures of
profile coherence to assess how the coherence of user
profiles interacts with trait inferences to influence memory
for profiles. Findings provide a better understanding of
specific profile content that makes profiles memorable and
the social-cognitive process utilized when extracting
information from profiles.
We live in the ―Information Age‖ and the sheer amount of
data that bombards us daily can be overwhelming. How
can humans, as information processors, quickly and
efficiently incorporate the important information and weed
out the less essential? Fortunately, humans are particularly
capable of this process when it comes to understanding and
processing information about other human beings. In fact,
this is one area where humans can still ―out-process‖
machines. For instance, humans are experts at face
recognition (Zhao, Chellappa, Phillips & Rosenfeld, 2003)
This is clearly an adaptive characteristic for our species. In
order for us to function in society we need to be able to
recognize whether another member of our society is angry
with us or is welcoming us as a friend. In the domain of
personality, interpreting information about other people’s
personality helps us predict their behaviors in the future.
For example, knowing who is selfish rather than generous
helps us decide who to ask for a favor should we need one.
In recent years, individuals have begun to represent
themselves online and create social networks that include
self-representations in the form of online profiles. Social
networks and online profiles have become critical
components of computer supported social interactions,
serving both social (e.g., dating site profiles) and functional
purposes (e.g., networking for work). Online profiles are a
unique means of self expression for users. Users may spend
a great deal of time creating profiles to convey their
personality to others, but how is that personality
information interpreted by perceivers? Can perceivers who
are bombarded with many sources of information both from
within the network itself and from other competing sources
adequately interpret trait information portrayed in profiles?
Additionally, are profiles
personality traits interpreted or remembered differently than
those that do not?
Traditionally, research surrounding human computer
interaction has provided important insight into cognitive
processes behind computing (Card, Moran & Newell,
1983). Design can be informed by the users’ cognitive
model. As our interactions in these domains become
increasingly social, additional research is needed to
understand users’ social cognitive model in order to begin
to answer these questions. If researchers understand how
users construe social information, software can be better
designed to facilitate social interactions.
that effectively portray
People need very little information to form impressions of
others’ traits. Work by Uleman, Newman & Moskwitz
(1996) demonstrates that perceivers make ―spontaneous
inferences‖ about traits when given a small amount of
behavioral information (Uleman, 1999). In the Uleman
work, study participants are told to study a number of
sentences describing behaviors. Some of the sentences
contain strong trait content. For instance, some subjects
receive ―John wondered where stars come from.‖ which
cues the trait ―curious‖. Later participants are asked to
remember the individuals based on either trait cues or cues
from the sentence. The key finding is that trait cues
(―curious‖) cue memory as well or better than actual
content from the sentence (―stars‖). The rationale is that
people are very good at extracting what information is
important and personality traits often are of high utility. It
may be prove more adaptive for us in the future to
remember that John is a curious person so we can ask him a
question than for us to remember something irrelevant
about stars. In the case of online profiles, this suggests that
when people are presented with profiles, they will
remember personality traits as well or better than the strict
content of the profiles.
It is important to keep in mind that not only do these studies
find that individuals infer personality traits, but they find
that they do so spontaneously. Perceivers need to merely
read about a behavior, and this is sufficient to trigger a trait
inference. According to Uleman (1999) these inferences are
1. They are often below conscious awareness.
2. They are not intentional (not implied by the
3. They are not controllable.
In other words, perceivers infer traits in spite of their
processing goals. This does not mean that controlled
(within conscious awareness and intentional) trait
inferences do not exist, or that automatic (below conscious
awareness and unintentional) processes cannot work in
parallel with controlled processes. However, Uleman’s
research suggests that both controlled and automatic
processes are at work in the impression formation process
and in fact, these processes are activated for different
reasons. It is intuitive to imagine both instances when users
examine profiles using controlled and spontaneous
processes. For example, a casual browser on a blogging site
might make inferences about profiles through spontaneous
processes while an online dater looking for a romantic
partner may use controlled processing. The individual
looking for a romantic partner may know very well what
traits they just inferred, but the casual browser may not
know that they made inferences at all. We suggest that the
casual browser makes more inferences than they think.
Spontaneous inferences are of particular interest to us
because they may guide users’ choices and behavior even
when those users can not elaborate them.
The “How” of Online Profile Processing
Users in computer medicated contexts must make many of
the same decisions as users in face-to-face contexts. They
often encounter many other user profiles and must form
impressions and simultaneously remember information
about these individuals. Does the impression formation
process operate in the same way as in a face-to-face
encounter? Computer Medicated Communication (CMC)
often has fewer cues than face-to-face encounters and
researchers propose two opposing theories that explain how
users will integrate content from these contexts. One theory
suggests that individuals will not encode as much
information about other users, and will disclose less
(Sproull & Kiesler, 1985). Walther (1996) and other
theorists however, suggest that users pick up on minimal
cues and encode whatever information is available.
Communication in CMC is, according to Walter,
―hyperpersonal‖ because the limited availability of cues
causes perceivers to pick up on whatever cues are available.
Users in CMC detect cues that may be overlooked in face-
to-face settings to make inferences about conversation
partners. For instance those in CMC use personal pronouns
as contextual cues as well as pauses and lulls in
These theories allow us to begin to understand how
people view those they meet and interact with online.
Although this research focuses primarily on the linguistic
cues that users parse, it is possible to imagine other social
cues that are utilized. For example, in the current research
we seek to explain the personality trait judgments people
make when they encounter others online and how quickly
this occurs. Thus our first study helps address the ―how‖ of
this fundamental social-cognitive process taking place
online. It also enables us to make suggestions to designers
and users based on social cognitive models.
The “What” of Online Profile Processing
Research shows that personal websites are fairly high
fidelity representations of personality (Vazire & Gosling,
2004). One goal for this work is to start breaking down
online personal representations into attributes that can then
be tied to the conveyance of personality. In that vein,
research has identified several attributes of online profiles
that are important to users in online contexts. For instance,
in an online dating context, Fiore (2002) analyzed 250,000
messages over an 8 month period and identified key profile
attributes that are important to men and women. Men are
interested in physical attractiveness and associated factors,
whereas women are interested in education level and
attractiveness. In online gaming interactions, users were
best matched to other users based on distinct player types
distinguished by their preferences for friendly versus
aggressive play (Schiano, Nardi, Gumbrecht, & Swartz,
2004). Again displaying a distinct preference for particular
profile attributes over others, in a chat room environment,
users prefer to match to similar others rather than others
who have good reputations (Jensen, Davis, & Farnham,
2002). Markus, Machelik & Schütz (2006) suggest that
users can form impressions of others from their websites
and they identify the elements of sites that help craft these
Like these previous efforts, we wanted to make
predictions about how the structure of online profile content
affects processing. With Study 2 we address the ―what‖ of
profile processing by examining the impact of the
coherence of profile content on memory for those profiles.
Study 2 continues to place an emphasis on social cognitive
Using memory as the dependent measure rather than
general ―preference‖, because it is less subjective and
explicit attitudes towards profiles may not be completely
representative (Greenwald & Banaji, 1995).
1. Avoiding domain specificity.
2. Considering interactions between profile structure
and the trait inference process.
Stimuli: Creating Profiles
For our study stimuli we created personal profiles using the
following four step process.
Step 1: Personal Descriptions. Real personal descriptions
were gathered from a popular blog site. Bloggers used these
descriptions as an introduction to their blog in response to
the item ―About Me‖. Descriptions ranged from favorite
movie quotes to descriptions of more broad personality
These profiles were then altered slightly to obfuscate the
identity of the individuals selected. Because the profiles
contained considerable variability in length, we also
standardized the length of the profiles so that each profile
contained between 30-60 words. Initially we selected 55
personal descriptions for inclusion in our pilot study.
Step 2: Add Photos. Photos were obtained using an
informed consent process. A separate group of participants
released their photos for inclusion in the study. Photos
were then matched with an appropriate personal
description. We combined the personal description with the
photo into a brief ―about me‖ profile overview (See Figure
1). These were our stimulus materials throughout the
Step 3: Pilot Study. In order to determine whether and
what personality traits were implied by our stimuli profiles,
we distributed the pilot profiles to 20 participants who were
asked to write down the three characteristics they believed
best defined the person in each profile. In order for the
profile to be included in our study, 50% or more of the
participants had to describe the profile using the same trait
or a close synonym. The trait terms used were the
colloquial terms rather than researcher driven names (i.e.
―geeky‖ not ―intelligent‖). The purpose of this piloting was
to ensure that the stimuli profiles included in the final study
did in fact contain traits that perceivers could extract. This
technique of trait listing is borrowed from previous research
on spontaneous trait inferences (Uleman et al., 1996;
Uleman, 1999). 32 profiles were included in the final
Step 4: Semantic Profiles. After creating trait implying
profiles in steps 1-3, we created ―semantic‖ versions for
each stimulus profile by removing the trait implication. For
instance instead of reading ―I am a typical chemistry major
attending MIT, with aspirations of either becoming a
college professor or becoming a pop star. I am an avid
player of videogames (especially Nintendo).‖, the semantic
version of the above profile reads, ―I am a typical college
student with aspirations of either becoming a teacher or
becoming a pop star. I am an avid player of video games
(especially Nintendo).‖ In this version, the trait implication
of ―geeky‖ is removed (or significantly weakened) by
diluting the implications of his academic affiliations (MIT),
his scientific major and his career aspirations. In all the
semantic profiles, we preserved as much of the content
from the original trait implying profile as possible. The
inclusion of semantic profiles in our study allows us to rule
out the suggestion that our effects are due to the profile
pictures alone. Each picture was assigned to a trait profile
and a matched semantic profile, allowing us to compare
memory for trait and semantic profiles.
Note that for our research purposes we created a
somewhat simplified profile, containing a photo and
personal statement. We felt that these abbreviated profiles
made our research tractable while maintaining sufficient
realism in that users often make choices based on these
types of profiles, such as whether this is a person who’s
blog they would like to read.
1. Individuals make personality trait inferences when
viewing online profiles.
2. These inferences can be “spontaneous”.
Demographics. Our first study included 31 participants
who were recruited via email. None of the participants
were known by the experimenters. Twenty-three male
participants and eight female participants took part in our
study. Twenty-two participants were Caucasian, 5 were
Asian, 3 were African-American and one was “other”.
Participants ranged in age from 19 to 49 years with an
average age of 33.5.
Participants reported having experience with social
networking sites. On average participants had 2.5 years of
experience and spent an average of 2.19 hours a week on
these sites. In addition, participants posted an average of
3.73 profiles in their lifetime.
Method. Participants were told that they would view a
number of online profiles that were ―About Me‖ sections
from a popular blogging site and also warned that some
details had been changed to protect the identities of the
individuals portrayed. After consenting to participate,
participants were given very general instructions to ―form
an impression‖ about the profiles they were to view. This
kept the experience as realistic as possible for participants
while not instructing them to remember any specific
information, thus allowing for spontaneous processing.
Using methods from social psychology, we adapted our
profiles to a cued-recall technique that is widely accepted
for detecting spontaneous inferences (Uleman, 1999). This
procedure utilizes Tulving’s encoding specificity principle
(Tulving, 1972; Uleman et al., 1996). If secondary
information is present while primary information is being
encoded, secondary information will serve as a cue for the
Figure 1. Sample Trait Profile. Implies “geeky”.
primary information. Therefore, if people make a
personality trait inference when they read a profile, then
traits should serve as memory cues to help remember those
profiles (Uleman et al., 1996). If this trait inference is
especially strong, then traits will be as good of a cue or
better as content from the profile. Trait cues will not help if
this personality trait was not present in the profile. This
procedure was administered using two study–recall phases.
Study Phase 1. During the first study phase participants
were asked to view 16 profiles. Profiles were presented on
screen for 10 seconds each. Participants had to press
―continue‖ after each profile to move on with the
experiment. This was to ensure each profile was viewed.
The order of the 16 profiles was randomized. Eight trait
profiles and eight semantic profiles were presented.
Recall Phase 1. In the recall phase participants are
presented with either trait cues or semantic cues in the form
of words shown on screen and asked to use these as
prompts to recall the profile information they read about
earlier. Semantic cues were words (e.g., ―Nintendo‖) that
were actually contained in the profile while trait cues were
trait words (e.g., ―geeky‖) that were not contained explicitly
in the profile but were cued in trait profiles. The trait cue
words were those provided by our pilot study participants
who initially evaluated the profiles for the presence of a
trait, which ensured that the trait cue words were well-
matched to the trait profile stimuli. Subjects were given a
blank space below the cue to provide as much of the
personal description as they could remember based on the
cue. After one practice trial, participants completed this
cued recall procedure for each of the 16 profiles shown in
the study phase of the experiment.
We randomized the stimuli but ensured there were the
same numbers of each type per cell to control for order
effects (Table 1). Each subject received eight of each
Study Phase 2- Recall Phase 2. The procedure described
above was repeated with 16 more profiles so that
participants saw a total of 32 profiles in two sets of study-
recall phases. The study was broken into two parts because
pilot testing demonstrated that participants did not have the
capacity to remember sufficient information from 32
Measures. As noted, this was a cued-recall task. Our
dependent measure was recall for the profile content when
cued. Raters who were blind to the participants’ condition
coded the recall on a scale of 0-3 (no recall- complete
recall). No recall meant that when cued, the participant was
unable to type in any amount of the profile ―About Me‖
statement. Complete recall meant that the participant
reproduced the entire ―About Me‖ statement when cued.
Most frequently, participants recalled some part of the
―About Me‖ statement.
Note that some words cued a profile that was not the
―intended‖ profile. For instance, when presented with the
trait word ―geeky‖, individuals may have recalled a
different profile than the profile the experimenters had in
mind. Because many profiles did imply similar traits, we
felt it necessary to account for this and included these
scores in our analysis. Less than 5% of the words recalled
fell into this category. Additionally, semantic and even trait
profiles may have cued other traits (e.g., the ―geeky‖ profile
in Figure 1 may have cued the trait ―achiever‖). However
because these traits were not provided as cues, we assume
this did not affect recall scores.
We were interested in what type of information helped cue
recall for participants. Results from our 2 x 2 ANOVA
show that there was no main effect of word, F(1, 30)=0.01,
p<0.987. There was however, a significant main effect of
profile such that trait profiles were preferentially recalled,
F(1, 31)=21.4, p<0.001. The trait word- trait profile pairing
was of particular interest to us. Simple effects of profile for
demonstrated a simple effect of profile for trait words, F(1,
31)=51.8, p<0.001. There was no simple effect of profile
when the cue was a semantic word, F(1, 31)=1.01, p=0.322.
Trait profiles were remembered better when the cue was a
trait word (M=1.02) but not when the cue was a semantic
word (M=0.52). Finally, there was an overall profile x
word interaction, F(1, 31)=35.3, p<0.001. This interaction
is presented in Figure 2.
These findings demonstrate that traits serve as better
recall cues than semantic words but only when profiles
imply relevant traits. Therefore, it seems that participants
remember the overall trait content of the profile better than
the actual semantic content of the profiles they are
presented with. They cannot however, extract trait content
if that trait is not implied.
Anecdotal Evidence for Spontaneous Trait Inferences.
It is also interesting to note that participants responding to
semantic cue words often spontaneously generated trait
information during recall. For instance, one subject recalls
―the seemingly self-centered girl who also liked Louis
Vuitton and diamonds‖, when prompted with the cue
―perfumes‖. Although the participant was simply asked to
recall the profile, the participant felt it diagnostic to note
trait information (―self-centered‖) in addition to actual
profile content. Another recalls, ―sports jock that will take
mother out to a seafood dinner‖ simply because the target
was wearing a track uniform although the target indicated
nothing about sports in his profile. Although we did not
Table 1. Presentation Matrix, Study 1. Numbers
represent profiles shown in each condition
Trait Cue 8 8
Semantic Cue 8 8
Figure 2: Recall based on condition. Memory ranges
from 0 = no recall of profile content to 3 = complete
recall of profile content. Error Bars= S.E.
Trait Word Semantic
Recall for All Profiles
systematically analyze this content, it seems that
dispositional processing of information is implicit and, as
we hypothesize, spontaneous.
Discussion, Study 1
We designed our first study to examine whether:
Individuals better remember people in online profiles using
trait inferences than from the actual content of the profiles.
These trait inferences can be ―spontaneous‖.
We demonstrated that, like people who meet others in
real world interactions, users in online communities are
skilled at extracting important trait information from messy
data. In fact, when users view profiles they may remember
information about personality traits of other users more than
the actual content of the profiles they view. However, this is
only true if these profiles imply this trait.
Not only do perceivers extract traits from online profiles,
but they also do so spontaneously. Without any explicit
instructions or processing goals, they still made inferences
about users’ personality traits. By simply reading a series of
brief profiles containing varying content, perceivers
remembered personality traits they inferred about other
users. Trait information was remembered preferentially to
Although absolute differences between conditions may
be small they were both statistically significant and
meaningful. Our overall numbers are low because of the
difficulty of using free recall rather than recognition to
identify other users. The recall task was challenging. Also,
trait profiles are remembered 30% more than semantic
profiles, and trait words trigger recall 20% more than
semantic words that were actually contained in the profile.
This finding led us to hypothesize that there may be
factors about the target profile that affect memory. We
therefore devised a second study to examine these factors.
Study 1 demonstrated that users make spontaneous trait
inferences when profiles clearly imply traits. Therefore, we
see a difference between profiles crafted to imply a clear
personality and those not designed to do so. Are there other
factors affecting memory for traits? We used Study 1 as a
building block to identify factors of profiles that might be
related to recall. Since trait profiles in Study 1 cohered on a
common trait, perhaps coherence in general is related to
recall for profiles. Coherent profiles may allow participants
to form more structured impressions, and lead to better
memory. Again, we are interested in identifying models
that are easiest to process for users. Coherency may be one
dimension that makes social information easier to process.
Hypothesis. There is a positive relationship between profile
coherence and overall recall.
We assessed coherence using three measures:
1. Overall Coherence: How well do profile elements
2. Number of Elements: How many particular
elements does the profile contain?
3. Specificity: How specific is each particular
Overall Coherence. Pilot study participants were asked to
rate the coherence of the 64 stimuli profiles (32 trait and 32
semantic). Profiles were broken into 3 segments. Three
participants were told to compare each part of the profile to
the other 2 parts and rate how well the parts ―went
together‖ using a 1-7 Likert scale. Interrater reliability was
good (α=0.97). Higher coherence scores are hypothesized to
be associated with higher recall.
Number of Elements. We asked a second group of pilot
participants to divide each profile into its constituent
elements. Nine participants were presented with all 64
profiles (trait and semantic). Participants were asked to
divide each profile into elements that were psychologically
meaningful for them by simply recording and labeling these
separate parts. Participants’ breakdowns ranged from 1-10
elements. Interrater reliability was good (α=0.81). For
example, the profile in Figure 1: ―I am a typical chemistry
major attending MIT, with aspirations of either becoming a
college professor or becoming a pop star. I am an avid
player of videogames, (especially Nintendo).‖ might be
broken down as, ―1) I am a typical chemistry major
attending MIT 2) with aspirations of either becoming a
college professor 3)or a pop star. 4) I am an avid player of
videogames, (especially Nintendo).‖
participant felt that these 4 items roughly cohered
psychologically, they broke the profile into 4 elements. We
expect that profiles with fewer items are associated with
greater coherence and will be remembered better.
Specificity: A third group of three pilot study participants
were assigned to rate the specificity of the profiles. Profiles
were again broken into 3 parts. Participants rated the
specificity of each part of the profile from 1-7 on a Likert
scale. The final specificity score for each profile was a sum
of the three ratings. Again interrater reliability was good
(α=0.93). We expect that profiles with higher specificity
scores will be remembered better.
For purposes of illustration in Table 2 we included text
from the profiles with the highest and lowest scores in each
For each coherence measure we calculated the correlation
between coherence (e.g., low to high specificity of profile
elements) of the profiles with the recall scores for those
profiles we collected in study 1. Correlations for trait and
semantic profiles are also examined separately for each
measure of coherence.
Overall Coherence. We found the predicted positive
correlation between overall recall and overall coherence
(r=0.19). However, since this correlation was not as strong
as expected, we looked at correlations for trait and semantic
profiles separately and saw this positive relationship was
stronger for trait profiles (r=0.22) and more modest
(although not negative) for semantic profiles (r=0.12).
Number of Elements. When we examined the correlation
between the number of elements in the profile and overall
recall for that profile, the predicted negative correlation did
not emerge (r=0.042). However, once the data were
separated into trait and semantic profiles, we see that the
presence of a trait mediates this effect. For profiles that
implied traits, as was hypothesized, the number of items in
the profile was negatively related to memory for the profile
(r=-0.28). However for profiles that did not imply traits, the
number of items in the profile was positively correlated to
memory for the profile (r=0.33). See Figure 3 for
scatterplots of the relationship between the number of
elements and recall for both trait and semantic profiles.
Specificity. There was no positive relationship between
specificity and recall for profiles (r=-0.09). Within the
separate profile conditions, effects are also nonsignificant
(r=-0.159) for trait profiles, and (r=-0.04) for semantic
Intercorrelation. As expected these items are related but
not the same. Specificity and the number of items are
significantly correlated, r=-0.41, p<0.001. More specific
profiles have fewer items. However, the overall coherency
was not related to the number of items, r=0.14, p=0.14 and
the overall coherency and the specificity were not related,
Individual Differences. Although we did not see the
predicted effects of either trait or semantic profiles when
looking at the relationship between specificity and recall,
we hypothesized that there are factors other than the
presence of a trait that mediate this relationship. People
differ in their responses and interests. Previously we
averaged across all participants, which may have caused us
to ignore effects at the level of the individual. We looked at
correlations between the average recall scores for each
profile established in Study 1 and each individual
participant’s responses to our three coherence measures.
This enabled us to determine if there were cognitive
mediators at the level of the person rather than the level of
the profile that drive the relationship between specificity
and recall. If some participants exhibited strongly positive
and others strongly negative correlations between recall and
specificity, we hypothesize that these mediators exist and
deserve more attention.
Table 2. Sample Profile Text: Highest and lowest
rated profiles for each category
attend class at
UCSD, im a junior
here. I really like
watching tv. I’m a
Umm, well here’s
helloo, =] I adore
couture, jazz, louis
Gucci, gold &
loads more. I think
my xanga site is
hot & appreciate it
if you read, u blog
& post some
I will smash your
face into a car
then take your
out to a nice
seafood dinner and
NEVER call her
i have id say about 5
friends but then again I
think I don’t need a
lot. i think if u have a
cuple TRUE friends
then ur fine!! I also
have glasses and
braces! I act different
to other ppl sumtimes.
The thing you need to
know is that 88% of
sites suck… that may
be yours also
thankfully, I balance
the world, so yay me
you inebriated simian
Im your typical girl,
that goes thorough the
normal ups and downs
in life... but i like to
add a romantic twist
some times....I like to
try new things...that
sometimes get me into
trouble...but it makes
my life so much more
fun and interesting!
Correlations ranged from r=-0.49 to r=0.69 for trait
profiles and r=-0.33 to r=0.58 for semantic profiles. For
illustrative purposes, Figure 4 displays sample participants’
correlations between recall and specificity for both trait and
semantic profiles. Individuals exhibit high correlations
between specificity and recall in both the negative and
Since individuals exhibited both significant positive and
negative correlations between specificity and recall within
trait and semantic conditions, we hypothesized that there
were also important effects for the overall coherence
measure and number of elements measure. The overall
coherence measure yielded diverse and strong correlations
in both the positive and negative direction ranging from r=-
0.58 to r=0.55 for trait profiles and r=-0.56 and r=0.64 for
semantic profiles. Additionally, when we looked at the
relationship between recall for each individual participant
and the number of elements in each profile, participants
displayed strong correlations in both the negative and
positive direction, from r=-0.45 to r=0.47 and r=-0.65 to
r=0.59 for trait and semantic profiles respectively. A graph
representing the highest and lowest correlations for all
measures of coherence is displayed in Figure 5. The
average overall correlation for semantic and trait profiles on
that dimension is also provided for reference.
Discussion: Study 2
Study 2, like Study 1 emphasizes the importance of
unifying information within a profile. In Study 1, when
traits were present, they served as powerful cues for the
observer. In Study 2, certain measures of coherency
provided evidence that perceivers remembered coherent
profiles better. Profiles that are internally consistent (all the
elements relate to each other) are most apt to be recalled.
Participants also remember profiles that had fewer elements
(another proposed measure of coherence), although this was
only true when a trait was present suggesting that traits may
serve as a mediating variable in some situations. We also
discovered that each user is unique and will not react to the
profile attributes with the same pattern of response. For
instance, highly specific profiles may be very memorable to
some but not contain the features that another user is
looking for at all.
Figure 3: Number of Elements x Recall for Trait and
Number of Elements
Number of Elements-Trait Profile
Number of Elements
Number of Elements- Semantic
Figure 4: Individual Response Patterns for
Specificity within Each Profile Condition
Specificity x Trait Profile
0 10 20
Specificity x Semantic Profile
Figure 5. Correlations for each level of analysis and
Hi Individual Overall Low Individual
Correlations-All Levels of Analysis
Specific results for each coherency measure are
Overall Coherence. Our overall coherence measure asked
participants to relate intra-profile coherence. To the extent
that the items within the profile matched well together, the
profile was remembered better. This was particularly true
when a trait was present.
Number of Attributes. When participants broke profiles
into their constituent elements, the presence of a trait served
as a mediator for the relationship between recall and this
measure of coherence. To improve recall, profiles that
contain traits should be condensed into a few trait implying
elements. Conversely, for those that do not imply a trait,
more elements seem to lead to a more memorable profile.
Specificity. Our findings for the specificity dimension were
not as hypothesized: there was no overall relationship
between profile specificity and memory. However, once we
broke the profiles down and examined individual patterns
of response, we determined that there were in fact,
correlations between specificity and memory, but because
they ranged in both directions, averaging across all
participants did not account for individual responses to
specific profiles. Additional research is needed to
determine the individual difference factors that drive the
relationship between specificity and recall.
As people become increasingly social in online domains,
we are able to study the social cognitive aspects of their
complex interactions. From these studies, we propose that
personality traits can be inferred spontaneously from online
profiles and that they are extracted preferentially to other
content (Study 1). We also identified information in
addition to trait content that allows users to process profiles
more effectively (Study 2). As a general trend, coherency
facilitates memory for profiles, implying that it allows for
more efficient processing of social information. This is
especially true when trait information is present in user
profiles. Finally, we argue that each user is unique and it is
important to attend to their unique patterns of responding.
Knowing how users make inferences in social networks
and computer mediated contexts has important theoretical
and practical applications. Users made trait inferences from
online profiles. Additionally, they formed these
impressions on the basis of little information and without
prior instruction sets. These studies suggest that users
engage in hyperpersonal communiation online and user
behavior in online domains replicates user behavior in
offline domains (Uleman, 1999; Uleman et al., 1996).
Impression formation online is a natural and automatic
process. Users will draw inferences about personalities of
the other people they encounter.
These studies also suggest that users process coherent
profile information more readily. Specifically, if profile
elements fit together, this aids in profile memory. If
profiles cue a trait then it is especially important for them to
be coherent. Simultaneously, it is important to keep in
mind that, despite general trends, people process social
information in different ways (e.g., memory for profiles is
facilitated by highly specific profile content for some users
but not for others). Our individual difference findings from
Study 2 highlight the importance of customization for the
individual user. We hope to use this study as a launch pad
to better categorize individual patterns of responding into
clear focus groups. For instance, there may be individuals
who prefer a certain attribute such as specificity and not
another. How can we a priori tag these individuals as part of
this preference group?
There are a number of services, typically aimed at dating
sites, that offer tips for the look and feel of the profile, but it
is reasonable to imagine many types of users wish to be
remembered. For instance, there are bloggers who want
people to return to their site to hear their views again. With
these studies, we show that memory for a profile has to do
with more than just look and feel, and again we point to the
importance of the promise to boost recall for their profiles
by 30% by their inclusion of a trait.
How then can these findings help designers facilitate
interpersonal interactions online and what suggestions can
we provide to users creating online profiles based on this
research? Personality trait inferences are natural in online
domains but they are facilitated only if a trait is implied in
the online profile. Therefore we suggest that for any
context where users want to be remembered, users and
designers create profiles and profile environments where
trait implications are natural and encouraged. For instance,
the blogging site that first displays the ―About Me‖ section
is organized for better memory than the site that first
displays demographic information that has less probability
of cueing a trait. In addition to organizational
enhancements, services could promote memory for profiles
by providing instruction sets that help users convey a trait,
even if this is as simple as telling users to do so explicitly.
As a somewhat more sophisticated alternative, a site
could solicit trait tags for profiles from other users as a way
to check whether profiles are in fact conveying a trait, and
if so, what trait they are conveying. Users could rate one
another on personality dimensions and adjust their profile
content based on their ratings. Finally, it may also be
possible parse the language used in profiles to identify the
strength of the traits conveyed using natural language
processing programs Natural language processing programs
have already been used in ecommerce to acquire implicit
and explicit user data such as mood, values (implicit) and
product references (explicit) from user postings and emails
(Paik, Sibel, Brown, Poulin, Dubon, & Amice, 2001). These
programs could be applied to mine both implicit and
explicit references to traits in user profiles.
In sum, we believe that people are social processors of
information and it is this social information that drives their
behavior. People demonstrate preferences for trait
information because they are trying to make sense of
complicated the social networks that surround them. Their
ability to quickly and efficiently extract traits is a skill that
is most essentially human and, as this study demonstrates,
carries over fluently to online interactions.
Ambady, N. & Rosenthal, R. (1993). Half a minute:
Predicting teacher evaluations from thin slices of nonverbal
behavior and physical attractiveness. Journal of Personality
and Social Psychology, 64, 431-441.
Card, S., Moran, T. & Newell, A. (1983). The Psychology
of Human-Computer Interaction. Hillsdale, NJ: Erlbaum.
Fiore, A. Romantic Regressions: An Analysis of Behavior
on Online Dating Systems: Program in Media and Cornell
(2002) Masters Thesis.
Greenwald, A.G & Banaji, M.R. (1995). Implicit social
cognition: Attitudes, self-esteem,
Psychological Review, 102(1), 4-27.
Jensen, C., Davis, J. P., & Farnham, S. D., (2002). Finding
Others Online: Reputation Systems for Social Online
Spaces CHI 2002, ACM Press (2002), 447-454.
Marcus, B., Machilek, F., & Schütz, A. (2006). Personality
in cyberspace: Personal websites as media for personality
expressions and impressions. Journal of Personality and
Social Psychology, 90, 1014-1031.
Paik, W., Sibel, Y., Brown, E., Poulin, M., Dubon, S. &
Amice, C. (2001). Applying natural language processing
(NLP) based metadata extraction to automatically acquire
user preferences. Proceedings of the First International
Conference on Knowledge Capture. Victoria, British
Riegelsberger, J., Counts, S., Farnham, S., & Phillips, B.
(2007). Personality Matters: Incorporating Detailed User
Attributes and Preferences into the Matchmaking Process.
In Proc. HICSS 2007.
Schiano, D.J., Nardi, B. A., Gumbrecht, M., Swartz, L.
(2004). Blogging by the Rest of Us. CHI 2004, ACM Press
Sproull L, Kiesler S. 1985. Reducing social context cues:
electronic mail in organizational communication. Manag.
Tulving E. (1972). Episodic and semantic memory. In E.
Tulving & W. Donaldson (Eds.), Organization of memory
(pp. 381-403). New York: Academic Press.
Uleman, J. S. (1999) Spontaneous versus intentional
inferences in impression formation. In S. Chaiken & Y.
Trope (Eds.), Dual process theories in social psychology.
New York: Guilford Press.
Uleman, J. S., Newman, L.S., & Moskwitz, G.B. (1996)
People as flexible interpreters: evidence and issues from
spontaneous trait inference. In M.P. Zanna (Ed.), Advances
in Experimental Social Psychology, 28, 211-279.
Walther, J. B. (1996). Computer-mediated communication:
Impersonal, interpersonal, and hyperpersonal interaction.
Communication Research, 23, 3-43.
Vazire, S. & Gosling, S. D. (2004). e-perceptions:
Personality impressions based on personal websites.
Journal of Personality and Social Psychology, 87, 123-132.
Zhao, W., Chellappa, R., Phillips, P. J., Rosenfeld, A.
(2003). Face recognition: A literature survey. ACM
Computing Surveys, ACM Press (2003), 399- 458.