The Expression of Personality in Virtual Worlds

To examine the expression of personality in virtual worlds (VWs), we tracked the behavioral and linguistic output of 76 students continuously over a six-week period in the VW Second Life (SL). Behavioral metrics in SL were consistent over time, but low stabilities were observed for linguistic metrics. To examine the ways in which personality manifested in SL, participant's Big Five scores were correlated with their virtual behavioral and linguistic metrics. For example, Conscientiousness was correlated with many metrics related to geographical movement, however, there was low overlap with findings from previous studies. We provide some reasons for this low concordance. Our study hints at the potential of leveraging VWs to understand not only the link between personality and behavior, but among other social and psychological phenomena as well. The Expression of Personality in Virtual Worlds Every morning, at exactly 7 o'clock, Stella treks to her farm to harvest and plant a new crop of peas, but in the popular FaceBook harvesting game FarmVille, she can do this without even breaking a sweat. And over in the virtual world Second Life where users create all of the world's content, Marcus is wondering whether his new mohawk hairdo would send the wrong message at the academic panel he is attending or whether it would be considered stylish in the context of Second Life. As virtual worlds (VWs) become mainstream, a critical psychological issue is whether and how personality manifests itself in VWs. The Expression of Personality Research in person perception has consistently shown that judgments of personality at zero acquaintance hold some degree of validity. This has been shown to be true for face-to-face encounters (Kenny, Horner, Kashy, & Chu, 1992), as well as judgments based solely on observations of an individual's bedroom or office (Gosling, Ko, Mannarelli, & Morris, 2002) or their music preferences (Rentfrow & Gosling, 2006). Similar research has also extended to computer-mediated communication (CMC). In particular, past findings have shown that somewhat accurate personality impressions can be formed based on an individual's personal website (Marcus, Machilek, & Schutz, 2006; Vazire & Gosling, 2004), FaceBook profile (Back, et al., 2010), email content (Gill, Oberlander, & Austin, 2006), and even an individual's email address (Back, Schukle, & Egloff, 2008). In exploring different methods of studying personality manifestation, some researchers (Mehl, Gosling, & Pennebaker, 2006) have illustrated the value of using natural observations to study personality as it manifests in everyday life. Collecting observations of natural behavior, however, is a daunting task where the tedium of continuous observations has only recently been offset by modern technology. Unobtrusive Observations in Virtual Environments Virtual Worlds provide unique affordances for studying the link between personality and behavior. For the purposes of this paper, we define VWs as graphical environments that enable geographically-distant individuals to interact via graphical avatars (i.e., digital representations of users). These environments are no longer academic prototypes, but have become mainstream interaction platforms. For example, the online game World of Warcraft has over 11 million active subscribers world-wide (White, 2008). Second Life is another example of a VW, and is unique in that users in Second Life create almost all the content (i.e., buildings, cars, dresses, hair styles, dance animations) in the world using scripting and modeling tools. This is in contrast to most online games where players can only use and interact with the content created by game developers. There are three unique affordances of VWs with regard to natural observations of behavior. First, VWs are already instrumented with high-precision sensors. The computer systems running VWs already track the movement and behavior of every avatar to make interaction possible. Second, these high-precision sensors can track behaviors continuously and longitudinally. And finally, all these observations can be performed unobtrusively, thereby eliminating the observer effect (Webb, Campbell, Schwartz, & Sechrest, 1966). Stability of Behavioral and Linguistic Metrics in Virtual Worlds Given that almost all studies of person perception in online environments have employed static or snapshot observations, such as personal websites (Marcus, et al., 2006), the stability of behavior and linguistic output in digital environments and its relation to personality manifestation is understudied. Important clues can be inferred from other studies however. For example, written language use in diaries, class assignments, and professional journal abstracts has been shown to be quite reliable over weeks and even years (Pennebaker & King, 1999). And in a study of everyday conversational language collected via automated voice recorders (Mehl & Pennebaker, 2003) across two-day periods separated by four weeks, linguistic measures were also observed to be highly consistent. Thus, these two studies show that both spoken and written language appears to be quite stable over time in physical contexts. In one of the few longitudinal studies of behavior in virtual environments, it was found that behavioral changes over time do occur as users acclimate to interacting via digital avatars, but that these changes occurred across all users (Bailenson & Yee, 2006); however, personality factors were not explored in that study. If we assume that this pattern generalizes to most VWs, then it implies that rank-order stability may be high even though absolute stability is low--i.e., over time, VW users may all explore less, but some users will always tend to explore more than other users. Consequently we seek to examine the stabilities of behavioral and linguistic output in VWs. RQ1a: Are behavioral and linguistic measures in virtual worlds stable over time in absolute terms? RQ1b: Are behavioral and linguistic measures in virtual worlds stable over time in relative terms? Manifestation of Personality in Virtual Worlds Research in person perception has documented the ways personality manifests itself in a wide variety of environments. Here, we will first consider past studies in behavioral correlates and then linguistic correlates of personality. In an early study of how personality manifested in normal face-to-face conversations (Funder & Sneed, 1993), coders rated participants in a social interaction using 64 behavioral categories. Acquaintances of participants then rated each participant using a Big5 personality inventory. Many significant correlations were observed between the personality ratings and the coded behaviors, most of which aligned closely with trait definitions of the personality factors. For example, Extraverted individuals spoke louder, with more enthusiasm and energy, and were more expressive. Agreeable individuals expressed sympathy, seemed to enjoy the interaction with their partner, and expressed interest in what their partner said. In another study, researchers explored the manifestation of personality in personal spaces (Gosling, et al., 2002). Thus, instead of observing behaviors directly, personal spaces hold identity claims (e.g., a poster of Nietzsche) and behavioral residues (e.g., a withered house plant) that reflect personality more indirectly. Nevertheless, researchers found significant correlations between coded personal space attributes and self-report personality ratings of individuals. In their study of bedrooms, it was found, for example, that individuals who scored high on Openness to Experience had more varied books and magazines. As another example, Conscientious individuals had more well-lit, neat, and well-organized bedrooms. These two studies were selected to illustrate how direct and indirect behavioral correlates of personality have been observed in the past. It is unclear, however, how these might translate into virtual worlds. For example, while individuals are able to interact in virtual worlds, many virtual worlds do not have user-controlled facial or hand gestures. And it is unclear how varied book collections translate into virtual worlds where people don't read virtual books. On the other hand, as we mentioned above, there are a plethora of behavioral metrics that virtual worlds provide, such as geographical movement that may nevertheless be significant personality cues. Unlike behavioral correlates, linguistic correlates of personality may translate more directly into virtual worlds. To provide an overview of findings in this area, we describe four studies that have all used the Linguistic Word Count and Inquiry software (LIWC; Pennebaker, Booth, & Francis, 2007) to examine linguistic correlates of personality. LIWC is a dictionarybased word count tool that counts the ratio of words in 70 linguistic categories. For example, the category "positive emotion" contains the words: happy, cheerful, joy, etc. The four selected studies span the past decade and examine different linguistic content: personal writing profiles (Pennebaker & King, 1999), self-narratives (Hirsh & Peterson, 2009), everyday conversations (Mehl, et al., 2006), and blog content (Yarkoni, in press). Table 1 A summary of previous linguistic correlates with personality factors Personality Factor Positive Correlates Negative Correlates Emotional Stability Articles Anger, Anxiety, Negative Emotions, First Person Singular Extraversion Social Processes, Positive Emotions Openness to Experience Articles, Exclusives First Person Singular, Present Tense, Past Tense, Social Processes Agreeableness First Person Singular, Inclusives, Family, Positive Emotions Articles, Anger, Negative Emotions Conscientiousness Achievement Causation, Exclusives, Anger, Negations, Negative Emotions To provide a concise and coherent summary of the findings (oftentimes with hundreds of comparisons) without being bogged down by idiosyncratic differences, we will summarize only correlates that were found to be significant in at least two studies (see Table 1). Of note, most linguistic correlates mirror trait definitions. For example, Emotional Stability is negatively correlated with negative emotions, and Agreeableness is positively correlated with social involvement and positive emotions. On the other hand, grammatical features such as articles and first person singulars also made frequent appearances (both were significant correlates in three out of the five personality factors) although their connection with personality is less obvious. Thus, in the present study, we were interested in the following research question: RQ2: What behavioral and linguistic correlates of personality in virtual worlds? Method