The boundary between real and fictional others in the medial prefrontal cortex is blurred in lonelier individuals

Abstract

People spend much of their free time engaging with narrative fiction. Research shows that, like real-life friends, fictional characters can sometimes influence individuals’ attitudes, behaviors, and self-beliefs. Moreover, for certain individuals, fictional characters can stand in for real-life friends by providing the experience of belonging. Despite these parallels between how people think of real and fictional others, it is unclear whether, and to what degree, their neural representations are similar. Does the brain treat psychologically close fictional others as it does close real-world friends, or are real others somehow privileged in their neural representation? In the present study, fans of the HBO series Game of Thrones performed a trait-evaluation task for the self, 9 real-life friends/acquaintances, and 9 fictional characters from Game of Thrones while undergoing functional magnetic resonance imaging. Using both brain decoding and representational similarity analysis, we found evidence of a categorical boundary between real and fictional others within the medial prefrontal cortex. However, the boundary between these categories was blurred in lonelier individuals. These results suggest that lonelier individuals may turn to fictional characters to meet belongingness needs, and this, in turn, alters the manner in which these categories are encoded within the social brain.

fictional characters, fMRI, narrative, loneliness, representational similarity analysis

That is part of the beauty of all literature. You discover that your longings are universal longings, that you’re not lonely and isolated from anyone. You belong.

—F. Scott Fitzgerald

During the height of the COVID-19 pandemic, opportunities for socializing with friends and acquaintances dwindled, and there is reason to believe that many people may have turned to fictional characters to fill this gap. In 2020, people in the United States reported spending an extra 30 min a day watching TV or reading for personal interest compared with the previous year (Bureau of Labor Statistics 2021). A similar increase was observed in the United Kingdom where adults spent over an hour a day watching streaming services during lockdown, a 71% increase over 2019 (Rajan 2020). This trend was also evident at the global level as indicated by subscriptions to streaming services surpassing 1 billion during the pandemic, a milestone spurred by a 26% increase in subscriptions over the previous year (Watson 2021). At the same time that an increase in engagement with fictional narratives appears to have been accelerated by the COVID-19 pandemic, so too is there evidence of a worsening loneliness epidemic (Cacioppo and Cacioppo 2018; Cigna 2020), especially among young adults (Lee et al. 2020; Weissbourd et al. 2021). Past research indicates that individuals are especially likely to seek a sense of social connection from mediated experiences when they are feeling socially rejected or lonely (Perse and Rubin 1990; Derrick et al. 2009), suggesting that these 2 recent trends may be related. With people seemingly spending more time than ever in the company of fictional characters, coupled with an increase in self-reported loneliness, it is worth considering just how fictional friends compare with real-life ones, especially among those who feel socially isolated.

Many people form emotional attachments to fictional characters just as they do their real-life friends (Cohen 2003; Brown 2015), and their beliefs, behaviors, and emotions are influenced by those characters in much the same way as they are by their friends (Cohen 2003; Gardner and Knowles 2008; Igartua 2010; Sestir and Green 2010; Kaufman and Libby 2012; Djikic and Oatley 2014; Shedlosky-Shoemaker et al. 2014; Slater et al. 2014). Most striking of all, perhaps, is evidence that fictional characters can provide the experience of belonging, in essence standing in for a real-life friend by making someone feel connected and less alone (Derrick et al. 2009; Gabriel et al. 2016). What, then, are people doing when they engage with narrative fiction, and how does it overlap with how they engage with the flesh-and-blood members of their actual social networks?

One prominent theory concerning the psychology of narrative fiction argues that it functions as an immersive simulation of social experiences, allowing one to exercise her social cognitive abilities much like a flight simulator helps prepare a pilot for actual flight (Mar and Oatley 2008; Mar 2018). Many of the same social cognitive processes used to understand real-world others are drawn on when engaging with narrative fiction (Zunshine 2006), and there is some evidence that experience with narrative fiction can translate to benefits for real-world social cognitive ability (Mar et al. 2006; Kidd and Castano 2013; Panero et al. 2016; Mumper and Gerrig 2017; Dodell-Feder and Tamir 2018). Consistent with this idea, meta-analyses demonstrate an overlap between brain regions implicated in theory of mind and story comprehension (Mar 2011), including the medial prefrontal cortex (MPFC), posterior superior temporal sulcus, and temporoparietal junction. Furthermore, some theorists have argued that because film and television are such new, sensory-rich technologies, humans’ neural architecture is ill-equipped to differentiate fictional people encountered through these mediums from real ones (Reeves and Nass 1996; Kanazawa 2002). Overlap in the cognitive/neural processes engaged when trying to understand both real and fictional others, along with the sensory realism of audiovisual narratives, may explain in part why fictional characters can, at times, stand in for friends. On the other hand, the fact that most people understand that fictional characters are not real suggests that there must still be some important differences in individuals’ representations of real versus fictional others. Although prior neuroimaging work has revealed differences in average neural activity between real-life friends and fictional characters within brain regions involved in social processing (Abraham et al. 2008; Ganesh et al. 2012; Broom et al. 2021), it remains an open question how the neural representations (i.e. distributed patterns of neural activity) of real and fictional others differ and whether loneliness modulates how the brain represents fictional others.

Why might loneliness be expected to impact how the brain represents other persons? A recent study by Courtney and Meyer (2020) demonstrated that feelings of closeness with others were reflected in greater self-other neural similarity in social brain regions including the MPFC. However, this neural mapping of others was modulated by participants’ perceived social isolation. One characteristic of loneliness is the sense that one is not similar to those around them (Russell 1996). Consistent with this definition, lonelier individuals in their study exhibited reduced self-other neural similarity with their friends, acquaintances, and even celebrities compared with less lonely individuals. Moreover, while in general these different social categories (i.e. friends, acquaintances, and celebrities) were represented distinctly in the MPFC, the boundaries between them were blurred in lonelier individuals (Courtney and Meyer 2020). Because people may be especially likely to turn to fictional characters for experiences of belonging when they are feeling lonely (Derrick et al. 2009), this raises the question as to whether the boundary between real-life friends and fictional characters might similarly be blurred in lonelier individuals. Furthermore, if loneliness leads people to turn more and more often to fictional characters for a sense of connection, could the reduction in self-other neural similarity with friends observed for lonelier individuals be counterbalanced by an increase in self-other neural similarity with the fictional characters substituting for them?

In the present study, we leveraged an existing data set (Broom et al. 2021) to investigate the multivariate neural representation of fictional characters and real-life friends and acquaintances evaluated during a commonly used trait-evaluation task (Krienen et al. 2010; Mitchell et al. 2011; Ganesh et al. 2012; Tamir and Mitchell 2012; Chavez et al. 2017; Kang et al. 2018; Chavez and Wagner 2020; Courtney and Meyer 2020) while participants underwent functional magnetic resonance imaging (fMRI). Reliability-based voxel selection (Tarhan and Konkle 2020) was implemented to define regions of interest (ROI) in which the identities chosen as targets for the study were reliably represented (9 friends/acquaintances and 9 fictional characters from the television series Game of Thrones). Within these ROIs, we explored whether real and fictional others could reliably be distinguished, using both representational similarity analysis (RSA) and multivariate pattern classification. Next, in light of recent work demonstrating that the mapping of social categories is blurred in lonelier individuals (Courtney and Meyer 2020), as well as work suggesting fictional characters can act as surrogates for real friends when people feel socially isolated (Derrick et al. 2009; Gabriel et al. 2016), we investigated whether loneliness blurred the boundary between real and fictional others as well as whether it led to increased self-other neural similarity for fictional characters.

Materials and methods

Participants

Twenty-six right-handed MRI eligible fans of the HBO series Game of Thrones were recruited to participate in the present study. Participants were recruited from the broader Ohio State community primarily through the posting of flyers around the campus area and e-mails sent on university listservs. Seven participants were excluded because of insufficient variance in character ratings across the 9 television show characters selected for inclusion in the study (more details below), leaving a total of 19 healthy right-handed participants (10 females; median age = 24; range: 18–37 years) with normal or corrected-to-normal visual acuity who underwent functional neuroimaging. Participants were scanned during the airing of the seventh season of the television series as it was expected that thoughts and feelings about the characters would be at their most accessible while a season was ongoing. All participants were self-reported fans of the series as determined by an adapted 6-item measure of consumer devotion (Ortiz et al. 2013) and were up to date on the current events of the series (i.e. had previously viewed all 60 episodes prior to the airing of the seventh season). Participants gave informed consent in accordance with the guidelines set by the Office of Responsible Research Practices at The Ohio State University.

Task stimuli

The following characters from Game of Thrones, the selection process for which has been described previously (Broom et al. 2021), were used as target stimuli for all participants: Bronn, Catelyn Stark, Cersei Lannister, Davos Seaworth, Jaime Lannister, Jon Snow, Petyr Baelish, Sandor Clegane, and Ygritte. In an online survey completed prior to the fMRI portion of the study, participants provided ratings on scales from 0 to 100 of familiarity, closeness, similarity to self, liking, emotional attachment, and the extent to which they viewed each character as a friend. These last 2 items, the second of which was adapted from a measure of parasocial interaction (Rubin et al. 1985), were highly correlated with the closeness measure (closeness and emotional attachment: r = 0.83; closeness and extent to which characters were viewed as friends: r = 0.84), indicating that subjects correctly understood the intended meaning of the closeness measure as indicating a feeling of interpersonal connection and attachment. To be included in the study, subjects had to have a range > 50 and standard deviation > 16 in their closeness ratings of the 9 target characters. These criteria were implemented to ensure participants had sufficient variance in their closeness ratings of the 9 target characters and to avoid the situation whereby a participant liked or disliked all 9 characters equally. Seven otherwise MRI eligible fans of the show were excluded for not meeting these criteria and did not complete the fMRI portion of the study.

In a separate online survey also completed prior to the fMRI portion of the study, participants provided information on 9 self-selected friends/acquaintances with whom they were personally familiar, who varied in subjective interpersonal closeness, and who were not romantic partners or immediate family members. They reported each target person’s name, their relationship to that person (e.g. friend and coworker), and how long they had known them. They also provided the same 0 to 100 scale ratings of familiarity, closeness, similarity to self, and liking as they had for the fictional characters. These 18 targets (9 fictional and 9 real), along with the self, were the target identities included in the fMRI task described below.

Procedure

Prior to undergoing fMRI, participants were tested to ensure they were familiar with the 9 target characters. They did this first by completing a matching task where they had to match each character’s name to the correct corresponding picture from a set of character photos. Next, they were asked to verbally provide one fact about each target character. Participants providing incorrect responses for any of the characters were asked to review biographies for those characters (adapted from fansites, e.g. http://www.westeros.org/GoT/Characters/) to refresh their memories. This was necessary for only 4 of the 19 participants, with 3 characters being the highest number of errors on the matching task. All 4 of these participants were able to provide an additional novel fact for each of the mismatched characters after reviewing their biographies suggesting that they recalled who the characters were.

Tasks and experimental design

All stimuli were presented using PsychoPy2 (Peirce 2007, 2009). While in the scanner, participants completed 2 versions of a standard trait-evaluation task widely used in studies on self-referential processing (Krienen et al. 2010; Mitchell et al. 2011; Moran et al. 2011; Ganesh et al. 2012; Tamir and Mitchell 2012; Kang et al. 2018; Chavez and Wagner 2020) with the first version serving as a functional localizer for the vMPFC. The data from the functional localizer task were not used in the analyses reported herein.

For the main task, participants were presented with 2 words arranged vertically in black text against a gray background. The top word displayed was “self” for the self-condition, or the name of one of the other 18 targets included in the study: 9 fictional characters from Game of Thrones (common across all participants) and 9 personally familiar friends and acquaintances (specific to each participant). The bottom word displayed was one of 36 valence-balanced trait adjectives (Anderson 1968), none of which overlapped with those included in the functional localizer. On each trial, participants made a yes/no judgment as to whether the trait word displayed accurately described the target being displayed. Each word pairing was displayed for 2,000 ms followed by 500 ms of fixation. Intermittent jittered fixation trials (2,500–5,000 ms) were also included and Optseq2 (Dale 1999) was used to create templates for each of the 6 runs that optimized event presentation. Each run of the main task consisted of 6 judgments for each of the 19 targets. The order of the run templates, as well as which target replaced which placeholder within the templates, was randomized, resulting in a unique order of trials across the experiment for every participant thereby obviating any potential trial-order effects. Including the functional localizer task, the total duration for these 2 tasks was ~47 min.

Post-scanner survey

Immediately after scanning, participants completed a final survey that consisted of additional ratings of the targets (9 friends, 9 fictional characters, and the self), as well as several individual-difference measures described below. First, because the seventh season of the series was ongoing, participants were asked whether their feelings toward any of the target fictional characters had changed since their original ratings (i.e. on the prescreening survey) because of events in the new episodes. If so, participants adjusted their original ratings to better reflect their current feelings, and these updated ratings were used in the analyses. Participants also completed the UCLA Loneliness Scale (Russell 1996), the interpersonal reactivity index (Davis 1983), and the Transportability Scale (Dal et al. 2004), which measure a general tendency to become transported into narratives (Green and Brock 2000). Finally, participants completed the Ten-Item Personality Inventory (TIPI; Gosling et al. 2003), a brief measure of the Big-Five personality domains, for all 19 targets.

Participants’ loneliness scores varied from 27 to 61 (mean = 43.6, SD = 10.4) and were measured using The UCLA Loneliness Scale. This measure has previously been demonstrated to have good internal consistency (α between 0.89 and 0.94) as well as good test–retest reliability over a 1-year period (r = 0.73; Russell 1996). Internal consistency in the present study was high and consistent with this previous work (α = 0.95). Moreover, loneliness scores in the present study were uncorrelated with trait identification (i.e. the fantasy subscale of the IRI), transportability, and how big of a fan a participant was of the series (Ps > 0.17). Loneliness scores were also uncorrelated with both average and maximum ratings of interpersonal closeness, liking, similarity to self, and familiarity for the 9 fictional characters included in the study (Ps > 0.14). Loneliness scores were negatively correlated with the maximum liking rating provided for real-life friends/acquaintances (r = −0.69, P = 0.001), but there were no other statistically significant correlations with average or maximum ratings of interpersonal closeness, liking, similarity to self, or familiarity (Ps > 0.08).

While the psychometric properties of the TIPI are diminished relative to longer measures of the Big-Five personality domains, it was necessary to use a brief measure in the current study because of the number of targets being rated. The TIPI has previously been demonstrated to converge with longer Big-Five measures and to have good test–retest reliability over a 6-week period (r = 0.72; Gosling et al. 2003).

Image acquisition

Magnetic resonance imaging was conducted with a Siemens Prisma 3 Tesla scanner using a 32-channel phased array coil. Structural images were acquired using a T1-weighted MP-RAGE protocol (176 sagittal slices; TR: 1,900 ms; TE: 4.4 ms; flip angle: 12°; 1 mm isotropic voxels). Functional images were acquired using a T2^*-weighted echo planar sequence (TR: 2,500 ms; 45 axial slices, TE: 28 ms; flip angle: 76°; 3 mm isotropic voxels). For each participant, we collected 1 run of the functional localizer task (138 whole-brain volumes) and 6 runs of the main task (166 whole-brain volumes per run).

Image preprocessing

Preprocessing was conducted using SPM12 (Wellcome Department of Cognitive Neurology, London) in conjunction with a suite of in-house tools for preprocessing and analysis (SPM12w; https://github.com/wagner-lab/spm12w). First, images were corrected for differences in acquisition time between slices and then realigned within and across runs via a rigid body transformation in order to correct for head movement. Images were unwarped to reduce residual movement-related image distortions not corrected for by realignment and then normalized into a standard stereotaxic space (3 mm isotropic voxels) based on the SPM12 EPI template that conforms to the ICBM 152 brain template space (Montreal Neurological Institute). Normalized images were spatially smoothed (4 mm full-width-at-half-maximum) using a Gaussian kernel to increase the signal-to-noise ratio (SNR). Volumes were inspected for scanner- and motion-related artifacts based on an examination of the realignment parameters and temporal SNR for each run in each participant.

Trait-judgment task analysis

GLMs were constructed with each of the 19 individual identities defined as a separate condition (i.e. 9 fictional characters, 9 real-life friends/acquaintances, and the self). GLMs incorporated covariates of no interest (a linear trend to account for low-frequency drift, movement parameters) and were convolved with a canonical hemodynamic response function to compute parameter estimates (β) and contrast images (containing weighted parameter estimates) at each voxel. Two GLMs were constructed using only half the data for each target (i.e. only odd runs or only even runs) for the purpose of reliability-based voxel selection (Tarhan and Konkle 2020). A third GLM consisted of the full data for all targets and was used in all analyses other than RBVS.

Reliability-based voxel selection

Unless otherwise specified, all analyses described below were conducted using the PyMVPA toolbox (Hanke et al. 2009) within JupyterLab (Kluyver et al. 2016). RBVS (Tarhan and Konkle 2020) was used to identify regions of the brain in which identity-specific information is reliably represented. The self-condition was excluded from this analysis, leaving 18 conditions (9 fictional characters and 9 real-life friends/acquaintances) for which RBVS was conducted using split-half data (i.e. odd and even runs). RBVS was restricted to the cortex through the use of a mask defined using the Harvard-Oxford probabilistic atlas (Desikan et al. 2006). In the first step of RBVS, voxel-wise reliability was computed by taking the vector of parameter estimates for the 18 conditions in one half of the data and calculating its correlation with the corresponding vector of parameter estimates from the other half of the data at every voxel in the cortex. Individual subjects’ maps of Pearson r values were then averaged together into a single group-level map of voxel-wise reliability. In the second step of RBVS, the average multivoxel pattern reliability across conditions was calculated at every threshold of voxel-wise reliability from r > 0 up to the r values of the most reliable voxels identified within the data. In other words, for each threshold, only the voxels with reliability exceeding that threshold were retained, and for each condition, the parameter estimates across all retained voxels in one half of the data were vectorized and correlated with the corresponding vector from the other half of the data. This multivoxel pattern reliability was then averaged across all 18 conditions for each subject, and the group-level multivoxel pattern reliability was computed as well by taking the average across all subjects. As recommended by Tarhan and Konkle (2020), a heuristic approach was taken to identify the appropriate reliability threshold. Voxel-wise reliability thresholds were plotted against condition multivoxel pattern reliability to identify a range of thresholds in which the latter largely flattens out (see Fig. S1A). These plots were examined alongside brain maps to determine what threshold best balanced condition multivoxel pattern reliability and good coverage of ROI.

Representational similarity analysis

RSA (Kriegeskorte et al. 2008) was used to investigate the neural representational structure of real and fictional others within a given ROI. Within each ROI, the parameter estimates across all 19 targets (9 fictional characters, 9 friends/acquaintances, and self) were demeaned at every voxel. That is, at every voxel included in the ROI, the mean activation at that voxel across all 19 targets was subtracted from the parameter estimate for each individual target. Next, pairwise neural similarity was computed by correlating the patterns of neural activity for every possible pair of targets and then performing a Fisher’s r–z transformation. To account for the nested nature of the data (i.e. the targets evaluated were nested within participant), linear mixed-effects modeling was implemented in the R statistical language (R Core Team, 2022) using the Lme4 package (Bates et al. 2015) with a random intercept for each participant. The lmerTest package (Kuznetsova et al. 2017) was used to calculate Satterthwaite approximated degrees of freedom and corresponding P-values for all linear mixed-effects models reported. Confidence intervals for each fixed-effect parameter were estimated using the confint function of the Lme4 package (1000 simulations). In some cases, warnings were raised regarding potential issues of singularity (i.e. some dimensions of the variance–covariance matrix being estimated as 0). In these instances, models were also tested using a maximum penalized likelihood approach implemented with the blme package (Chung et al. 2013). In all instances in which both approaches were used, the same conclusion was supported (e.g. confidence intervals were identical after rounding); therefore, we report only the results using Lme4 and lmerTest.

To test the question of whether fictional characters and real-life friends/acquaintances were reliably differentiated, we examined whether pairwise neural similarity was significantly greater within category than between categories. Pairs of real-life friends and pairs of fictional characters were coded as 1, and pairs of targets containing both a friend and a character were coded as −1. Similarly, to test whether one of these categories was significantly higher in self-other neural similarity, real-life friends were coded as 1, and fictional characters were coded as −1. Moderation by loneliness was investigated by entering loneliness scores and their interaction term into the models described above.

Because ratings of interpersonal closeness, liking, similarity to self, and familiarity were generally highly correlated (see Supplementary Tables S1 and S2 for all correlations) and exhibited adequate average internal consistency (average α for real others = 0.90 and average α for fictional others = 0.79), these measures were combined into a composite score of psychological closeness. For both real and fictional others, though especially the latter, familiarity was not as highly correlated with the other ratings (average rs between 0.48 and 0.67, and between 0.25 and 0.33, respectively) as the other ratings were with one another (average rs between 0.81 and 0.88, and between 0.69 and 0.82, respectively). Whereas people can generally choose to spend more time with those they feel closer to, this is not the case with fictional characters from a television series for whom familiarity is confounded with screen time. To be consistent across social categories (i.e. real and fictional others) as well as consistent with previous work that examined familiarity, similarity, and liking together (Thornton and Mitchell 2017), we included all 4 dimensions of psychological closeness collected in our composite score. However, we note that composite scores excluding familiarity were extremely highly correlated with scores including familiarity (r = 0.96, P < 0.001) suggesting that excluding this measure would have little impact on the results.

Each of these 4 ratings was z-scored within category (i.e. separately for real-life friends and fictional characters), and then the average of the 4 z-scored ratings was computed for each target. Some previous work suggests that targets similar in psychological closeness should be more similar in their neural representations (Thornton and Mitchell 2017). Therefore, psychological closeness scores were modeled as the absolute difference between a pairs’ scores. However, according to construal level theory, people who are psychologically closer should be more likely to be represented concretely, i.e. in specific, idiosyncratic detail rather than more abstractly (Trope and Liberman 2010). Consistent with this idea, recent evidence suggests that individuals represent their own mental states more distinctly than others’ (Thornton et al. 2019), and objects higher in reward value more distinctly than those lower in reward value (Londerée and Wagner 2021). Conversely, the density hypothesis posits that psychologically close others should be especially similar to one another, whereas distant others should be especially distinct, because of the observation that liked others tend to cluster around the densest part of the distribution along many different traits, whereas distant others need only be perceived as extreme on a single trait to be disliked (Alves et al. 2016). To account for the possibility that psychological closeness may be asymmetrically associated with pairwise neural similarity (i.e. one end of the spectrum may be especially distinct in neural representation), psychological closeness scores were also modeled as the mean of a pairs’ scores. Lastly, psychological closeness scores were modeled as the difference (not absolute difference) between a pairs’ scores to investigate between-category pairs specifically, meaning pairs consisting of 1 real-life friend and 1 fictional character. The real-life friend’s psychological closeness score was subtracted from the fictional character’s psychological closeness score, thus the highest scores reflect pairs with a close character and distant friend and the lowest scores reflect pairs with a distant character and close friend. The purpose of modeling pairs’ scores in this way was to determine whether psychologically close fictional characters were especially similar to distant real-life friends/acquaintances and close real-life friends were especially dissimilar from distant fictional characters, which might suggest that others are represented on a spectrum of “realness” and that favored fictional characters might be especially close to this border as some previous work indicates they are perceived as more “real” relative to less favored characters (Gardner and Knowles 2008).

Classification analysis

To test whether real versus fictional others could be accurately decoded, PyMVPA’s C-SVM classifier using a linear kernel (LinearCSVMC) was implemented. Prior to conducting the classification analysis, data were demeaned for every target. That is, for a given target, the mean activation across all voxels was subtracted from the parameter estimate at every voxel for that target. This ensured that decoding accuracy was not because of differences in univariate activation across targets. For every participant, 2 targets were left out (1 real and 1 fictional), the classifier was trained on the remaining 16 targets, and then tested on the 2 left out targets. This process was repeated for every possible combination of 2 targets left out, and the accuracy scores were averaged across all combinations. Significant decoding at the group-level was determined by a 1-sample t-test comparing the distribution of accuracy scores to chance (0.5). Significant associations with loneliness scores were examined with Pearson correlations.

Control analyses

Though loneliness scores were not associated with ratings of psychological closeness, it remained a possibility that differences in the neural representations of fictional characters for lonely individuals relative to less lonely individuals could be because of systematic differences in the characters one gravitates toward. To account for this possibility, we tested whether similarity in loneliness was associated with similarity in which characters participants were psychologically close to. For every possible pair of participants, we calculated the absolute difference between their loneliness scores and the Euclidean distance between their composite psychological closeness scores for the 9 target fictional characters. Statistical significance of the correlation between pairwise similarity in loneliness and pairwise similarity in psychological closeness to the 9 characters was determined through a Mantel (1967) test with 10,000 permutations. A Mantel test is a nonparametric approach that involves creating a null distribution of correlation coefficients by taking one variable, shuffling participants, recomputing the pairwise metric for that variable, and then calculating the correlation coefficient between the 2 pairwise variables across n permutations. The true correlation coefficient is then compared with the null distribution of correlation coefficients.

Because some of the characters in Game of Thrones exhibit extreme behaviors that are likely to be outside of the lived experience of most participants as well as the people in their social networks, we investigated whether there were differences in how negatively the fictional characters’ personality traits were perceived relative to real-life friends/acquaintances and the self. To ensure that the TIPI ratings were reliably associated with positive/negative trait perceptions, we first conducted linear mixed-effects models with a random intercept for participants to determine which of the 10 personality ratings significantly predicted ratings of liking. We focused on liking specifically because, of the dimensions of psychological closeness collected in this study, it is the one most clearly associated with valence. For example, someone may be viewed negatively but still be highly familiar, or someone may be viewed as highly similar to the self while still being viewed negatively if one views themself negatively. Next, we computed a positivity score for each target, which was the average of all the TIPI ratings significantly associated with liking coded such that higher values reflected more positive trait perceptions. Paired-sample t-tests were conducted to determine whether the self and real others (averaged) were viewed significantly more positively than fictional others (averaged).

In the event that fictional characters were viewed more negatively than the self or real others, we conducted additional analyses to ensure that all results held controlling for trait perceptions. Dissimilarity in trait perceptions was modeled as the Euclidean distance between a pair of targets TIPI ratings. Dissimilarity in trait perceptions was then included as an additional predictor in all significant results to investigate whether those results could be explained by differences in trait perceptions. Finally, analyses were also conducted with Dissimilarity in trait perceptions as the dependent variable to determine whether the neural results were also reflected in behavioral trait ratings.

Results

Definition of ROI using reliability-based voxel selection

RBVS (Tarhan and Konkle 2020) was used to identify brain regions in which identity-specific information was reliably represented. Voxel-wise reliability thresholds were plotted against condition multivoxel pattern reliability (Fig. S1A) and examined alongside brain maps to determine what threshold best balanced condition multivoxel pattern reliability and good coverage of ROI. A voxel-wise reliability threshold of 0.23 was chosen because of the corresponding condition multi-voxel pattern reliability being comparable to higher thresholds up to ~0.60 as well as because of adequate coverage of clusters within the MPFC and precuneus (PC)/posterior cingulate cortex (PCC): 460 voxels and 122 voxels, respectively. A reliable cluster of voxels was also found within the visual cortex (Fig. S1B). However, this cluster was discarded and all remaining analyses focused on the ROIs identified within the MPFC and PC/PCC (Fig. S1C) because of prior work implicating these regions as central to the neural representations of others (Wagner et al. 2019). There were no significant differences between the 3 categories (self, real-life friends/acquaintances, and fictional characters) in average condition multi-voxel pattern reliability in either the MPFC ROI (F_2,36 = 0.65, P = 0.53) or PC/PCC ROI (F_2,36 = 2.58, P = 0.09), despite the self-condition not being included in the RBVS procedure.

Are the neural representations of real and fictional others distinct?

To test whether the MPFC and PC/PCC reliably distinguish between real and fictional others, linear mixed-effects modeling was implemented to determine whether within-category pairwise neural similarity (Fisher’s z transformed Pearson correlation) was greater than between-category pairwise neural similarity. Pairs of real-life friends/acquaintances and pairs of fictional characters were coded as 1, and pairs consisting of 1 friend and 1 character were coded as −1. Within the MPFC, within-category pairwise neural similarity was significantly greater than between-category pairwise neural similarity (B = 0.25, SE = 0.006, t₍₂₉₀₅₎ = 40.42, P < 0.001, 95% CI = 0.24–0.26; Fig. 1B), suggesting that real and fictional others are reliably differentiated in this region. Real and fictional others were also represented distinctly within the PC/PCC (B = 0.20, SE = 0.006, t₍₂₉₀₅₎ = 32.28, P < 0.001, 95% CI = 0.19–0.22; Fig. S2B).

Fig. 1

Results of the RSA conducted within the MPFC. A) ROI in the MPFC. B) Raincloud plot depicting the distribution of pairwise neural similarity values for within-category (real other with real other or fictional other with fictional other) versus between-category (real other with fictional other) pairs. C) Differences in within-category versus between-category pairwise neural similarity for individuals high versus low in loneliness. D) Raincloud plot depicting the distribution of self-other neural similarity for real-life friends and fictional characters. E) Differences in the relative difference in self-other neural similarity between real-life friends and fictional characters for individuals high versus low in loneliness. A median split on loneliness was used for visualization purposes. Asterisks indicate the statistical significance of the interactions between loneliness scores and the contrast coding used for within-category versus between-category pairs or real-life friends versus fictional characters, respectively. ^**P < 0.01. ^***P < 0.001.

Linear SVM classification analyses further supported the idea that there is a clear boundary between real and fictional others in their neural representations. Within the MPFC, a 1-sample t-test suggested that classification accuracy for real and fictional others was well above chance (mean accuracy = 0.92, range = 0.75–1.0, t₍₁₈₎ = 23.94, P < 0.001). Classification accuracy was similarly high within the PC/PCC (mean accuracy = 0.90, range = 0.72–1.0, t₍₁₈₎ = 19.96, P < 0.001). Indeed, classification accuracy was not only statistically significantly above chance at the group level but for each individual participant in each ROI as indicated by binomial tests comparing individual distributions of correct and incorrect classifications to chance (all Ps < 0.001).

Are real others or fictional others higher in self-other neural similarity?

To test whether real-life friends/acquaintances or fictional characters were reliably higher in self-other neural similarity, real others were coded as 1, and fictional others were coded as −1. Within the MPFC, linear mixed-effects modeling indicated that real-life friends/acquaintances were significantly higher in self-other neural similarity than fictional characters (B = 0.16, SE = 0.02, t₍₃₄₀₎ = 9.11, P < 0.001, 95% CI = 0.12–0.19; Fig. 1D). Real others were also significantly higher in self-other neural similarity than fictional others within the PC/PCC (B = 0.13, SE = 0.02, t₍₃₄₀₎ = 6.82, P < 0.001, 95% CI = 0.10–0.17; Fig. S2D). Consistent with previous work (Courtney and Meyer 2020), these findings suggest that one way in which cortical midline structures “map” social categories is in terms of their neural representational similarity to the self.

Does loneliness modulate the neural mapping of real and fictional others?

To test whether loneliness modulates the finding that real and fictional others are distinct in their neural representations, loneliness was entered as a predictor into the model along with the contrast described above (i.e. within-category pairs = 1, between-category pairs = −1) and their interaction term. Within the MPFC, there was a significant interaction between the within-versus-between pairs contrast and loneliness scores (contrast: B = 0.41, SE = 0.03, t₍₂₉₀₃₎ = 15.27, P < 0.001, 95% CI = 0.35–0.46; loneliness: B = −0.0002, SE = 0.0006, t₍₂₉₀₃₎ = 0.29, P = 0.77, 95% CI = −0.001 to 0.001; interaction: B = −0.004, SE = 0.0006, t₍₂₉₀₃₎ = 6.00, P < 0.001, 95% CI = −0.005 to −0.002) such that the difference between within-category pairwise neural similarity and between-category pairwise neural similarity was attenuated for individuals higher in loneliness (Fig. 1C). Fig. 2 illustrates this finding by showing the representational similarity matrices for the least lonely participant and loneliest participant in the sample. Whereas clear boundaries between real and fictional others are evident in the representational similarity matrix for the least lonely participant, the boundary between these 2 social categories is almost nonexistent for the loneliest participant. The results of the classification analysis further support the idea that the boundary between real and fictional others is blurred in the MPFC for lonelier individuals.

Fig. 2

Matrix plots depicting pairwise neural similarity in the MPFC ROI (pictured to the left) for the least lonely participant and loneliest participant in the sample. Whereas the boundary between real and fictional others is clearly visible to the eye in the least lonely participant (light squares for within-category pairs and dark squares for between-category pairs), this boundary is not perceptible in the case of the loneliest participant.

With regard to the neural decoding analysis, there was a significant association between overall classification accuracy and loneliness scores: r = −0.47, P = 0.04. Furthermore, breaking down the classification results by category indicated that loneliness scores were associated with classification accuracy for real-life friends and acquaintances (r = −0.58, P = 0.009) but not classification accuracy for fictional characters (r = 0.24, P = 0.32). In other words, greater loneliness was associated with higher rates of misclassifying real others as fictional others, but not with misclassifying fictional others as real others (see Fig. S3).

A significant interaction between the within-versus-between pairs contrast and loneliness scores was also found within the PC/PCC (contrast: B = 0.27, SE = 0.03, t₍₂₉₀₃₎ = 9.81, P < 0.001, 95% CI = 0.21–0.32; loneliness: B = −0.0002, SE = 0.0006, t₍₂₉₀₃₎ = 0.26, P = 0.79, 95% CI = −0.001 to 0.001; interaction: B = −0.001, SE = 0.0006, t₍₂₉₀₃₎ = 2.38, P = 0.02, 95% CI = −0.003 to −0.0002; Fig. S2C). While there was no significant association between overall classification accuracy within the PC/PCC and loneliness scores (r = −0.32, P = 0.18), breaking down the classification results by category revealed a pattern of results consistent with that found within the MPFC: there was no association between loneliness and classification accuracy for fictional characters (r = 0.06, P = 0.79), but there was an association between loneliness and classification accuracy for real-life friends/acquaintances (r = −0.46, P = 0.047; see Fig. S3).

To test whether loneliness modulates the finding that real-life friends/acquaintances are higher in self-other neural similarity than fictional characters, loneliness was entered as a predictor into the model along with the social-category contrast (i.e. real other = 1, fictional other = −1) and their interaction term. Within the MPFC, there was a significant interaction between the social-category contrast and loneliness scores (contrast: B = 0.37, SE = 0.07, t₍₃₃₈₎ = 5.08, P < 0.001, 95% CI = 0.23–0.51; loneliness: B = 0.0002, SE = 0.002, t₍₃₃₈₎ = 0.11, P = 0.91, 95% CI = −0.031 to 0.004; interaction: B = −0.005, SE = 0.002, t₍₃₃₈₎ = 3.03, P = 0.003, 95% CI = −0.008 to −0.002) such that the difference in self-other neural similarity between real-life friends/acquaintances and fictional characters was attenuated for individuals higher in loneliness (Fig. 1E). There was no interaction between the social-category contrast and loneliness scores within the PC/PCC (contrast: B = 0.24, SE = 0.09, t₍₃₃₈₎ = 2.86, P = 0.004, 95% CI = 0.08–0.41; loneliness: B = 0.001, SE = 0.002, t₍₃₃₈₎ = 0.57, P = 0.57, 95% CI = −0.003 to 0.005; interaction: B = −0.003, SE = 0.002, t₍₃₃₈₎ = 1.32, P = 0.19, 95% CI = −0.006 to 0.001; Fig. S2E).

Is psychological closeness associated with the neural representations of real and fictional others?

Because our findings above suggest that there is a clear boundary between real and fictional others within cortical midline structures, associations with psychological closeness (a composite score of interpersonal closeness, liking, similarity to self, and familiarity) were investigated for fictional characters and real-life friends/acquaintances independently. As explained in the Materials and Methods section above, we tested both whether pairs of targets similar in psychological closeness (absolute difference between 2 scores) were more similar in their neural representations and whether one end of the spectrum of psychological closeness (mean of 2 scores) was represented more similarly than the other. Within the MPFC, there was no evidence that fictional characters similar in psychological closeness were higher in pairwise neural similarity (P = 0.43). However, greater mean psychological closeness scores were associated with decreased pairwise neural similarity within the MPFC (B = −0.08, SE = 0.02, t₍₆₆₄₎ = 3.63, P < 0.001, 95% CI = −0.12 to −0.03), suggesting that psychologically distant characters are represented especially similarly to one another, whereas psychologically close characters are especially distinctive in their neural representations. There were no significant associations between psychological closeness and pairwise neural similarity within the PC/PCC for fictional characters (Ps > 0.57).

For real-life friends/acquaintances, there was also an association between mean psychological closeness scores and pairwise neural similarity within the MPFC, though in the opposite direction (B = 0.05, SE = 0.02, t₍₆₆₄₎ = 2.56, P = 0.01, 95% CI = 0.01–0.09), suggesting that psychologically close friends/acquaintances are represented especially similarly to one another, whereas psychologically distant friends/acquaintances are especially distinctive in their neural representations. However, there was also an association between dissimilarity (absolute difference) in psychological closeness scores and pairwise neural similarity within the MPFC (B = −0.05, SE = 0.02, t₍₆₆₆₎ = 2.86, P = 0.004, 95% CI = −0.08 to −0.01), showing that friends/acquaintances who are similar in psychological closeness are more similar in their neural representations. Moreover, when both the mean and absolute difference models of psychological closeness were entered as predictors simultaneously, only the absolute difference model remained a statistically significant predictor of pairwise neural similarity within the MPFC (absolute difference in psychological closeness: B = −0.04, SE = 0.02, t₍₆₆₅₎ = 2.15, P = 0.03, 95% CI = −0.07 to −0.001; mean psychological closeness: B = 0.04, SE = 0.02, t₍₆₆₃₎ = 1.77, P = 0.08, 95% CI = −0.004 to 0.08), though mean psychological closeness remained marginally associated. Within the PC/PCC, mean psychological closeness was associated with increased pairwise neural similarity (B = 0.05, SE = 0.02, t₍₆₆₄₎ = 2.11, P = 0.04, 95% CI = 0.001–0.09) and dissimilarity in psychological closeness was marginally associated with decreased pairwise neural similarity (B = −0.03, SE = 0.02, t₍₆₆₈₎ = 1.85, P = 0.07, 95% CI = −0.07 to 0.002), but when both predictors were entered simultaneously neither explained significant additional variance beyond the other (Ps > 0.10).

Within the PC/PCC, greater psychological closeness of real-life friends/acquaintances was associated with lower self-other neural similarity (B = −0.07, SE = 0.03, t₍₁₅₁₎ = 2.38, P = 0.02, 95% CI = −0.12 to −0.01). There was no significant association between psychological closeness and self-other neural similarity in the PC/PCC for fictional characters, nor for fictional characters or real-life friends/acquaintances within the MPFC (Ps > 0.11). None of the associations between pairwise neural similarity or self-other neural similarity and psychological closeness were moderated by loneliness (Ps > 0.30).

Are psychologically close fictional characters “realer” than distant ones?

Because psychological closeness was associated with pairwise neural similarity in the MPFC for both real-life friends/acquaintances and fictional characters, and because some behavioral work suggests that favored fictional characters are perceived as more “real” than less favored ones (Gardner and Knowles 2008), we focused on only between-category pairs (i.e. 1 fictional character and 1 real-life friend/acquaintance) to see whether psychological closeness predicted which characters were most similar to which real-life friends/acquaintances. We expected that real-life friends and fictional characters would be represented on a spectrum of “realness” with the closest friends being least similar to fictional characters and the closest fictional characters being most similar to real-life others. To test this possibility, we modeled psychological closeness as the difference (not absolute difference) between a fictional character’s psychological closeness score and a real-life friend/acquaintance’s psychological closeness score so that the highest values corresponded to close characters and distant real others and the lowest values corresponded to close real others and distant characters. Difference in psychological closeness scores between fictional characters and real-life friends/acquaintances was associated with greater pairwise neural similarity within the MPFC (B = 0.03, SE = 0.007, t₍₁₅₁₉₎ = 4.29, P < 0.001, 95% CI = 0.02–0.04), showing that for between-category pairs the greatest neural dissimilarity was between close friends and distant fictional characters, whereas the greatest neural similarity was between close characters and distant friends/acquaintances. This association was not moderated by loneliness (interaction P = 0.71).

Is loneliness associated with which characters an individual is drawn to?

Because loneliness modulated the manner in which person knowledge was represented in the MPFC and PC/PCC (i.e. the boundary between real and fictional others was blurred for lonelier individuals), we tested for the possibility that lonelier individuals are systematically drawn to similar characters, which, in turn, might account in part for the differences observed. For every possible pair of participants, we calculated the absolute difference between their loneliness scores and the Euclidean distance between their psychological closeness scores for the 9 target fictional characters. Results indicated that there was no significant association between similarity in loneliness and similarity in which characters participants were psychologically close to (r = −0.09, permuted P = 0.40). This suggests that the blurring of the boundary between real and fictional others observed in midline cortical structures cannot be explained by systematic differences in which fictional characters one is drawn to according to their level of loneliness.

Controlling for trait perceptions

To determine whether the results reported above could be explained by differences in trait perceptions, we first examined whether there were differences in positive trait perceptions of fictional characters compared with the self and real others. Linear mixed-effects models predicting ratings of liking from each of the 10 TIPI personality ratings showed that higher agreeableness, higher conscientiousness, lower neuroticism, and higher openness to experiences were significantly associated with greater liking. For ratings of extraversion, however, we found that only the rating reflecting higher extraversion (“I see [target] as extraverted, enthusiastic”) was significantly positively associated with liking (B = 3.53, SE = 0.86, t₍₃₄₀₎ = 4.09, P < 0.001), whereas the rating reflecting lower extraversion (“I see [target] as reserved, quiet”) was not significantly associated with liking (B = 0.90, SE = 0.86, t₍₃₄₀₎ = 1.04, P = 0.30). Positive trait perceptions were calculated as the average of the 9 TIPI ratings that were significantly associated with liking, with each rating re/coded so that higher scores reflected more positive perceptions. Results of paired-samples t-tests indicated that perceptions of one’s own traits and averaged perceptions of real-life friends and acquaintances’ traits were both significantly more positive than averaged perceptions of fictional characters’ traits: t₍₁₈₎ = 4.41, P < 0.001 and t₍₁₈₎ = 2.88, P = 0.01, respectively.

Because fictional characters were perceived more negatively than the self and real others in terms of their traits, we conducted additional analyses to ensure that all significant results reported above held controlling for trait perceptions (see Materials and Methods section for additional details). We found that, with the exception of 2 analyses, all findings reported herein remained statistically significant controlling for trait perceptions (see Supplementary Materials, pp. 2–5). The association between mean psychological closeness and pairwise neural dissimilarity in the PC/PCC for real-life friends/acquaintances became marginally significant after controlling for trait perceptions (P = 0.08). Similarly, the association between psychological closeness and self-other neural similarity in the PC/PCC for real-life friends/acquaintances became marginally significant after controlling for trait perceptions (P = 0.099). Finally, we note here that for real-life friends/acquaintances similarity in trait perceptions was significantly associated with pairwise neural similarity in the MPFC and PC/PCC (see Supplementary Materials, p. 4), which is consistent with some previous work suggesting that trait perceptions are associated with neural representations of person knowledge in midline cortical structures (Hassabis et al. 2014; Thornton and Mitchell 2018).

Discussion

The current study demonstrates that the boundary between real and fictional others within midline cortical structures is blurred for lonelier individuals. The neural representations of real and fictional others were highly distinguishable within the MPFC and PC/PCC, 2 regions commonly implicated in encoding social knowledge (Wagner et al. 2019), including individuals’ social connections to others (Parkinson et al. 2017; Thornton and Mitchell 2017; Courtney and Meyer 2020). While in some ways it is not surprising that these brain regions distinguish personally familiar friends and acquaintances from fictional characters existing in a medieval fantasy world, the differences between the 2 worlds in which these friends and characters are encountered only make it all the more striking that the boundary between real and fictional others was blurred in lonelier participants. The difference in pairwise neural similarity between within-category pairs (i.e. pairs of friends/acquaintances or pairs of characters) and between-category pairs (i.e. pairs with 1 friend/acquaintance and 1 character) within the MPFC and PC/PCC was attenuated in lonelier individuals such that within-category similarity was lower and between-category similarity was greater relative to less lonely individuals. Relatedly, the difference in self-other neural similarity between real and fictional others within the MPFC was altered in lonelier individuals such that self-other neural similarity with real-life friends/acquaintances was lower and self-other neural similarity with fictional characters was greater relative to less lonely individuals. Moreover, the influence of loneliness on the neural representations of others was independent of any association with psychological closeness, suggesting that while loneliness may warp the mapping of social categories in the MPFC and PC/PCC (including in terms of their similarity to the self) it may not alter the content of the information being encoded within social categories. Finally, though participants tended to view fictional characters’ traits more negatively than their own or those of their friends/acquaintances, we found that the main findings reported herein held even after controlling for trait perceptions.

Modulation of neural representations of others by loneliness

Loneliness, or perceived social isolation, is associated with a host of self-defeating cognitive biases, depression, and increased morbidity and mortality (Cacioppo and Hawkley 2009; Cacioppo and Cacioppo 2018). Recent work has begun to characterize how lonely individuals’ neural representations of their social connections differ from those of less lonely individuals (Courtney and Meyer 2020), and the present findings build on this work by demonstrating that the blurring of boundaries between social categories that occurs in the brains of lonelier individuals extends to the line between real and fictional others. Humans are a highly social species with a fundamental need to belong (Baumeister and Leary 1995). When a belongingness motivation is activated, evidence suggests that individuals often seek to address their social needs from entities outside of their real-world social networks, which can include engaging with fiction (Derrick et al. 2009; Gabriel et al. 2016) or anthropomorphizing nonhuman objects including pets (Epley et al. 2007, 2008). The fuzzy boundary between real and fictional others in the MPFC and PC/PCC may reflect the consequence of looking to fictional characters for a sense of belonging when it is lacking from one’s real-world social network. In other words, processing narrative fiction when in a state of high need to belong, whether one is conscious that they are seeking to satisfy this motivation or not, may impact the nature of the representation that is stored in memory such that fictional characters, in general, more closely resemble real friends and acquaintances. Consistent with this idea, one of the ways in which fictional narratives provide a sense of belonging is by making one feel as though they belong to the same group as the characters in a story (Gabriel and Young 2011). The blurred boundary between real and fictional others resulted both from an increase in between-category similarity (i.e. real and fictional others becoming more similar in their neural representations) and a decrease in within-category similarity (see Fig. 1). Theorizing on social surrogacy suggests that though activities such as watching TV or reading for pleasure are ostensibly nonsocial, they nevertheless fulfill belongingness needs by immersing individuals in social worlds and thus serve to protect individuals from the negative consequences of social rejection and isolation (Gabriel et al. 2016). Engaging with narrative fiction may be an adaptive way to manage feelings of loneliness (at least in the short term). Therefore, it is possible that the finding that the neural representations of fictional characters more closely resemble those of real-life friends and acquaintances for lonelier individuals need not necessarily be viewed as deleterious as the finding that members of one’s real-world social network are less similar and thus are not “mapped” as a cohesive group in the MPFC or PC/PCC. Indeed, the results of the pattern classification analysis showed that loneliness was associated with misclassification of real others but not fictional others.

The blurred boundary between real and fictional others was also reflected in their relative self-other neural similarity within the MPFC. We previously reported large univariate differences in activation within the MPFC, and in particular the ventral MPFC, such that the greatest activation was observed for the self, followed by real-life friends/acquaintances, followed by fictional characters (Broom et al. 2021). Here, we show a similar pattern for self-other neural similarity defined in terms of multivoxel patterns of neural activity rather than univariate activation. In general, self-other neural similarity was much greater for real-life friends/acquaintances compared with fictional characters, but this was attenuated for lonelier individuals such that real others were less similar to self and fictional others were more similar to self in their neural representations. Though this finding certainly overlaps with the blurring of the boundary between real and fictional others, it need not be the case that the 2 findings are necessarily dependent, as it just as easily could have been the case that fictional characters were more similar to real-life friends/acquaintances within multidimensional representational space in a way orthogonal to self-other neural similarity. Thus, increasing self-other neural similarity with fictional characters may reflect ways in which lonely individuals seek a sense of belonging from characters that is at least partly independent of the finding that characters more closely resemble real-life friends/acquaintances. Many people turn to fiction for moments that resonate with their own feelings and experiences, and lonely individuals may be especially likely to be impacted by these moments if they are lacking in their real-world interactions. Relatedly, Courtney and Meyer (2020) have posited that self-other neural similarity may reflect the extent to which another has been incorporated into one’s sense of self (Aron et al. 1992), suggesting that reduced self-other neural similarity for real-life friends/acquaintances within the MPFC for lonelier individuals may reflect the fact that members of their real-life social network are not incorporated into their self-identity to the same extent as for less lonely individuals.

Neural representations of others and psychological closeness

In addition to being represented distinctly from one another, real-life friends/acquaintances and fictional characters also differed in how their neural representations within the MPFC were associated with psychological closeness. Friends and acquaintances who were more similar in psychological closeness were also more similar in their neural representations, demonstrating that the MPFC encodes information about individuals’ social connections to others, including, as other work shows, their interpersonal closeness (Courtney and Meyer 2020) as well as their familiarity, similarity to self, and how well they are liked (Thornton and Mitchell 2017). For fictional characters, on the other hand, greater mean psychological closeness was associated with decreased pairwise neural similarity. In other words, psychologically closer fictional characters were more distinct in their neural representations. This finding adds to a growing body of evidence suggesting that distinctiveness in neural representations (i.e. greater dissimilarity to other, related neural representations) is a meaningful outcome indicating that the object being represented may be brought to mind in finer-grained detail. For example, recent evidence suggests that tastier foods are represented more distinctly than less tasty foods (Londerée and Wagner 2021) and people represent their own mental states more distinctly than those of others (Thornton et al. 2019). According to construal level theory, psychologically closer objects are more likely to be represented in terms of their concrete, idiosyncratic detail as opposed to their abstract, invariant features (Trope and Liberman 2010). It may be the case that favored fictional characters stand apart from the rest because they are brought to mind in greater, individuating detail than more distant fictional characters who, as a result, tend to look more alike in their neural representations. Moreover, the association between psychological closeness and pairwise neural similarity for both real and fictional others was independent of the influence of loneliness, suggesting that the tendency for a handful of favored fictional characters to stand apart from the rest may be typical, at least for individuals who derive enjoyment from narrative fiction as was the case for the present sample. In other words, it is not the case that psychologically close fictional characters are especially distinctive for lonelier individuals, rather this finding appears to reflect how people in general represent favored characters from a series they enjoy.

In addition to standing apart from one another, psychologically close fictional characters were represented especially similarly to real others, and more specifically psychologically distant real others. The lowest pairwise neural similarity for between-category pairs, on the other hand, was observed for pairs of close real others and distant fictional characters. This suggests that, when brought to mind within the same context, real and fictional others may be represented on a spectrum of “realness” with distant fictional characters being the least real and close friends being the realest. This finding is consistent with work showing that favored fictional characters are perceived as “realer” than other fictional characters for which people have similar levels of knowledge (Gardner and Knowles 2008). Moreover, Gardner and Knowles (2008) demonstrated that social facilitation effects occur when people are in the presence of their favorite fictional character, but not someone else’s. Again, loneliness did not moderate this effect, suggesting that psychologically close fictional characters being represented more similarly to real others relative to more psychologically distant fictional characters is typical of how people in general engage with fiction: certain characters capture individuals’ attention, eliciting feelings of interpersonal closeness or perceptions of similarity that result in those characters feeling more “real.” Here we provide evidence that the perceptions of realness reported in previous work (Gardner and Knowles 2008) may be reflected in neural representations of fictional characters that more closely resemble those of members of one’s real-life social network.

Limitations

One limitation of the present study is that the fictional characters focused on were drawn from a fantasy story with settings and extreme events far removed from participants’ lived reality in Ohio in the United States in 2017. Though we controlled for trait perceptions measured on the Big-Five personality domains, it is possible that the boundary between real and fictional others identified in this study was driven by differences in trait perceptions that were not collected (e.g. perceptions of one being “evil”) rather than the fictional nature of the characters per se. In addition, though our results suggest that the association between psychological closeness and pairwise neural similarity in the MPFC differs between real and fictional others, this carries with it the caveat that the selection process differed between these 2 social categories (i.e. real-life friends and acquaintances were self-selected, whereas the fictional characters were chosen by the experimenters and constant across all participants). Moreover, there is the additional difference that participants have met and had face-to-face interactions with members of their real-world social network, but cannot interact with fictional characters, though this limitation is true of any work focused on the psychological/neural impacts of fictional narratives. Finally, the generalizability of our results may be limited to individuals who enjoy fiction. Future work should investigate whether similar effects are observed for individuals who form parasocial attachments to nonfictional social surrogates (e.g. one’s favorite podcaster, athlete, or TV personality).

Conclusion

Part of what draws people to narrative fiction is the chance to indirectly experience extreme, dark situations without the risk of actual consequences (Green et al. 2004). When individuals watch a show like Game of Thrones, they can vicariously experience deadly battles without putting themselves in harm’s way. But engaging with fiction also offers protection from other sorts of harms. One can grow emotionally attached to a fictional character without fear of rejection or judgment. The findings of the current study suggest that when people turn to fictional characters for a sense of belonging that is lacking from their real-world social connections, there are downstream consequences on the manner in which fictional characters and real-life friends/acquaintances are represented within the social brain: the boundary between the 2 becomes blurred with the neural representation of fictional characters coming to resemble that of real-world friends.

Acknowledgments

The authors thank the Center for Cognitive and Behavioral Brain Imaging (CCBBI) at The Ohio State University.

Author contributions

Timothy W. Broom (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Visualization, Writing—original draft, Writing—review & editing) and Dylan D. Wagner (Conceptualization, Methodology, Software, Supervision, Writing—review & editing)

Conflict of interest statement: None declared.

Data availability statement

All code and materials are available upon reasonable request.