AR filters change how we think, not who we like: effects of role and age embodiment on cognitive performance and age bias

1 Introduction

Digital technologies allow individuals to adopt identities that differ radically from their real-world selves. In virtual environments, social behavior often mirrors real-world dynamics, meaning that existing biases and prejudices can reappear and lead to discrimination (Banakou and Chorianopoulos, 2010). Luckily, it can also promote positive change:embodied identification with virtual avatars in VR has been shown to reduce a wide range of social biases (Hasler et al., 2017; Oh et al., 2016; Gonzalez-Liencres et al., 2020). This includes several prominent forms of bias, such as racism and sexism, but also less frequently addressed forms of bias like ageism (Banakou et al., 2018; Oh et al., 2016). While, Ageism is not as frequently discussed it has become an increasingly relevant issue, driven by both technological shifts, and a growing older population (Butler, 1980; Bodner, 2009; Nelson, 2005; Stypińska et al., 2023). Evidence suggests that enabling people to experience embodied identification with older adult avatars can meaningfully reduce age-based prejudice (Oh et al., 2016; Banakou et al., 2018). Most research on embodied identification has focused on fully immersive VR experiences, where the user’s entire body is replaced by a virtual avatar within a simulated environment. Typically, this involves a head-mounted display that completely occludes the user’s vision of the physical world. However, fully immersive VR is not the only mixed-reality experience that can promote embodied identification. Recent research suggests that embodied identification with augmented reality (AR) filters—such as those available on social media platforms like Snapchat—may also promote attitudinal and behavioral changes (Jarrell et al., 2021; Leong et al., 2021; Genay et al., 2021; Fribourg et al., 2021). This is despite the fact that many AR experiences only offer partial virtual transformation: users remain embedded in the real world while selected aspects of their appearance are digitally modified. Still, studies have shown that AR filter embodied identification can influence users’ self-perception (Fribourg et al., 2021) and even modulate cognitive processes such as creativity (Leong et al., 2021). Moreover, embodied identification with “gender-swapping” Snapchat filters has been associated with both increases and decreases in perceived gender bias, depending on the direction of identification and the participant’s gender group (Jarrell et al., 2021).

While most theories on embodied identification have been developed in the context of fully immersive virtual reality it seems to be somewhat present in AR experiences. Which is extremely noteworthy as the growing popularity of AR filters on smartphone platforms; such as Snapchat, Instagram, and TikTok, gives millions of daily users the ability to embody and identify with multiple virtually transformed versions of themselves (Keach, 2019). AR filters already generate a wide variety of altered personas for people to embody and identify with by virtually reshaping a person’s face (Fribourg et al., 2021), or by altering perceived social categories such as gender (Harbison, 2019), race (Bell, 2017), or age (Keach, 2019). In VR, comparable transformations have been shown to both decrease and increase social biases, depending on the context, duration, and type of embodiment (Groom et al., 2009; Lopez et al., 2019; Banakou et al., 2020). Yet, despite their global reach, empirical evidence on how these mixed-reality experiences affect users’ remains limited. Given that AR filters have become a routine form of digital self-expression, used by millions of people worldwide each day, their psychological and social effects deserve closer examination. Are these filters subtly reshaping attitudes and reinforcing stereotypes, or could they be harnessed to promote empathy and reduce prejudice? If AR embodiment indeed influences perception and behavior, it may represent a promising yet underexplored avenue for bias-reduction interventions, complementing the outcomes of fully immersive VR studies (Banakou et al., 2016; Oh et al., 2016; Gabbiadini et al., 2016).

The present study seeks to address this gap and establish a foundation for future investigations. To do so, it attempts to reproduce key findings from previous VR studies on age-based embodiment using readily accessible AR technology—specifically, Snapchat face filters. Short, smartphone-based interactions were deliberately chosen over more elaborate and longer AR setups, as these filters represent the most familiar, popular, and widely available form of AR experience among the general public today (Alkhamisi and Monowar, 2013; Haines, 2021). If AR embodiment exerts any measurable psychological or behavioral influence, it is essential to first understand its impact in this most pervasive and ecologically valid context. Accordingly, the study examines how virtually aging the user and/or presenting them with the appearance of an intellectual professional through AR filters affects both age-related biases and cognitive performance. This approach parallels earlier VR research in which participants embodied older avatars—most notably, representations of Albert Einstein—and showed corresponding changes in age-related attitudes and cognitive task performance (Oh et al., 2016; Banakou et al., 2018). We also aim to build on past research by separating the impact of age and professional appearance, which we found was often conflated in previous VR and AR studies. Specifically, earlier experiments investigating virtual intellectual identities often used avatars or filters that combined age-related and professional cues—such as white hair, wrinkles, and lab coats—to represent scientists (Leong et al., 2021; Banakou et al., 2018; Guegan et al., 2016). Yet, both age and profession independently carry societal expectations about intellect, authority, and competence, and may interact in complex, intersectional ways.

Building on this observation, we adopted an intersectional framework (Crenshaw, 1989; Crenshaw, 1991) to disentangle and re-examine the unique and combined effects of virtual aging and virtually wearing scientific attire. Participants used a Snapchat AR filter that either virtually aged their appearance, dressed them as a scientist, combined both manipulations, or left their appearance unchanged in a control condition. Cognitive performance was measured through a timed mental arithmetic task, while implicit and explicit attitudes toward older adults were assessed using an Implicit Association Test (IAT) and questionnaires. This design allowed us to evaluate how everyday AR interactions—similar to those millions of users engage with on social media—might shape perception, performance, and social bias, and to explore whether AR embodied identification could be harnessed as a scalable medium for bias-reduction interventions, extending insights from previous VR research (Kishore et al., 2021; Gonzalez-Liencres et al., 2020).

We examined how various cognitive and age-based bias metrics were impacted by embodied identification across four experimental conditions (a control condition, a scientist condition, an elder adult condition, and an elder scientist condition). Based on prior literature, we formulated the following hypotheses which were the basis of our study:

H1 Age manipulation hypothesis. Interacting with an AR filter that virtually ages one’s appearance will reduce age-based bias toward older adults. This reduction may manifest as (a) more positive implicit associations in the Age IAT (H1a), (b) more favorable explicit evaluations of older adults (H1b), (c) a greater willingness to engage with older adults (H1c), and/or (d) stronger perceived self–other similarity with older adults (H1d).

H2 Role manipulation hypothesis. Interacting with an AR filter that presents the user as a scientist will enhance cognitive performance, such that participants in the scientist condition will achieve higher scores on a timed mental arithmetic task compared to controls (H2).

H3 Intersectional hypothesis. The combination of age and profession cues will produce interaction effects beyond the sum of their parts. Specifically, participants embodying the older scientist identity may display both improved cognitive performance and reduced age-related bias, as the scientist role counteracts stereotypes of cognitive decline typically associated with aging (H3).

H4 Identification moderation hypothesis. The strength of embodied identification will moderate the effects of the AR manipulations, such that higher identification with the virtual self will amplify both cognitive and attitudinal outcomes (H4).

2 Related works

2.1 Ageism and its modern impact

Similar to racism or sexism, ageism—a term coined by physician Butler, (1980) — describes prejudicial attitudes toward older adults. Ageism can lead to discriminatory practices, both individual and institutional, that reduce older adults’ opportunities for a satisfactory life and undermine their personal dignity. Butler writes that ageism transforms “aging from a natural process into a social problem in which the elderly individual bears the detrimental consequences (Butler, 1980).” Ageism has been called the ultimate prejudice and last discrimination because everyone, if they are lucky to live a long life, may be a target of it (Palmore, 2001). Ageism can affect a person on an individual, institutional, and societal level in multiple ways that can have devastating impacts on their mental and physical health, opportunities, and overall quality of life. For example, Bodner (2009) brought up how the ageist stereotype that older employees are less efficient than younger ones can lead to discriminating hiring, mandatory retirement, and toxic environments. We can see this discrimination in action when reviewing Stypińska et al. (2023) interviews from Silicon Valley, an area dubbed the most ageist place on earth. During their investigations, tech workers reported that the industry considered people over 35 to be too old for the industry, and there was an unspoken rule that no-one over 35 would be hired. In another report, Pillemer et al. (2021) proposed a conceptual model to outline how ageism can contribute to elder abuse. As ageism has been seen as a likely contributing factor to younger adults neglecting, exploiting, and maltreating their elders (Quinn and Tomita, 1997). Levenson (1981) explored how ageism was a key factor in medical mistreatment finding that the medical field implicitly trains doctors to treat older patients with bias, apathy, and even disdain. In their review of ageism studies, Nelson (2005) notes that medical ageism can lead to doctors seeing older patients as untreatable despite the fact that many diseases associated with ageism can be easily addressed. Ageism only seems to be growing as of late with the digital age and stereotypes about older adults and technology fueling it (Comunello et al., 2023). However, digital technology has shown itself to be a new tool for combating societal biases. For example, VR and AR systems have been constructed to utilize embodied identification with virtual avatars to reduce societal biases against real-life social groups.

2.2 Identification with virtual characters and avatars

Identification is the mechanism through which individuals receive, experience, and interpret information from media (Cohen, 2001). Initially, the concept was used to describe how audiences could emotionally and cognitively connect with fictional characters, adopting their perspectives and temporarily sharing their points of view. With the advent of interactive digital media such as video games, this process evolved: users could now act through and control the characters they identified with, effectively merging self and avatar into a single experiential entity (Klimmt et al., 2010). Van Looy et al. (2012) termed this phenomenon identification through embodiment, or embodied identification. This merger may occur because embodiment itself entails a transformation of bodily boundaries (Murray and Sixsmith, 1999). Our internal sense of the body is inherently plastic and can be reshaped by sensory manipulations or altered perceptual contexts (Murray and Sixsmith, 1999). During embodiment, individuals may experience a temporary “corporeal shift” in which the virtual body becomes integrated into their body schema, fostering an empathetic link between self and avatar. Farmer and Maister (2017) showed that it is the act of controlling a virtual character that produces such embodied identification, leading the virtual body to be incorporated into the user’s conceptual self. Building on this, Klimmt et al. (2010), Klimmt et al. (2009) found that users controlling a soldier avatar subsequently displayed stronger mental associations with traits such as bravery, strength, and patriotism. On a smaller non-gaming related scale, Fribourg et al. (2021) demonstrated that subtle manipulations of facial traits using AR Snapchat filters could shift self-perception—for instance, participants rated themselves as more intelligent or competent when wearing a virtual glasses filter.

However, the effects of embodiment extend beyond self-perception. Banakou et al. (2020) demonstrated that embodiment can influence motor resonance mechanisms: white participants embodying Black avatars unconsciously mimicked the movements of a Black confederate more than those of a white confederate, a form of social mimicry sometimes referred to as the “chameleon effect.” In another study, Banakou et al. (2018) found that participants embodying an Albert Einstein avatar showed enhanced cognitive processing during a logic task, leading to higher performance scores. Similarly, Guegan et al. (2016) reported that participants who embodied “inventor” avatars during a brainstorming session produced ideas that were both more fluent and more original. Comparable effects have been observed with AR technologies. In an experiment using Zoom-based AR filters, Leong et al. (2021) found that participants embodying a “baby” or “inventor” version of themselves achieved slightly higher creativity scores on verbal Alternate Uses and Verb Generation tasks. These findings collectively indicate that embodied identification can occur across a wide range of VR and AR systems, and that such embodiment can foster positive cognitive and attitudinal outcomes (Genay et al., 2021). Yet, the effects are not universally beneficial. For instance Peck et al. (2018) replicated the phenomenon of stereotype threat in VR: male participants embodying female avatars performed worse on visuospatial working-memory tasks, suggesting that virtual embodiment can also amplify negative social stereotypes. Given the ubiquity of AR filters in daily communication—particularly on platforms such as Snapchat and TikTok—these findings raise important questions about their broader societal impact. Millions of users interact with such filters daily, often without awareness of their potential cognitive or social effects. Despite this global reach, most empirical work has focused on immersive VR contexts (Banakou et al., 2018; Farmer and Maister, 2017; Lopez et al., 2019), while studies of short, socially embedded AR interactions remain rare. Moreover, when AR filters are examined, research typically emphasizes changes in self-perception or individual behavior (Leong et al., 2021; Miller, 2019; Youn, 2019; Alsaggaf, 2021) rather than their potential influence on broader social biases and intergroup attitudes.

2.3 Reproducing and expanding VR/AR bias research

Embodied identification has a long and well-documented history as a means of reducing aggression and prejudice rooted in implicit bias (Kishore et al., 2021). Jarrell (2020) conducted several studies examining how identification with video game characters influences players’ implicit attitudes. Their findings showed that even brief interactions—as short as two minutes—can induce embodied identification and significantly alter racial and gender-based biases. Moreover, their subsequent work demonstrated that embodied identification operates in an intersectional manner, affecting multiple dimensions of bias simultaneously (Jarrell et al., 2021). This is particularly relevant because most VR and AR studies of embodied identification have focused on long and deeply immersive experiences.

For example, Gonzalez-Liencres et al. (2020) found that male participants who embodied a female avatar experiencing domestic abuse reported greater empathy toward victims and lower prejudice toward women overall. The authors attributed this shift to the fact that participants’ felt vulnerability and emotional engagement during the simulation, which was powerful enough for some to experience the fear of potential assault. Gonzalez-Liencres et al. (2020)’s work fits within a larger sub-field of embodied identification know as “empathic computing (Jinan et al., 2025)”. VR has long been explored as an “empathy-generating machine” that has the possibility to impact human rights attitudes (Buie, 1981). A claim which has been substantiated as a systematic review of VR head-mounted display systems conducted by Lee et al. (2024) have found that the technology can significantly impact ‘cognitive’ empathy. This effect has been used in a wide range of scenarios such as simulating visual impairments to promote empathy for the disabled (Guarese et al., 2023), to simulating the POVs of both aggressors and victims to educate middle and high school students on the impacts of bullying (Gu et al., 2022).

However, VR experiences that focus on evoking empathy have been widely criticized by marginalized people and researchers. Studies that focus on empathy as a metric for evaluating such systems have been argued to center the emotions of the participant placed in the simulation rather than promote true understanding with the simulated subject groups (Bollmer, 2017). For example, in a VR visual impairment study by Zwoliński et al. (2025) though 72% of participants believed “they could predict what may be difficult for a person with a vision disorder” it should be noted that none of the participants truly experienced what it was like to live with a vision disorder. They only experienced the effect of living without a vision disorder and then rapidly being given a temporary vision disorder. As Silverman (2015) warns when discussing non-VR blindness simulations, this can “mislead people about blindness, because it highlights the initial trauma of becoming blind rather than the realities of being blind.” These empathy experiences fail as they do not capture the fact that people living with disabilities will often master alternative techniques that could impact what they find to be difficult (Ubel et al., 2005). Furthermore, there is debate regarding if empathetic experiences benefit the groups being replicated within them. One meta-analysis revealed “there is little empirical evidence of a correlation between VR exposure and an increase in empathy that motivates pro-social behavior (Sora-Domenjó, 2022).” However, a more recent systemic review suggests empathetic computing designed effectively can enhance education for healthcare professionals and better engage students with historical lessons (Jinan et al., 2025). Regardless, effective simulations seem to require intense, highly researched, and/or detailed scenarios. These empathy affecting embodied simulations are difficult to replicate in the few seconds or minutes typical of AR embodied filter interactions.

Yet, there is growing evidence suggesting that short AR embodied filter interactions may reproduce some of the pro-social psychological effects that have been observed in longer, more immersive VR experiences. A study by Leong et al. (2021) using an AR filter depicting a scientist with bright white hair produced results similar to Banakou et al. (2018)’s Einstein and Guegan et al. (2016)’s scientist avatar VR experiments. Leong et al. (2021) found that even a short interaction with an AR filter could result in improved cognitive performance in verbal generation tasks. Together, these studies suggest that both VR and AR embodiments can modulate cognitive abilities through identification with intellectual personas. Crucially, Banakou et al. (2018) strengthened the argument that embodiment could reduce implicit ageism, an effect previously claimed by Oh et al. (2016). Their study found that embodying avatars representing average everyday elder adult individuals was more effective at reducing age-related bias than exposure to static images or videos of real-life elder adults (Oh et al., 2016). To date, however, no comparable AR studies have directly tested this bias-reduction effect. The work of Jarrell and Peillard (2023) provides preliminary evidence that identification with AR filters might relate to changes in perceived levels of bias. In their study men who embodied and identified with themselves while using a filter claiming to “turn them into a woman” reported lower levels of agreement with gender stereotypes about appearance and behavior. While this finding does not establish a causal link between AR embodiment and bias reduction, it demonstrates that AR filters can serve as a sensitive probe for attitudinal change. Importantly, not all VR embodiment studies yield positive outcomes—some have even reported increased bias following identification—underscoring the need to examine both the potential and the pitfalls of AR-based embodiment paradigms.

Lopez et al. (2019) found that embodiment can also amplify bias under certain conditions. In their study of embodied identification during a virtual Tai Chi lesson, participants embodying female avatars exhibited significantly higher implicit gender bias than those embodying male avatars. The authors attributed this increase to stereotype activation—a process akin to stereotype threat—whereby existing negative stereotypes are triggered rather than mitigated. In this case, the stereotype that women are less athletic or physically competent may have contributed to participants’ reinforced gender bias. Similarly, Banakou et al. (2020) reported that the valence of the embodiment experience—whether positive or negative—plays a critical role in determining its outcome. White participants who embodied a Black avatar and subsequently faced negative social feedback from other virtual characters exhibited stronger racial bias afterward, suggesting that adverse experiences can lead participants to psychologically reject their virtual bodies and, by extension, the social group those bodies represent. Together, these findings underscore that embodied identification is a double-edged phenomenon: it can both reduce and reinforce bias, depending on context, affective tone, and social feedback. This ambiguity is especially concerning given the widespread use of AR filters that allow millions of mobile users to adopt virtual identities tied to real social categories—including gender, race, and age—on platforms such as Instagram, TikTok, and Snapchat (Haines, 2021; Dean, 2022; Dixon, 2022). Despite their massive reach and cultural influence, the potential of these short, everyday AR interactions to alter implicit or explicit biases remains largely unexamined.

Building on this body of work, the present study extends findings from Leong et al. (2021) by examining whether embodied identification through AR filters can also influence age-related attitudes and cognition. While Leong et al.’s AR study demonstrated that embodying a scientist filter enhanced creativity and cognitive performance, it did not address bias directly. Nonetheless, the scientist filter used in their experiment featured distinctive white, unruly hair reminiscent of Albert Einstein—a trait that symbolically links scientific expertise with older age. Given that white hair is also a salient visual marker of older identity, it is plausible that such filters could simultaneously activate age-related representations. This observation aligns with Oh et al. (2016) and Banakou et al. (2018), who showed that embodying older avatars in VR reduced implicit ageism and improved cognitive performance. If similar outcomes can be reproduced using short, smartphone-based AR interactions, it could lower the barriers to implementing bias-reduction interventions. Such experiences could be delivered remotely—via a simple Zoom call or mobile app—without the need for expensive and logistically demanding VR headsets.

To explore this possibility, the present study combines principles from Banakou et al. (2018), Oh et al. (2016), Guegan et al. (2016), and Leong et al. (2021). Participants engaged in a brief interactive task while experiencing embodied identification through one of several Snapchat filters designed to manipulate perceived identity. Unlike previous work where the “scientist” identity was inherently linked with an older adults appearance (e.g., white hair and wrinkles), we separated these traits into distinct conditions. One filter portrayed participants as older adults, another as scientists, and a third combined both manipulations, allowing us to isolate and compare the effects of each. This approach enabled us to test whether short, accessible AR experiences could replicate the cognitive enhancements and bias-reduction effects previously observed in immersive VR contexts. Specifically, we examined how brief embodied identification with an older adult, a scientist, or an older scientist AR filter influenced both cognitive performance and implicit and explicit attitudes toward older adults.

3 Methods

The present study aimed to replicate and extend previous VR and AR research on embodied identification, ageism, and cognitive performance. Unlike earlier work employing fully immersive VR systems or prolonged AR sessions, we used brief, smartphone-based AR interactions to explore whether short exposures could produce comparable effects. In addition, the study adopted an intersectional framework, examining how virtual embodiment along two identity dimensions—age and profession—might interact to shape bias and cognition. Participants were randomly assigned to one of four groups: a no-change control condition, a scientist condition, an older adult condition, and an older scientist condition. The randomized assignment was counter balanced throughout the running of the study. Initial assignments were randomized, then randomization percentages were manipulated if a particular group was overrepresented. Breakdowns of the final participant groups were; no filter-19, elder filter-32, scientist filter-29, elder scientist filter-25. Each group had the minimum number of participants for comparison (Gall et al., 1996). This design allowed us to isolate and compare the unique and combined influences of these social identities on both cognitive performance and age-related bias.

To evaluate these effects, we used both implicit and explicit measures. Implicit age bias was assessed through an Implicit Association Test (IAT) comparing associations between the concepts “young” and “old” with positive and negative traits. Explicit attitudes toward older adults were measured through Likert-scale questionnaires assessing warmth, competence, and general evaluations of older people. Participants also reported their willingness to interact with older adults and the degree to which they perceived similarity between themselves and older individuals. Cognitive performance was assessed using a time-constrained mental arithmetic task (a series of true and false mathematical problems) to capture effects on attentional and working-memory processes. Finally, participants rated their level of embodied identification with the AR filter, reflecting the extent to which they felt ownership, agency, and self-location in their virtual appearance.

3.1 Materials and applications

The study was conducted using an online survey combined with scannable Snapchat QR codes that granted access to custom AR filters. Each participant used Snapchat either on their personal smartphone or on a device provided by the experimenters. All filters were created by the researchers using Lens Studio, Snapchat’s official development platform. The older adult filters applied a semi-transparent 3D facial mask that tracked and moved according to the user’s head position. The model was designed from a heavily modified image of Albert Einstein which was edited to remove the iconic mustache and hide the more masculine features so that mainly the wrinkles and folds of skin would appear on a participant’s face. This way the aging effect could be applied to participants of different races and genders as seen in Figure 1. This approach echoes the study by Banakou et al. (2018) which used an Einstein avatar. This makes the figure less recognizable and more generic, in line with Oh et al. (2016). In-fact the modified filter was tested to be relatively applicable across multiple races and genders so the filter does not bias participants sharing the same race or gender as Einstein. Additionally, the filter applied a whitening effect to the user’s hair to simulate aging, similar to the visual manipulation used by Guegan et al. (2016). The scientist filters overlaid a 2D image of a white lab coat and blue bow tie across the user’s upper torso. This design choice aimed to evoke a general cultural, perhaps even cartoonish, stereotype of a scientist without anchoring it to a specific profession or environment. To reinforce the scientific theme around the face—the area most salient to users—we added a 3D model of safety goggles that tracked head movements in real time. Together, these elements created four distinct experimental conditions: a control condition with no visual manipulation, an older adult condition, a scientist condition, and an older scientist condition combining both overlays. Examples of these filters are shown in Figure 1. Participants completed questionnaires on their demographics, past Snapchat & filter use, perceived level of identification, perceived math performance skill, and three different measures for explicit elder bias. Questionnaires either came from or were modified from existing validated studies and Confirmatory Factor Analysis (CFA) tests were conducted on relevant factor questionnaires. During the mental math tasks participants answered two different series of 60 true or false math questions which were modified from an online mental math quiz and tested to ensure similar levels of difficulty between the two tests. A breakdown of the questionnaires and the mental math questions used for the tutorial game and the experimental games can be found in Tables 1–6 in the SUPPLEMENTAL DATA section. Furthermore a copy of all the study questions and forms can be found along with all used data from this study. For more information please see the OPEN ACCESS TO RESEARCH DATA section.

TABLE 1. Part 1 of the Full Questionnaire with all survey items.

TABLE 2. Part 2 of the Full Questionnaire with all survey items.

TABLE 3. Part 1 of the math questions used for the (1) tutorial true/false mental math filter game. Showing the question, the answer, and the percent that correctly answered the question during the pre-study test to determine question difficulty.

TABLE 4. Part 2 of the math questions used for the (1) tutorial true/false mental math filter game. Showing the question, the answer, and the percent that correctly answered the question during the pre-study test to determine question difficulty.

TABLE 5. Part 1 of the math questions used for the (2) experimental true/false mental math filter game. Showing the question, the answer, and the percent that correctly answered the question during the pre-study test to determine question difficulty.

TABLE 6. Part 2 of the math questions used for the (2) experimental true/false mental math filter game. Showing the question, the answer, and the percent that correctly answered the question during the pre-study test to determine question difficulty.

3.2 Participants

Participants could complete the study either remotely or in person. Those who participated online accessed the experiment through a Qualtrics survey distributed via email, WhatsApp, and public recruitment platforms such as SurveySwap and SurveyCircle. A power calculation was done using two f values (f = 0.270 and f = 0.465) from two similar past studies. Using G*Power our sample size estimates were generated and rounded up to 55–154 to account for possible invalid participants. Our study ended up with a total of 117 valid volunteers (108 remote, nine in-person) participating in the study. To ensure data quality and engagement, several attention-check questions were embedded within the survey. Each participant interacted with one of the Snapchat filters and completed a short arithmetic game. The application automatically generated a unique 14-character alphanumeric code based on the participant’s assigned condition and performance on the cognitive task. Participants were instructed to enter this code, along with their numerical test score, into the Qualtrics form. The encoding system, which participants were unaware of, was designed as a checksum-like mechanism: the code could be decoded by the researchers to verify that the reported score matched the actual in-app results. This procedure allowed us to detect random or inconsistent entries, ensuring participants genuinely engaged with the tasks. Participants whose codes were invalid or inconsistent with the expected outputs were excluded from analysis. Additional exclusion criteria were applied to detect low-effort responses in the Implicit Association Test (IAT), following the IAT’s internal response quality algorithm. Because the study offered no monetary compensation, the risk of disengaged participation was low (Qiao et al., 2021; Mason and Watts, 2010). After applying all quality controls, data from 12 participants were removed from analysis. For the study, 105 (34 male, 71 female) participants were used for data analysis.

The study was global but a majority our respondents listed France 19%, The United Kingdom 19%, the United States 16%, Italy 6%, Sweden 4%, or the Netherlands 4% for one of their nationalities. Of our participants, 60% were aged 18–24, 26% were aged 25–30, 6% were aged 31–34, 2% were aged 35–40, and 6% were 41 and above. We also asked for participants’ level of completed education. Most of the participants (83%) had a college associates degree or higher, with 41% having a Bachelor’s degree and 36%, a Master’s or PhD. A participant’s study type (in-person vs. remote) and various demographics were analyzed for possible impact on results. No significant differences were found between demographic groups except elder participants were less likely to agree with negative perceptions of elder adults (SE = 0.114, z = − 2.93, p = . 003) Still, participants were evenly distributed among the experimental groups so this difference did not have a significant impact on the final results. We also asked participants Snapchat related questions. Such as “On average how much time do you spend on image sharing social media platforms like SnapChat or Instagram?” Overall 6.5% said ‘I never go on such platforms’, 18% ‘1–2 h per week’, 24% ‘3–5 h’, 25% ‘6–10 h’, 20% ‘11–20 h’, and 6.5% ‘21–40 h’. Lastly we asked participants “Do you use image filters or image manipulation applications to modify your pictures?” In our sample 26% of participants reported they had ‘never or basically never’ used them, 50% had used them ‘a couple of times but not frequently’, 6.5% used them ‘about half of the time’, 12% used them ‘frequently but not all the time’, and 5.5% used them ‘on nearly all’ their photos. The study was reviewed and approved by the IMT Atlantique Ethics Board.

3.3 Procedure

A full overview of the experimental workflow is presented in Figure 2. After providing informed consent, participants first completed a demographics questionnaire and a brief survey on their prior experience with Snapchat and AR filters. All participants then engaged with a baseline AR filter designed to produce no visual alteration, allowing them to familiarize themselves with the interface and game mechanics. Participants were informed they were taking part in a study on filter use and cognitive performance. They were given a general breakdown of the purpose of the study and informed both before and after the study that they could drop-out or remove their data from the study at any point for any reason.

To encourage bodily engagement and initial identification, participants were instructed to perform three facial movements—smiling, raising their eyebrows, and opening their mouths—for 10 seconds each. The system only progressed when the expected facial movement was detected, ensuring that all participants completed the brief embodied identification exercise. This 30-s calibration task was conceptually inspired by full-body embodiment routines commonly used in VR studies (Banakou et al., 2016; Lopez et al., 2019), but adapted to the short and spontaneous nature of typical social-media filter interactions. Its purpose was to make users aware of their control and ownership over their AR self while maintaining ecological validity.

Following the tutorial test, participants then completed a tutorial mental arithmetic game implemented within the AR interface. The task consisted of true-or-false equations of varying difficulty, adapted from an online math IQ game. Although not an official intelligence measure, mental arithmetic has long been used as a proxy for fluid intelligence and working memory capacity (Baddeley, 2012). It is also sensitive to self-representation and embodiment effects: for example, Banakou et al. (2018) found that embodying an Einstein avatar improved reasoning performance, while Adam and Galinsky (2012) showed that wearing symbolic professional attire can enhance attention and cognitive control. In this study, the math task served as a proxy for cognitive performance, analogous to the Tower of Hanoi reasoning task used in previous VR embodiment experiments. Participants answered 60 questions, each with a 12-s limit, interspersed with three short 10-s breaks after every 15 questions to prevent fatigue. The total task duration was approximately 12 min and 30 s. At the end of the test, participants received a fabricated “IQ” score (computed as twice the number of correct answers plus 30 points). These inflated results obscured the true scoring logic, thus reducing the chance of poor results negatively impacting bias (Lopez et al., 2019; Silverman et al., 2015). A 14-character alphanumeric Base-36 code encoded each participant’s performance and filter version, enabling verification of data integrity. All participants completed the math test twice. This first test under the baseline (no-change) condition was a tutorial to introduce them to the system.

After the test, participants completed an Implicit Association Test (IAT) assessing baseline implicit age bias. We reproduced the validated Harvard Age IAT (Project Implicit, 2011) using the IATgen framework (Carpenter et al., 2019). The test measured associations between “young people” and “old people” with positive and negative attributes. Positive scores indicated a bias favoring young–good over old–good associations, while negative scores indicated the reverse. The full attribute sets for each category are summarized in Table 7. After completing the IAT, participants reported their perceived identification with their unaltered AR self, their confidence in their performance, and their self-assessed mathematical ability. Then participants completed a second test in their assigned experimental condition (no filter (control), elder filter, scientist filter, and elder scientist filter).

TABLE 7. Implicit Association Test (IAT) categories and items.

Source: Faces from Project Implicit Age Attitude IAT, Nosek et al. (2007), used under Project Implicit terms.

Participants were randomly assigned to one of four experimental groups: (1) control (no filter), (2) older adult, (3) scientist, or (4) older scientist. Meaning the four groups were as follows: 1) No filter tutorial–no filter experimental, 2) No filter tutorial–elder filter experimental, 3) No filter tutorial–scientist filter experimental, 4) No filter tutorial–elder scientist filter experimental. After launching their assigned filter, participants repeated the same short facial interaction sequence—smiling, raising their eyebrows, and opening their mouths—for 10 seconds each. This facial interaction sequence was done for every filter interaction, control, elder, scientist, and elder scientist. This repetition ensured consistent engagement with the AR system and reinforced a sense of agency and ownership over their newly altered AR appearance before continuing the study.

Following this re-familiarization phase, completed the second iteration of the math test under their assigned AR condition, using a new but equivalently difficult question set. Figure 3 illustrates the interface during gameplay, rest breaks, and score reporting. After completing the second task, participants once again completed the age-bias IAT to measure post-condition implicit bias. They then rated their sense of identification, confidence, and perceived ability using the pre-validated Embodied Identification Scale (Van Looy et al., 2012).

Finally, participants completed a series of questionnaires designed to measure explicit age-related attitudes. These included: (1) a Perception of Older Adults scale and (2) an Interaction Intention scale adapted from Oh et al. (2016), both measured on seven-point Likert items (1 = “Completely Disagree” to 7 = “Completely Agree”); and (3) the single-item Inclusion of Other in the Self (IOS) scale (Gächter et al., 2015), which assessed perceived closeness with older adults (1 = “Completely Separate,” 7 = “Extremely Similar”).

All multi-item constructs (Embodied Identification, Perception of Older Adults, and Interaction Intention) were validated via Confirmatory Factor Analysis (CFA). The final retained factors and questionnaire items resulting from the CFA are presented in Table 8. Results demonstrated convergent validity (AVE > 0.50) and discriminant validity (AVE exceeding inter-factor correlations). The overall model fit was excellent: χ 2(41) = 68.43, p = . 005; RMSEA = 0.079, 90% CI [0.044, 0.112]; CFI = 0.993; TLI = 0.990 (Kline, 2015).

TABLE 8. Final analyzed factors and question items from the results of the standard Confirmatory factor analysis (CFA) Kline, (2015), for identification and explicit ageism testing factors.

4 Results

4.1 Implicit older bias impact

Using the software provided by IATgen.org (Carpenter et al., 2019), we computed the IAT results for both the first (baseline) and second (post-condition) sessions to assess participants’ implicit associations before and after interacting with the AR filters. To quantify the impact of each filter condition, we calculated a difference score (ΔIAT) by subtracting each participant’s baseline IAT score from their post-condition IAT score. This pre–post subtraction method is consistent with standard procedures for evaluating intervention effects in IAT research (Lopez et al., 2019; Banakou et al., 2016; Peck et al., 2013). In our scoring convention, positive IAT values indicated a bias favoring younger adults (pro-young bias), whereas negative values indicated a bias favoring older adults (pro-old bias). Consequently, a decrease or negative shift in ΔIAT values represented a reduction in implicit age bias. Overall, IATgen.org reported that participants’ average baseline IAT score was significantly positive (M = 0.49, SD = 0.29, t ( 106 ) = 14.88, p < . 0001), indicating a clear implicit preference for younger adults relative to older adults across the sample.

To examine the effects of the experimental manipulations and participants’ levels of identification, we conducted a two-way linear regression analysis with Filter Type (None, Older Adult, Scientist, Older Scientist) and Level of Embodied Identification as predictors. The latter variable was computed from participants’ responses to the five items of the embodied identification questionnaire. Specifically, we calculated a weighted composite score by summing each participant’s 1–7 Likert ratings across the five items, then centering these sums by subtracting the overall mean identification score across all experimental conditions. This centering procedure allowed the variable to represent participants’ relative level of identification compared to the average user in the study. This approach offered a more sensitive measure of identification effects within the context of our data. For instance, a participant reporting a mean identification rating of 3 on the 7-point scale would ordinarily appear below the midpoint of the scale; however, if the overall sample average was 2, this same participant would actually reflect a comparatively high level of identification. By using this deviation-from-mean weighting, we were able to model identification as a continuous factor interacting with filter type, accounting for natural variation in participants’ engagement with AR embodiment. A Shapiro-Wilk test was performed and showed that the distribution of the IAT score results did not depart significantly from normality (W = 0.99, p = 0.69) Results of the regression model revealed no significant interaction between identification and filter type on the change in IAT scores (ΔIAT). Identification did not significantly predict implicit bias change for any of the experimental filters: Older Adult (SE = 0.01, t = 1.31, p = . 19), Scientist (SE = 0.01, t = 1.04, p = . 30), or Older Scientist (SE = 0.02, t = 0.76, p = . 45). A more in-depth breakdown of the stats can be seen in Table 9. Consequently, none of our hypotheses regarding the effect of filter embodiment on implicit bias (H1a, H2, H3, H4) were supported.

TABLE 9. IAT Score Change by Filter Group and Level of Embodied Identification stats.

Regardless of the filter condition or participants’ level of identification, changes in implicit bias (ΔIAT) between the baseline (no-change) and post-experimental sessions did not differ significantly across conditions. In other words, interacting with filters that altered age, professional attire, or both did not produce measurable changes in implicit bias compared to the control condition, in which participants completed a second IAT after using an unmodified filter and performing the same cognitive task. Figure 4 illustrates the ΔIAT results, showing the difference between post-experimental and baseline IAT scores for each filter condition as a function of participants’ reported level of embodied identification. No clear trend was observed across levels of identification, and the slopes for all filter types remained non-significant. Figure 5 presents the overall IAT scores before and after the experimental interaction. Although participants in the Scientist filter condition appeared to exhibit a slightly reduced pro-young bias following the manipulation, this difference did not reach statistical significance and thus did not support a measurable reduction in implicit ageism. In both figures, embodied identification is displayed on a normalized 7-point scale corresponding to the questionnaire items, ranging from −3 (Extremely Low Identification) to +3 (Extremely High Identification), with 0 representing the group mean (Neutral Identification).

4.2 Explicit older bias impact

To analyze explicit attitudes toward older adults, we employed a Structural Equation Model (SEM) rather than a linear regression approach, allowing the multi-item constructs to be treated as standardized latent factors. The SEM included a two-way interaction between the Filter Group (None, Older Adult, Scientist, Older Scientist) used during the second experimental phase and participants’ reported Level of Embodied Identification with that filter. A Shapiro-Wilk test was performed for the averages of Perceptions of Older Adults, their Intentions to Interact with Older Adults ratings, and the Similarity with Older Adults rating. They showed that the distribution of Perceptions of Older Adults did not depart significantly from normality (W = 0.98, p = 0.21) however Intentions to Interact with Older Adults (W = 0.97, p < . 01) and Similarity with Older Adults (W = 0.94, p < . 01) did depart significantly. The interaction with Level of Embodied Identification did not significantly predict either participants’ Perceptions of Older Adults. Using a robust estimator (WLSMV) due to non-normality we also found that Level of Embodied Identification did not significantly predict Intentions to Interact with Older Adults.

Since Similarity with Older Adults was found to be significantly different from normality a Kruskal-Wallis test was used to examine the filter group differences. It revealed there was no significant difference between the groups (χ 2(3) = 4.28, p = . 232). Due to the reported non-normality, interaction effects were explored by stratifying participants into low vs. high identification groups and testing filter differences separately. So, to examine if there was an interaction effect between the filter groups and Embodied Identification we created a median split on Embodied Identification (Low vs. High), then conducted two Kruskal-Wallis tests on the two groups. Both the Kruskal-Wallis tests for the High Embodiment (χ 2(3) = 5.28, p = . 15) and Low Embodiment (χ 2(3) = 0.60, p = 0.90) groups were not significant. Suggesting there was no significant difference between the filter groups alone or modified by idenfication as illustrated in Figure 6, which plots the Filter × Identification interaction, and Figure 7, which displays the main filter effects. However, the graphs suggested a potential main effect of the scientist manipulation regardless of age-related changes. In order to ensure we completed a thorough review of all possible effects we conducted a Mann Whitney U test comparing two groups; participants who used filters featuring scientific attire (Scientist and Older Scientist) versus those who did not (None and Older Adult). It revealed there was no significant difference between the groups (W = 1347.5, p = 0.08). Due to the reported non-normality, interaction effects were again explored by stratifying participants into low vs. high identification groups. Both the Mann Whitney U tests for the High Embodiment (W = 264.5, p = 0.081) and Low Embodiment (W = 415.0, p = 0.47) groups were not significant. Suggesting there was no significant difference between the filter groups alone or modified by identification. Figure 8 illustrates the filter group difference and Figure 9 illustrates the Embodied Identification by Filter Group difference. These findings did not support our hypothesis that virtual aging would reduce explicit age bias (H1d).

4.3 Cognitive performance impact

To assess cognitive performance, we compared participants’ scores on the experimental math test to those obtained during the initial tutorial session, following the same analytical approach as for the IAT. Although the two tests contained different items, their difficulty levels were pre-validated to ensure equivalence. Specifically, a separate pilot study was conducted with over ten independent participants who answered all available questions. The frequency of correct responses for each question was then analyzed, and questions of comparable difficulty were distributed between the two versions of the test. Easier questions appeared earlier and more difficult ones later, preserving a similar progression of challenge across both versions of the test (tutorial vs. experimental). The questions in the two test groups were completely different with no repeat questions. This way no participant would be able to remember a question from the tutorial section and improve their answer in the experimental section. As all participants completed both the tutorial and experimental math tests, and all math tests had an easy to hard progression, with all questions tested to ensure the progression from easy to hard questions within each test was statistically equal, any learning/practicing bias was equally applied to each experimental condition group. Thus nullifying any learning/practicing bias effects participants could have experienced.

Cognitive improvement was quantified by subtracting each participant’s tutorial score from their experimental score, yielding a difference measure (ΔPerformance) that reflected whether performance improved, remained stable, or declined during the second test. This approach allowed us to examine whether exposure to a particular AR filter affected participants’ likelihood of improvement between sessions. In addition, participants’ self-reported confidence and perceived mathematical ability were compared between the two testing phases to evaluate potential meta-cognitive or self-perceptual effects of the filters. A Shapiro-Wilk test was performed and showed that the distribution of the IQ score results did not depart significantly from normality (W = 0.98, p = 0.20).

Similar to the implicit and explicit bias analyses, we first conducted a two-way linear regression analysis to examine the impact of filter condition x level of identification on cognitive performance. Identification with the Older Adult (SE = 0.41, t = 0.27, p = . 79), Scientist (SE = 0.41, t = 1.63, p = . 11), and Older Scientist (SE = 0.47, t = 1.14, p = . 26) filters did not significantly predict changes in cognitive performance under this two-way divided statistical set-up. A more in-depth breakdown of the stats can be seen in Table 10. However, visual inspection of Figure 10 suggested that participants identifying with the scientist filters showed a pattern of improved performance relative to non-scientist conditions. To test this observation, we conducted a two-way linear regression comparing Scientist (Scientist and Older Scientist) versus Non-Scientist (None and Older Adult) filters, with Level of Embodied Identification as a continuous moderator. This model indicated that stronger identification with a scientist-related filter significantly predicted higher gains in math scores (S E = 0.30, t = 2.03, p = . 04). However, as shown in Figure 11, this effect appeared to be influenced by a single high-performing outlier. After removing the outlier, the interaction between filter type and identification level was no longer significant (S E = 0.31, t = 1.63, p = . 11) as shown in Table 11. Though the main effect of identification remained significant (S E = 0.15, t = 2.10, p = . 04). Further exploratory analyses removed two-way linear regression division of the filters and focused on simple one variable linear regression models. Where only identification as the manipulating variable for cognitive improvement was analyzed with each filter group. Consistent with our hypotheses(H2,H4), higher embodied identification significantly enhanced math performance for the Scientist (S E = 0.29, t = 2.00, p o n e - w a y = . 03) and Older Scientist (S E = 0.29, t = 2.00, p o n e - w a y = . 03) groups, but not for the Older Adult (S E = 0.33, t = 0.49, p = . 63) or Control (S E = 0.31, t = 0.18, p = . 86) groups, as shown in Table 12. When considering only participants in the Scientist condition, the relationship between identification and performance reached conventional significance (S E = 0.21, t = 3.02, p < . 01), as illustrated in Figure 12 and Table 13. Model coefficients indicated that for users of scientist filters, a one-point increase in embodied identification predicted an average improvement of 0.62 points in math score. Given the adjusted identification range (i = 7–35, centered at M = 14.67), this implies that a participant with maximal identification ( i = 35 ) could be expected to improve their performance by approximately 14.7 points—equivalent to seven additional correct responses out of thirty, or roughly a 25% increase. In contrast, a participant with average identification ( i = 14.67 ) would be predicted to improve by only 2.7 points (one additional correct response). Although the overall group comparisons did not yield statistically significant differences, a consistent positive association was observed between embodied identification and cognitive improvement among participants using scientist-related filters. This association appeared strongest in the Scientist and Older Scientist conditions, but the effect size was sensitive to the inclusion of a single high-performing outlier. Taken together, these results do not provide causal evidence that embodying a scientist persona improves cognitive performance. However, they do indicate that higher identification with scientist-themed AR filters was associated with greater performance gains on the math task, even within brief, smartphone-based interactions. This pattern aligns with the proposed Role Manipulation and Identification Moderation hypotheses (H2, H4), but should be interpreted cautiously given the limited sample size and the influence of individual variability.

TABLE 10. IQ Score Test Differences by Filter Group and Level of Identification stats.

TABLE 11. Post-Outlier IQ Score Test Differences Between Non-Scientist (Control & Elder) and Scientist (Scientist & Elder Scientist) Filters by Level of Identification stats.

TABLE 12. IQ Score Test Differences for each individual Filter Group by Level of Identification stats.

TABLE 13. IQ Score Test Differences of only Scientist (Scientist & Elder Scientist) Filters by Level of Identification stats.

5 Discussion

5.1 Augmenting age and older bias

Past VR embodiment studies have shown that embodying and identifying with an older adult avatar can reduce age-based biases (Oh et al., 2016; Banakou et al., 2018). The present study sought to reproduce these effects using AR by virtually aging participants rather than immersing them in a fully controlled virtual body. Unlike in previous VR research, however, our manipulation of facial age did not significantly influence implicit or explicit measures of ageism. Several factors may account for this discrepancy, the most prominent being the brevity and simplicity of our AR experience compared to the more immersive and sustained exposure typical of VR embodiment protocols. Nonetheless, prior short-form AR studies have demonstrated that filter-based manipulations can influence perceived social biases—for instance, identification with gender-swapping filters has been linked to changes in perceived agreement with gender stereotypes (Jarrell and Peillard, 2023). This suggests that the null effect observed here may be specific to the emotional or cognitive nature of virtual aging. It is possible that the experience of appearing older was affectively negative for some participants. Anxiety toward aging is common, and seeing one’s aged face could have elicited discomfort or self-directed ageist reactions. Previous research indicates that when embodiment experiences are perceived as unpleasant or threatening, they may increase bias rather than reduce it (Banakou et al., 2020). Even in the absence of overt negative emotion, contextual factors such as task difficulty could have influenced participants’ reactions. Many participants reported finding the arithmetic task demanding, which may have reinforced societal stereotypes associating aging with cognitive decline (Butler, 1980; Bodner, 2009). Thus, rather than inducing empathy toward older adults, the combination of virtual aging and cognitive challenge might have implicitly primed deficit-based views of aging. Alternatively, participants may have transferred their frustration or discomfort from the task onto stimuli associated with the filter itself—an effect previously observed when negative embodiment experiences generalized to related out-group perceptions (Banakou et al., 2020).

Another possible explanation is that the manipulation inadvertently induced stereotype threat. Stereotype threat occurs when individuals become acutely aware of a negative stereotype about a social group they identify with, which can in turn affect their performance or attitudes (Peck et al., 2018). In this context, interacting with Snapchat—a platform widely associated with youth culture—may have heightened awareness of age-related stereotypes and activated stereotype threat. Technology use is frequently portrayed as a domain of young people, and older adults are often stereotyped as less competent or less adaptable to digital environments (Comunello et al., 2023; Stypińska et al., 2023). Prior work has shown that stereotype threat can paradoxically increase bias expression when participants are reminded of negative group associations. For example, Groom et al. (2009) found that racial bias increased when individuals experienced racially charged interview scenarios, and Schulze et al. (2019) reported heightened gender bias when participants embodied female avatars in work environments that implicitly activated gender stereotypes. Although we expected, following Banakou et al. (2018) and Leong et al. (2021), that aging-related cues such as white hair or wrinkles might foster empathy and reduce bias, it is plausible that the mere act of using a youthful, image-based app like Snapchat primed the opposite effect. By situating older-looking faces within a platform strongly coded as youthful (Dixon, 2022), the experience may have amplified rather than mitigated awareness of intergenerational stereotypes. At this stage, it remains unclear whether the lack of observed bias reduction reflects this stereotype threat mechanism, the brevity of the interaction, or deeper differences between AR and VR embodiment. What seems likely is that AR-based virtual aging and VR avatar embodiment represent distinct psychological experiences: while both rely on visual self-modification, only the latter may consistently produce the immersive identification required for durable bias reduction. Developers and researchers should therefore be cautious in assuming that AR filters can replicate the full psychological impact of VR embodiment. Nevertheless, our null results contribute valuable insight to the emerging literature on AR-based self-representation, highlighting that some effects—such as gender-bias modulation (Jarrell and Peillard, 2023) or creativity enhancement (Leong et al., 2021)—may transfer across modalities, whereas others, like age-bias modulation, may require more prolonged or contextually tuned interactions. This study thus provides a useful reference point for future work exploring the boundaries of embodied identification in augmented reality.

5.2 The impact of virtual clothes

Virtual aging in AR did not significantly affect implicit or explicit measures of ageism. Instead, from the data graphing it seemed as if embodying a scientist persona through AR filters would most likely have an impact on participants’ explicit perceptions of older adults. With scientist filter users showing higher average scores of Similarity to Elder ratings. However, further analysis revealed no significant difference as a result of virtual costuming. This is in contrast with previous studies which showed dressing as an elderly person, especially an older scientist could significantly impact elder bias (Banakou et al., 2018; Oh et al., 2016). Embodying a scientist avatar in VR has been shown to produce notable psychological effects, even when the clothing is virtual rather than physical (Guegan et al., 2016). While Leong et al. (2021) showed that prolonged AR filter interactions designed to simulate professional or creative identities could influence users, the present study suggests this effect is not guaranteed within AR systems. The results of the present study seem to suggest that brief, low-fidelity interactions—lasting only a few seconds, like those common on social medial platforms—cannot elicit measurable changes to biases from single session interactions. However, this does not mean these brief low-fidelity interactions are not capable of electing any measurable significant cognitive changes.

Although identification with the filter did not predict bias-related outcomes, it emerged as a meaningful predictor of cognitive performance. Participants who used the scientist filter and reported higher levels of embodied identification tended to achieve better results on the mental arithmetic task. Importantly, this improvement occurred despite no corresponding increase in self-reported confidence or perceived mathematical ability, suggesting that participants were not consciously aware of their enhanced performance. This pattern aligns with previous work showing that embodied identification can subconsciously modulate cognitive functions, potentially by increasing attentional control or self-focus (Jones et al., 2019; Van Stockum and DeCaro, 2014). Another possible mechanism is the activation of a positive stereotype threat, whereby internalizing a socially valued role—such as that of a scientist—temporarily boosts task engagement and performance (Peck et al., 2018; Shih et al., 1999). These findings reinforce the potential of AR filters as lightweight tools for enhancing learning and performance. Similar to prior creativity-based interventions (Guegan et al., 2016; Leong et al., 2021), our results suggest that symbolic digital self-representation can momentarily shape cognition in measurable ways. However, this cognitive enhancement effect also raises broader ethical and psychological questions. If embodying a scientist can enhance performance, could embodying stigmatized or negative roles have the opposite effect? For instance, would donning a virtual “prisoner” outfit increase antisocial associations, or could repetitive exposure to beauty filters amplify body dissatisfaction and cosmetic surgery intentions, as suggested in prior studies (Alsaggaf, 2021; Ramphul and Mejias, 2018; Eshiet, 2020; Brucculieri, 2018)? Our findings highlight that AR filters are not neutral—they can modulate users’ cognition and self-perception even after brief exposure. While our effects were immediate and short-lived, the long-term consequences of repeated interactions remain unknown. Future research should therefore examine the durability, accumulation, and contextual boundaries of AR embodiment effects, especially as filter use becomes increasingly integrated into everyday digital self-expression.

6 Limitations

The main limitation of our study was the time, scope, and involvement of the interaction. Participants only interacted with their AR selves for around 12 min, but in most of that time their focus was split between viewing their manipulated selves and taking the mental math test. Participants did not primarily dedicate their time to interacting in playful ways with their filtered selves. Furthermore, at not point did they communicate with others, as is common on social media apps that use AR filters such as Snapchat. So, they did not receive feedback about their AR selves. This is important because experiencing different treatment from other people based on appearance could have had a strong impact on participant’s biases. A key aspect of many bias reduction interventions includes exposing participants to the feeling of living as another person, including the daily struggles a person can experience (Gutierrez et al., 2014). The experiment was also limited to only changing the face and did not include a full body change like similar VR experiments (Oh et al., 2016; Banakou et al., 2018; Banakou et al., 2016; Gonzalez-Franco et al., 2020; Lopez et al., 2019). Furthermore, while we separated the aging and scientific apparel experiences, we basically treated both experiences as equal. That is to say, we treated the experience of seeing yourself virtually aged as being comparable to seeing yourself virtually clothed. This may not be the case as each requires very unique manipulations. The scientist filter; for example, required multiple separate clothing pieces to work together (the goggles, lab-suit, and tie) while aging placed virtual wrinkles and white hair on the avatar. Was one experience more complex or more realistic? We did not explore this. Future studies could include a manipulation check on the perceived realism and affective valence of the filters to assess whether the “older” condition was perceived as positive, neutral, or negative. This would help determine whether the absence of significant effects stems from the filter’s technical limitations or from participants’ emotional reactions to their new self-representation. Also the study did not explore possible long term changes or how repeated interactions with the filter could impact users. It is far more common for Snapchat users to interact with a filter several times over their use of the system. Further research could explore if long term use of the filter impacts results.

7 Conclusion

This study sought to replicate prior research on VR-based embodiment and ageism reduction (Oh et al., 2016; Banakou et al., 2018) by translating it into an augmented reality context. Rather than allowing participants to control an older avatar, we virtually aged their own faces with AR filters, while also exploring how virtual role manipulation—specifically, dressing users as scientists—might influence cognition and bias. By isolating these two dimensions of virtual identity change (age and professional role), we were able to compare their respective and combined effects, an aspect not fully disentangled in earlier VR and AR work (Banakou et al., 2018; Leong et al., 2021; Genay et al., 2021). Our results indicate that, unlike VR-based aging embodiment, brief AR aging experiences did not significantly reduce age-related biases. However, participants who virtually “wore” scientific attire did perform better on a cognitive task—particularly when they reported higher levels of identification with their virtual appearance. These outcomes align with previous findings in both immersive and real-world contexts showing that what we wear, even virtually, can shape how we think and act (Adam and Galinsky, 2012; Van Stockum and DeCaro, 2014; Leong et al., 2021). Overall, the study underscores that even minimal AR interactions can modulate self-perception and behavior. The absence of effects from virtual aging highlights the limits of AR compared to fully embodied VR experiences, while the cognitive and attitudinal effects of virtual costuming emphasize the power of symbolic self-representation. These findings hold promising implications for the design of digital environments—from virtual classrooms and collaborative platforms to social or professional interactions—where AR filters could be leveraged to foster engagement and performance. As digital self-presentation becomes increasingly pervasive, both researchers and developers should remain mindful that virtual identities, however transient, can meaningfully influence how users perceive themselves and perform.

8 Open access to research data

In accordance with the Horizon 2020 Pilot on Open Research Data free access to all used data from this study will be made available on the lead researcher’s website: https://mariejarrell.wordpress.com/. Interested individuals can reach out to the lead researcher at m.a.jarrell@soton.ac.uk if they encounter any issues or require any further assistance.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Ivana Hromatko, University of Zagreb, Croatia

Reviewed by: Antonino Casile, University of Messina, Italy

Renan Guarese, Royal Institute of Technology, Sweden