An investigation of how gender shapes the appearance and judgment of apologetic faces

Abstract

Do people have mental representations of what apologetic faces look like? Do representations differ by gender? We used reverse correlation to (a) generate images that approximate mental representations of apologetic faces, (b) determine whether these images are rated highly on apology-related characteristics, and (c) see if ratings differ by gender of the image generator, target face, and/or image rater. Faces generated from male and female base faces to look apologetic were rated as more apologetic, remorseful, and sad than the base face, demonstrating these mental representations can be approximated using reverse correlation. Findings suggest visually represented apologies express multiple apology-related characteristics. Study 2 revealed the visual templates of faces generated by the gender ingroup appeared more apologetic than those of the gender outgroup; women-generated female faces were most apologetic, and men-generated female faces were least apologetic. Findings highlight gender differences in mental representation, but not perception, of female apologetic faces.

Apologies help to repair social bonds and improve relationships following transgressions (Bippus & Young, 2019; Schumann, 2018). Much of the apology research has sought to identify verbal components that are most likely to lead to positive relationship outcomes and forgiveness (e.g., Bippus & Young, 2019; Kirchhoff et al., 2012; Lazare, 2004; Lewicki et al., 2016; Tavuchis, 1991), and common themes have emerged. For example, acknowledgement of responsibility for the transgression and an offer to compensate for the wrongdoing are viewed as critical components of an apology (Lewicki et al., 2016). For transgressors, apologies relieve negative emotions (Riek, 2010), and are associated with less severe punishment (e.g., legal repercussions, penalties; Mungan, 2012). For those receiving an apology, however, accurately perceiving sincerity will help them avoid being transgressed against in the future (Lewicki et al., 2016). Although far less research has examined the nonverbal aspects of apology (cf. George et al., 2023; Hornsey et al., 2020; ten Brinke & Adams, 2015), evidence suggests that nonverbal and verbal cues interact to influence the perception of sincerity (Park & Guan, 2009; ten Brinke & Adams, 2015; Yamamoto et al., 2021). So, how do we know—both as the apologizer and victim—what to communicate nonverbally, how it should be expressed, and how to discern when it is sincere?

To begin to answer these questions, we first need to understand nonverbal apologies. That is, is there such a thing as an apologetic facial expression, and if so, what does it look like, and how is it perceived? Evidence from George et al. suggests that we hold a mental representation of an apologetic male face, and that it is perceived largely as sad (George et al., 2023). However, theory and research suggest that face perception differs based on the social group membership (e.g., gender) of the target and the perceiver (Hehman et al., 2017, 2019; Xie et al., 2019). Gender differences also exist in both the expression and perception of emotions (Deng et al., 2016; Fischer & LaFrance, 2015; Hall & Matsumoto, 2004) and apology (Polin et al., 2024; Schumann & Ross, 2010; Wei & Ran, 2019). Because of this, we wanted to extend previous research by examining the role of gender in the conceptualization and perception of apologetic faces. The main goals of the current research were therefore to use a reverse correlation method to approximate people’s mental representations of a female apologetic face, to determine whether the subsequently generated visual templates of male and female apologetic faces received high ratings on apology-related characteristics, and to test whether ratings differed depending on the gender of the image generator, the target face, and/or the person rating the image. By establishing a foundational understanding of nonverbal apologies, we will be better able to study and predict the circumstances under which apologies are optimally successful for both transgressors and perceivers.

Nonverbal Communication of Apologies

Although we tend to think about communication as mainly verbal, nonverbal channels are also used to get our message across to others. Nonverbal cues encompass a variety of techniques that can either be used alone or in concert with speech, and can strengthen or weaken one’s message (Nešic & Nešic, 2015). For example, changes in vocal tone can indicate tension or anxiety (Givens, 2015), and body language including avoidant gaze or expansive posture can detract from or contribute to one’s confidence in their message (Burgoon et al., 2016). When it comes to apologies, initial evidence suggests that embodying emotions such as shame and regret through crying (Hornsey et al., 2020) and displaying sadness (Claeys et al., 2013) can improve how apologies are perceived and judged. Although each of these cues contributes to our understanding of the nonverbal aspects of apology, facial expressions—described as the most influential form of nonverbal communication (Awasthi & Mandal, 2015)—have received far less attention in the apology literature (cf. George et al., 2023). In the current research, we aimed to build on this topic by approximating people’s mental representation of an apologetic face using a reverse correlation method to generate a visual template. Additionally, we sought to test whether these mental representations differ depending on gender.

Reverse Correlation

Reverse correlation methods seek to identify possible associations between features of a stimulus and how participants judge said stimulus (Todorov et al., 2011). Although different forms of the reverse correlation paradigm have been used in neurophysiology, audition, and vision research (e.g., Ahumada & Lovell, 1971; Gosselin, & Schyns, 2001; Okazawa et al., 2018), the approach used in the current research has become popular among social and cognitive psychologists as a way to estimate people’s mental representations of various constructs through the generation of a template or image (Dotsch & Todorov, 2012). In an initial image generation phase, participants are shown between 400 and 700 trials comprised of face pairings. They are instructed to select which face in each pair has more of a given characteristics (e.g., apologetic). Individual face stimuli differ by random noise patterns that are layered over a single base image known as the base face. Within each trial, faces are paired such that they have inverse or opposite random visual noise patterns (Dotsch & Todorov, 2012). The visual noise patterns from the selected images are then averaged, resulting in a final image known as a classification image (CI). This image is meant to approximate people’s mental representation of a face exhibiting the characteristic in question (Brinkman et al., 2017). Next, new participants, naïve to how the image was generated, are asked to rate the CI from the first phase. These ratings provide insight into how the visual template, or CI, is perceived.

There are several benefits of using reverse correlation. Demand characteristics are avoided because the differences between stimuli are difficult to detect across trials, and factors influencing participant responses are often implicit or beyond conscious awareness. Also, the process is data-driven, making it particularly useful for researching abstract constructs that are otherwise difficult to verbally articulate. For example, reverse correlation has been used to approximate people’s mental representation of faces communicating trustworthiness and dominance (Dotsch & Todorov, 2012), social ingroup and outgroup members (Bjornsdottir et al., 2019), friendship networks (Kunst et al., 2022), and mate attractiveness (Karremans et al., 2011).

Of greater relevance, George et al. (2023) recently used a reverse correlation method to approximate people’s mental representation of an apologetic face. Using a male base face, George et al. provided initial evidence that a visual template of an apologetic face could be generated using participant responses, with the generated image subsequently being rated high on apology and characteristics typically associated with being apologetic, including regret, remorse, and sadness. One question to emerge from these findings is whether people’s mental representation of a female apologetic face differs from that of a male apologetic face. It is also unclear whether men and women hold different mental representations and/or whether gender differences might emerge in the rating phase due to gender-based differences in perception of the visual templates generated using reverse correlation. These questions have theoretical and practical implications for our understanding of how people mentally represent apology, by elucidating the role of gender and gender stereotypes in perceptions and representations.

Gender and Mental Representations of Apology

We live in a gendered society, and theory and research suggest that gender shapes how we perceive ourselves and others (Diekman & Schmader, 2021; Martin & Mason, 2022; Morgenroth & Ryan, 2018). The historical and persistent power and privilege afforded to men over centuries have become ingrained in the collective unconscious, and a preference or bias for men has become automatic and implicit (Cheryan & Markus, 2020). Many researchers have argued that men represent the standard against which all other social categories are compared (Bailey et al., 2020; Hamilton, 1991; Hegarty & Buechel, 2006; Hegarty & Pratto, 2004; Merritt & Harrison, 2006). This implicit assumption that men are the default form is so robust that even shifting to the use of gender-neutral language has not eliminated pro-male bias (Lassonde & O’Brien, 2013).

Gender bias has contributed to disparities in work, wages, and voting, and these differences may have downstream consequences for how we think about apologies and conceptualize apologetic faces. Despite evidence that men apologize less than women (Schumann & Ross, 2010), they are more likely than women to be seen offering large-scale public apologies (Wei & Ran, 2019). This is because pro-male bias in the workforce has led to a greater proportion of men, relative to women, in high-status jobs such as CEO and scientist (Ceci et al., 2009; Eagly & Steffen, 1984; Wei & Ran, 2019). When groups, companies, or individuals transgress, apologies are increasingly being offered publicly by men who typically hold powerful and high-status positions (Cowen & Montgomery, 2020), leading to repeated exposure of apologetic men. Apologies are also thought to vary as a function of offense severity (Wohl et al., 2011), and because public admissions of wrongdoing reach a large audience of victims and nonvictims alike, public apologies likely contribute to implicit associations between men and severe transgressions (Wohl et al., 2011).

One benefit of a public apology is that it is likely to improve the chances of being forgiven, because willingly lowering oneself to a position of low power, dominance, and/or social status is associated with successful apologies (MacLachlan, 2013). Contrast effects suggest that it may also be easier to recognize submission—and therefore potentially apology—in high-status individuals because of the discrepancy with their typically highly dominant demeanor (Thayer, 1980). Together, these biases may lead people to have a clearer mental representation of an apologetic man as compared to an apologetic woman, and to perceive apology more easily on a male face.

Unlike men, women often occupy less visible and lucrative occupations, such as teacher or nurse (Ceci et al., 2009; Eagly & Steffen, 1984). These positions are typically low in status and power, which is consistent with gender stereotypes of women as subservient across domains such as social status, personality, and even body posture (De Lemus et al., 2012). Women are stereotyped as communal (vs. agentic) and warm (vs. competent; Martin & Slepian, 2021), and these relationship-focused qualities may help explain the finding that women apologize more than men (Schumann & Ross, 2010).

Additional support for the hypothesis that nonverbal apologies may differ across genders comes from face perception research. Findings suggest that as a neutral face is morphed to appear dominant, it is rated as more masculine, whereas when the same face is morphed to appear submissive, it is rated more feminine (Oosterhof & Todorov, 2008). Masculine faces also tend to have features associated with dominance, such as square or well-defined jaw lines and lower brows, while feminine faces often have more submissive features such as softer jaw lines and higher brows (Adams et al., 2015; Keating, 1985).

The same structural features associated with dominance and submissiveness can also be traced to gender-stereotypical emotional expressions. For example, sadness is thought to be associated with women more so than men, in part because the softer features of a prototypical woman are perceived as low in dominance—as are sad faces (Adams et al., 2015). Given that the visual template of an apologetic face created from a male base face (George et al., 2023) was found to be high on both submissiveness and sadness, it seems possible that a female face may be more readily rated as being apologetic. Women are also stereotyped as being expressive (Adams et al., 2015), and emoting functions as a socially sanctioned way for women, relative to men, to signal affiliation (Fischer & LaFrance, 2015). Given stereotypes that women are more submissive and less dominant than men (Adams et al., 2015), that women are emotionally expressive (e.g., Deng et al., 2016; Fischer & LaFrance, 2015), and that sadness and low dominance co-occur in face perception (Knutson, 1996), one might predict that apology would be perceived more easily on women, as compared to men, leading to higher apology ratings of an apologetic image generated from a female face.

Person perception research has also demonstrated that people’s own social group memberships, including gender, can influence perceptions. Women perceive faces as more expressive than men and are more accurate in identifying mixed emotions (Hall & Matsumoto, 2004). Although women and men identify individual emotions equally well, women are more attuned to subtle differences within a single emotion category (Hall & Matsumoto, 2004). In sum, the gender of both targets and perceivers can potentially influence apologies through people’s mental representations and perceptions of an apologetic face. We examined these possibilities in the current research.

The Present Research

The first goal of this research was to build upon previous findings (George et al., 2023) and further examine whether mental representations of male and female apologetic faces can be approximated using reverse correlation. We presented participants with stimuli derived from a female (Studies 1 and 2) and male (Study 2) base face to develop apologetic visual templates that were subsequently rated by an independent sample of participants on apology-related characteristics. We predicted that images created from these base faces would be perceived as significantly more apologetic than their neutral base face. Such findings would add to the literature suggesting that people have a mental representation of an apologetic face, and that such a face can be approximated using the reverse correlation procedure regardless of the gender of the target image.

The second goal was to replicate and extend previous findings suggesting people’s visual template of an apologetic male face is rated highly on apology-related characteristics, including sadness and submissiveness (George et al., 2023). If these characteristics are key components of an apologetic face, we predict that they would emerge regardless of the gender of the base face. By contrast, if seeing a face exhibiting counter-stereotypical characteristics, such as a sad and submissive male face, results in higher ratings on the characteristics because of contrast between the apologetic face and base face, we expect these findings would only replicate when using a male base face.

The final goal was to examine whether apology ratings are moderated by the gender of the image generator or image rater. Although analyses were largely exploratory, we tentatively predicted a few key patterns. First, because women perceive faces as more expressive and are more accurate in identifying mixed emotional expressions relative to men (Hall & Matsumoto, 2004), we hypothesized that an apologetic face generated by women would be rated as more apologetic than an apologetic face generated by men. Given that judgments can also be biased by group membership, it seemed possible that there would be a moderating effect of participant gender on face generation and ratings, with perceptions of gender ingroup faces differing from those of gender outgroup faces. These possibilities were examined in Study 2. Finally, we predicted that participants would demonstrate an ingroup bias in Study 2 such that faces generated by ingroup members would appear more apologetic, and that participants would rate ingroup faces as more apologetic than outgroup faces.

Materials, data, and syntax for both studies can be found on the Open Science Framework (OSF; https://osf.io/nme57/?view_only=7c9dc04ea5e94422ac0a80f3c51788a2).

Study 1

A female base face was created on WebMorph (DeBruine, 2017) by averaging eight randomly selected White female faces from the Karolinska Face Database (Figure 1; Lundqvist et al., 1998). This is the same database from which the male base face was created for previous reverse correlation studies (e.g., George et al., 2023; Oliveira et al., 2019; Petsko et al., 2021).

Figure 1.

Female base face.

Study 1a: Image Generation

Method

Participants and procedure

An initial sample of 118 participants was recruited from an undergraduate research participant pool at a large racially and ethnically diverse Canadian university. After excluding four people (n = 3 did not finish; n = 1 failed the attention check), the final sample included 114 participants (M_age = 19.68, SD = 2.92; women = 81, men = 31, other = 2) who received course credit for their participation. The racial/ethnic composition of the sample included participants who self-identified as South Asian (n = 25), White (n = 24), Middle Eastern (n = 15), Black (n = 13), East Asian (n = 10), Southeast Asian (n = 9), Latin or South American (n = 5), mixed race (n = 5), or a race that was not listed (n = 4). Sample size was determined to be adequate based on norms within the reverse correlation literature, which suggest that as few as 30 participants per condition are needed to produce high-quality resulting images (e.g., Bjornsdottir et al., 2019; Lick et al., 2013).

Participants completed the study online. After signing up, they were given a link to access the survey at their convenience. Participants provided consent, followed by demographic information. Next, they completed 400 reverse correlation trials. Each trial consisted of a pair of faces, one with a unique visual noise pattern superimposed on the base face, and the other with the inverse visual noise pattern superimposed on the base face. In each trial, both images were displayed on the screen simultaneously and participants were asked to select which face appeared more apologetic. Once all trials were complete, participants were debriefed and thanked for their participation.

Measures

Reverse correlation

Each trial of the reverse correlation paradigm consisted of two images. The images were generated using the “rcicr” package (Dotsch, 2017) in R (R Core Team, 2020) by superimposing random visual noise patterns on the original base face.

Conscientious Responders Scale

Due to the large number of trials and repetitive nature of the task, we included five questions from the Conscientious Responders Scale (CRS; Marjanovic et al., 2014) to break up the task and ensure participants were not responding randomly or using patterned responding. These questions were presented at random intervals and served as an attention check. Instructions directed participants to select a specific response (e.g., “In response to this question, please choose option number three, slightly disagree”). Participants who passed three or more attention check questions were included in the final dataset (Marjanovic et al., 2014).

Results

Using the “rcicr” package (Dotsch, 2017) in R (R Core Team, 2020), participant responses were averaged to generate a group-level CI of an apologetic female face (Figure 2).

Figure 2.

Apologetic female face.

Study 1b: Image Rating

Method

Participants and procedure

One hundred and twenty participants recruited from an undergraduate research participant pool completed the image rating phase of Study 1. Because this phase included a small number of ratings, we felt that the study was not long enough to warrant using an attention check, and therefore all participants were retained (M_age = 19.95, SD = 4.20; women = 94, men = 26). The racial/ethnic composition of the sample included participants who self-identified as South Asian (n = 33), White (n = 27), Middle Eastern (n = 21), Black (n = 12), Southeast Asian (n = 11), East Asian (n = 9), Latin or South American (n = 3), and mixed race (n = 4). A priori power calculations indicated that a sample size of 90 was required to detect an effect size of d = 0.30, which was based on findings from previous reverse correlation research (Lloyd et al., 2017) at 80% power. We oversampled by 30 participants because the study was conducted online and we were not including attention checks.

Participants began the study by providing consent and basic demographic information. Next, participants were shown either the apologetic CI generated in Study 1a or the original base face, ¹ and were asked to rate it on a scale from 0 (not at all) to 100 (very) on the following characteristics: apologetic, regretful, remorseful, sad, submissive, dominant, trustworthy. ² After rating the image on all characteristics, they repeated the procedure for the other image (i.e., the CI or base face, depending on counterbalancing condition). The order of characteristics was randomly presented.

Results

The White female apologetic face generated using reverse correlation was perceived as significantly more apologetic (M = 72.00, SD = 19.44) than the base face (M = 36.68, SD = 23.91), t(119) = 12.92, p < .001, d = 1.18. This provides evidence that participants have a mental representation of an apologetic face and that a visual template can be generated using reverse correlation to approximate such a mental representation (see George et al., 2023).

Results from multiple paired-sample t tests indicate that regretful, remorseful, sad, and submissive ratings were all significantly higher for the apologetic face than the base face, whereas dominance was lower for the apologetic face than the base face (Table 1). Ratings for trustworthy did not differ. Replicating previous findings with a male base face (George et al., 2023), the apologetic CI generated from this female base face received ratings on the characteristic “sad” (M = 81.65, SD = 22.23) that were significantly higher than ratings on the characteristic “apologetic,” t(119) = −5.26, p < .001, d = 0.48. ³

Table 1.

Ratings and comparisons of the apologetic classification image and base face: Study 1b.

	Image
	Apologetic face		Base face
Construct	M	SD	M	SD	t	d
Apologetic	72.00	19.44	36.68	23.91	12.92***	1.18
Sad	81.65	22.23	40.52	24.74	14.27***	1.30
Regretful	74.38	22.01	35.84	24.09	12.82***	1.17
Remorseful	67.95	24.38	37.69	23.16	9.88***	0.90
Submissive	60.95	22.63	49.42	22.43	4.16***	0.38
Trustworthy	48.16	22.44	50.96	22.70	−1.07	0.10
Dominant	28.64	22.00	48.47	22.96	−7.04***	0.64

Note. Bonferroni adjustments were made for p values and accounted for seven comparisons in total (i.e., comparisons of the apologetic and base faces for seven characteristics).

^***p < .001.

Discussion

Reverse correlation was used to generate a visual template of an apologetic female face. Results provide evidence that this technique can be used to produce an image of a face that was subsequently rated as more apologetic than the base face from which it was created. This pattern replicates and extends previous findings using a male base face (George et al., 2023), such that the apologetic female face received the highest ratings on the characteristic “sad,” followed by “regretful,” “remorseful,” “apologetic,” and “submissive.” Ratings on the characteristic “trustworthy” did not differ across the base and apologetic faces, and the apologetic face was rated as less dominant than the base face. The female apologetic face was rated particularly high on the characteristic “sad” relative to the base face, providing further evidence of the key role of sadness in nonverbal expressions of apology (George et al., 2023). Although the CI was rated as more submissive than the base face, the magnitude of this difference was less than what might have been predicted based on previous findings with a male face (George et al., 2023). This may be because, while expressions of sadness are gender-nonspecific, appearing submissive is counter-stereotypical for men but not women, and could therefore contribute to a stronger contrast effect for men. However, because only women’s faces were used in Study 1, direct gender comparisons could not be made. We sought to address this in Study 2.

Study 2

The first goal of Study 2 was to replicate Study 1 with faces of White women and men by examining whether an apologetic face could be created using reverse correlation from both a female and a male base face. To do so, we used a male base face derived from morphing and averaging men’s faces from the Karolinska Face Database (Figure 3) as well as the female base face used in Study 1 (Lundqvist et al., 1998). These base images were then used to produce new apologetic faces of a man and a woman using reverse correlation. We predicted that both faces would be rated as more apologetic than their respective base faces, and that both would receive high ratings on apology-related characteristics, specifically, “sad,” “regretful,” “remorseful,” and “submissive.”

Figure 3.

Male base face.

We also examined three potential moderators of apology ratings: the gender of the raters, the target image, and the image generator. We examined the possibility that the gender of the raters would influence perceptions of apology, with women perceiving a greater degree of all characteristics relative to men, who may not be as aware of subtle differences in expression due to gender differences in ability to identify complex mixed emotions (Hall & Matsumoto, 2004).

We also predicted a main effect of target gender; however, it was unclear how this effect would manifest. It was possible that the apologetic male faces would be rated as more apologetic than the apologetic female faces because they match a prototype developed from repeated exposure to high-powered men offering public apologies (MacLachlan, 2013; Wei & Ran, 2019; Wohl et al., 2012). Alternatively, the apologetic faces of women could be rated as more apologetic than the apologetic faces of men because they match a prototype developed from extensive and repeated exposure to women apologizing for even minor transgressions (Schumann & Ross, 2010).

The gender of the image generator was also of interest. Stereotypes suggest that women not only experience and express emotions to a greater extent than men (Plant et al., 2000), but they are also more attuned to emotion perception (Hall & Matsumoto, 2004). Therefore, one might predict that women are likely to generate more expressive faces than men. However, intergroup research has consistently shown that people hold ingroup biases and often attribute negative traits and characteristics to outgroup relative to ingroup members (e.g., Dotsch et al., 2008; Hewstone et al., 2002). This could suggest that, regardless of generator gender, other-gender faces would be judged more negatively. Importantly, however, it is unclear in the current context whether apologies are construed negatively (e.g., associated with transgressions) or positively (e.g., to repair relationships). Our test of the role of generator gender in apology ratings was therefore exploratory.

Study 2a: Image Generation

Method

Participants and procedure

A sample of 140 undergraduate students participated in the image generation phase of Study 2 in exchange for course credit. Three participants were removed because they responded randomly (n = 1), did not complete the study (n = 1), or did not indicate their gender identity (n = 1). ⁴ Therefore, 137 undergraduates (M_age = 19.82, SD = 3.89; women = 70, men = 67) were included in the final dataset. The racial/ethnic composition of the participants included White (n = 34), South Asian (n = 25), Black (n = 20), Middle Eastern (n = 19), East Asian (n = 14), Southeast Asian (n = 11), Latin or South American (n = 3), mixed race (n = 7), or another race not listed (n = 4). Because image generation was divided into four conditions (men generating male faces, men generating female faces, women generating male faces, and women generating female faces), we aimed to recruit a total of 120 participants based on norms within the relevant literature (e.g., Bjornsdottir et al., 2019; Lick et al., 2013).

The procedure was similar to the image generation phase in Study 1, with a few exceptions. First, participants of each gender were randomly assigned to one of two conditions (the base face of a woman or a man). Second, Study 2 was conducted in person, and participants completed the study on desktop computers located in a laboratory on campus. As many as 20 participants completed the study at the same time, under the supervision of a research assistant who ensured that participants worked quietly and independently. Each participant was given a condition-specific URL that took them to the website where the survey was hosted. After providing informed consent and demographic information, participants proceeded to the reverse correlation trials. Participants were thanked and debriefed upon completion.

Measures

Reverse correlation

The stimuli used in the current study were generated from White male and female faces. The female base face was identical to that used in Study 1. The male base face was sourced from the Karolinska Face Database (Lundqvist et al., 1998). All images were grayscale, and visual noise patterns were generated in R (R Core Team, 2020) using the “rcicr” package (Dotsch, 2017) and layered on top of the female and male base faces separately.

Conscientious responding

The CRS (Marjanovic et al., 2014) used in Study 2a was identical to that used in Study 1.

Results

The “rcicr” package in R (Dotsch, 2017) was used to create an average face for each participant based on their responses. These individual averages were then combined with same-gender participant averages within each condition, resulting in four CIs: a male face generated by men, a male face generated by women, a female face generated by men, and a female face generated by women (Figure 4).

Figure 4.

Male and female apologetic faces separated by generator gender.

Note. Images in Column A are the apologetic CIs generated in Study 2a by male participants. The apologetic CIs generated in Study 2a by women are in Column B.

Study 2b: Image Rating

Method

Participants and procedure

Two hundred and ninety-three surveys were completed by undergraduate students in exchange for course credit. Participants were removed from the data because they did not consent (n = 1), did not finish (n = 6), failed the CRS (n = 60), completed the study more than once (n = 27), or because they did not identify with a binary definition of gender (n = 1). After exclusions, the total sample size was 198 (M_age = 21.62, SD = 5.21; women = 98, men = 100). The racial/ethnic composition included South Asian (n = 50), White (n = 49), Middle Eastern (n = 26), Black (n = 20), East Asian (n = 19), mixed race (n = 16), Southeast Asian (n = 11), Latin or South American (n = 4), race not listed (n = 2), and Polynesian (n = 1). A sample of 175 was set as the minimum sample required to detect a medium effect of d = 0.30 at 80% power prior to data collection (Lloyd et al., 2017).

The procedure was similar to that of Study 1b, except participants in the current study rated six faces: the female and male base faces, the female apologetic CI generated by female and male generators, and the male apologetic CI generated by female and male generators. Each face was rated on all characteristics before moving on to the next, and the order in which the faces were presented was randomized. Participants were sent a link to the study upon signing up and asked to complete it online on a desktop computer. Five adapted CRS questions (Marjanovic et al., 2014) similar to those used in Study 1 assessed attention (see Supplemental Material for more details). Again, these were randomly selected and presented among the other trials at specific intervals. The online study consisted of an informed consent, image rating trials, and debriefing.

Results

To examine whether reverse correlation could be used to create apologetic faces from both a female and male base face, we compared ratings of apologetic and base images for female and male faces separately, collapsing across the gender of generators and participants. Consistent with our hypothesis and findings from previous studies, the apologetic faces of men (M = 67.29, SD = 20.78) and women (M = 63.41, SD = 18.53) were significantly more apologetic than the base faces of men (M = 28.20, SD = 23.27) and women (M = 28.66, SD = 22.56), t(197) = 17.62, p < .001, d = 1.25 and t(197) = 17.33, p < .001, d = 1.23, respectively. The male apologetic face was significantly more apologetic compared to the female apologetic face, t(197) = −2.97, p = .013, d = 0.21, while the male and female base faces did not significantly differ, t(197) = 0.26, p = .999, d = 0.02.

To test whether ratings on other apology-related characteristics replicated using the newly generated apologetic faces, we compared the female and male CIs to their respective base faces separately. Multiple paired-sample t tests confirmed that apologetic faces were again perceived as significantly more sad, regretful, remorseful, and submissive, as well as less dominant, than the base face from which they were generated, regardless of the target face gender (all ps < .001, see Table 2 for details). Unlike Study 1, ratings on the characteristic “trustworthy” significantly differed, with the female apologetic face (M = 40.44, SD = 20.59) rated less trustworthy than the base face (M = 50.13, SD = 22.72), t(198) = −5.14, p < .001, d = 0.36. Consistent with the previous studies, there was no difference in ratings on the characteristic “trustworthy” for the male faces. Once again, the apologetic CIs were rated as more sad than apologetic for the male and female faces, t(197) = −9.64, p < .001, d = 0.68 and t(197) = −9.98, p < .001, d = 0.71, respectively.

Table 2.

Mean characteristic ratings and comparisons across faces: Study 2.

	Apologetic male face	Male base face
Characteristic	M (SD)	M (SD)	t	d
Apologetic	67.29 (20.78)	28.20 (23.27)	17.62***	1.25
Sad	78.63 (17.47)	31.42 (24.01)	23.94***	1.70
Regretful	68.15 (18.78)	28.24 (24.79)	19.46***	1.38
Remorseful	68.31 (18.17)	30.06 (23.54)	18.67***	1.33
Submissive	56.80 (21.59)	36.40 (24.19)	9.09***	0.65
Trustworthy	39.62 (20.10)	38.75 (22.62)	0.45	0.03
Dominant	31.14 (17.55)	58.04 (24.36)	−13.31***	0.65
	Apologetic female face	Female base face
Characteristic	M (SD)	M (SD)	t	d
Apologetic	63.41 (18.53)	28.66 (22.56)	17.33***	1.23
Sad	75.59 (18.76)	30.24 (24.14)	21.85***	1.55
Regretful	65.00 (19.48)	26.14 (21.70)	19.46***	1.38
Remorseful	64.25 (19.26)	28.35 (22.71)	17.74***	1.26
Submissive	58.01 (21.24)	39.55 (22.48)	8.63***	0.61
Trustworthy	40.19 (20.40)	50.07 (22.64)	−5.32***	0.38
Dominant	27.79 (17.93)	55.58 (23.53)	−13.97***	0.99

Note. All p values within each gender category were Bonferroni-adjusted for seven comparisons (i.e., comparisons of the apologetic and base faces for seven characteristics per image gender).

^***p < .001.

Moderation of apology ratings

Finally, we examined whether and how gender affected apology ratings. A 2 (rater gender: man, woman) × 2 (generator gender: man, woman) × 2 (target gender: man, woman) mixed design analysis of variance (ANOVA), with rater gender representing a between-group factor, and generator and target genders representing within-group factors, revealed a significant main effect of generator gender, F(1, 196) = 76.21, p < .001, and target gender, F(1, 196) = 8.80, p = .021. These effects were qualified by a significant Generator Gender x Target Gender interaction, F(1, 196) = 82.55, p < .001. Follow-up analyses revealed that men generated an apologetic female face (M = 51.54, SD = 26.94) that was subsequently rated as less apologetic than any of the other apologetic faces: male-generated male face: M = 67.41, SD = 23.44, t(197) = −7.88, p < .001, d = 0.56; female-generated male face: M = 67.16, SD = 23.48, t(197) = −7.96, p < .001, d = 0.57; female-generated female face: M = 75.28, SD = 20.80, t(197) = −10.88, p < .001, d = 0.77. By contrast, women generated an apologetic female face that was subsequently rated as more apologetic than the other apologetic faces: male-generated male face: t(197) = 4.41, p < .001, d = 0.31; female-generated male face: t(197) = 4.80, p < .001, d = 0.34. All p values were Bonferroni adjusted. No other effects were significant, ps > .212. ⁵

Discussion

The apologetic faces generated in Study 2 were rated as significantly more apologetic than the base faces from which they were created. This is consistent with previous findings and with the possibility that people have a mental representation of an apologetic face which can be estimated using reverse correlation. Both the male and female apologetic faces showed comparable patterns in the ranking of the apology-related characteristics that were perceived. Participants rated both apologetic faces as highly sad, regretful, and remorseful. For the first time, ratings for trustworthiness differed across the base and apologetic faces, however, this was only true for female faces. The female apologetic face was rated as less trustworthy than the base face, whereas this difference was not significant for male faces.

Although gender moderated apology ratings, the effect was not consistent with our predictions. Generator and target gender were the only significant main effects to emerge, and these were qualified by an interaction. Having the same gender as the target did not influence apology ratings; men and women identified similar levels of apology for each face, suggesting perhaps women are not better at differentiating subtle or complex emotions. Despite similarities in how men and women perceive apologetic faces, the findings do seem to support the theory that women are particularly expressive (Hall & Matsumoto, 2004). This is seen in the interaction between generator and target gender, with women generating female faces that were later seen as more apologetic than all other faces. However, this expressiveness may be limited to interactions with ingroup members. Men generated female faces that were rated significantly less apologetic than all other faces, suggesting perhaps women express apologies differently when they are directed at men relative to women. Alternatively, this interaction could be driven by an ingroup bias such that generators created apologetic faces of ingroup members that were later rated as more apologetic than those created of outgroup members.

General Discussion

We set out to replicate and extend previous findings demonstrating that nonverbal apologies can be estimated using reverse correlation. We wanted to see if participants would generate visual templates of female faces that would be judged as sharing characteristics with those of male faces. Apologetic faces were created from female (Studies 1 and 2) and male (Study 2) base faces and, as predicted, the apologetic faces in each study were rated by an independent group of participants to be significantly more apologetic than their respective base faces. This provides evidence that using reverse correlation successfully led to the creation of visual templates from male and female base faces that were subsequently perceived as apologetic.

A second goal of this research was to determine whether these visual templates would receive similarly high ratings for apology-related characteristics. Across both studies, and in line with previous work (George et al., 2023), the characteristic “sad” received the highest ratings for each of the apologetic faces, providing further evidence that displaying sadness is important when apologizing. Regret and remorse—constructs that have received attention in the verbal apology literature—were also rated highly by participants judging male and female apologetic faces. Thus, the characteristics “regretful” and “remorseful” were perceived to a similar degree as “apologetic.” It is possible that this is because regret, remorse, and apology are equally important characteristics to display while apologizing. Alternatively, these three constructs may be less discernable or more difficult to identify because, unlike sadness, they represent more abstract concepts.

Trustworthiness was also examined in the present research. Promises not to reoffend are important aspects of successful verbal apologies (Kirchhoff et al., 2012; Lewicki et al., 2016; Schlenker & Darby, 1981; Struthers et al., 2008), and this might be conveyed nonverbally by appearing trustworthy. However, consistent with previous findings (George et al., 2023), ratings on trustworthiness did not vary consistently across apologetic and base faces. The female face in Study 2 was the only instance in which ratings on trustworthiness significantly differed from the base face, and surprisingly, the female apologetic face was rated as less trustworthy than the base face. Mean ratings suggest that this was because the female base face was perceived as more trustworthy than other faces, and not because the female apologetic face (whose mean rating was in line with ratings on the characteristic “trustworthy” for male faces) was perceived to be less trustworthy than other faces. Although speculative, one possible reason for this is that women are generally perceived as more trustworthy than men, however, seeing an apologetic female face may bring to mind the frequency with which women apologize (Schumann & Ross, 2010). This, in turn, may lead to beliefs that women transgress (i.e., betray trust) more than men. If this is the case, being sensitive to situations that may harm a relationship and offering an apology may ironically be a detriment to building and maintaining trust. Discrepancies in trustworthiness ratings may also result from how participants were thinking about the act of apologizing. An individual thinking about how they would signal trustworthiness following a transgression may perceive an apologetic face as more trustworthy than someone who is imagining they are being apologized to by someone who has recently betrayed trust. Investigating other characteristics known to vary across genders (e.g., emotions such as sadness) is another avenue for future research, as this could provide further insight into the relationship between trustworthiness, gender, and apologies.

Submissiveness and lack of dominance were also examined as potential characteristics of apologetic faces. The characteristic “submissive” was not rated particularly high across each of the studies, with mean ratings ranging from 54 to 60 on a scale from 0 (not at all) to 100 (very submissive). In Study 2, apologetic male faces were rated as more submissive than their base faces, supporting the predicted association between constructs and suggesting that apologies may be thought of as submissive behaviors. Alternatively, the association between gender and stereotypically gendered apology-related characteristics may be weaker than originally expected. Future research aimed at parsing the relationship between constructs in an apology context as well as in other domains would prove beneficial. Measuring the degree to which participants rate themselves on gender-related characteristics and submissiveness may provide additional data that could be used to scale ratings of target faces to better inform this relationship (Wood & Eagly, 2015). Such self-report ratings could be included in mediation and moderation models to examine how one’s self-perception might shape how they perceive others.

Finally, the gender moderation analyses in Study 2 only partially supported our hypotheses. Mental representations of men and women differed for female faces, such that women generated a visual template of a female apologetic face that was rated as more apologetic than that of men. This is consistent with research showing that women are more expressive than men (Adams et al., 2015; Fischer & LaFrance, 2015), however, it is possible that expressiveness is specific to interactions with other women, and that the interaction between generator gender and target gender represents an ingroup bias specific to women. Female faces generated by men were rated as significantly less apologetic than those generated by women. Given the current findings that men and women perceive faces similarly, it seems possible that women express differing levels of nonverbal apologies when addressing men versus women.

Yet another possible explanation for gender differences in mental representations is that men more readily identify an apologetic expression on female faces. That is, men may require less signal (e.g., fewer visual cues) to identify apology on a female face, whereas women may require more. Such an explanation aligns with findings that women apologize more frequently than men due to a lower threshold for identifying transgressions (Schumann & Ross, 2010). In other words, if women apologize more for low-severity transgressions, the facial expressions that accompany these apologies could also be perceived as less severe. Men would subsequently experience apologies from women as less expressive on average, whereas women may incorporate their personal experiences of apologizing and hold mental representations of an apologetic female face based only on transgressions that meet a specific severity threshold. It is also possible that men simply make more varied selections during the image generation phase because they lack the ability to detect more subtle or nuanced emotional signals (Hall & Matsumoto, 2004). When responses are averaged to generate a group-level CI, such variability could lead to a face with less refined signals of apology, and this may be particularly true for outgroup (i.e., female) faces.

Despite women having a mental representation of an apologetic female face that was more expressive than that of men, men and women judged apologetic faces similarly, regardless of the face’s gender. These findings have potential implications for interpersonal relationships. For example, romantic partners are likely to encounter instances that necessitate an apology. Le et al. (2020) found that, even outside of an apology context, accurate perception of a partner’s appeasement expressions (e.g., embarrassment, shame, submissiveness, guilt) was linked to greater relationship quality. Appeasement expressions map onto our understanding of verbal apologies, and our findings complement this by demonstrating that perceivers can identify such characteristics on novel faces. Within apologies, it seems possible that one’s ability to identify apology-related characteristics can lead to higher quality romantic relationships.

Limitations and Future Directions

One main limitation of the current research is that participants were constrained by the characteristics they were asked to rate. Although the characteristics in the present research were derived from a larger list used in previous work (George et al., 2023), it is interesting to consider whether similar results would emerge if a larger and/or different pool of characteristics were used. Qualitative research in which participants describe or list characteristics they would anticipate in apologetic faces of men versus women may be one way to address this limitation.

Given the number of ratings that participants were asked to make, combined with our research goals, we decided a priori to have participants assess group-level CIs (averaged across all generators’ individual-level CIs), as opposed to the individual-level CIs generated by each participant in the first phase. Although we felt the benefits of taking this approach outweighed the disadvantages, it is important to note that a main limitation of using only the group-level images is that this can increase the risk of Type I error (Cone et al., 2021). It would be useful to use multiple approaches to assess the robustness of our findings in future research.

Our findings suggest that gender may play a role in how one thinks about and perceives apologetic faces. This was particularly evident in apology ratings of female faces, which differed based on generator gender. Previous studies demonstrated that women identify transgressions at a lower threshold than men, leading to more frequent apologies for less severe offences (Schumann & Ross, 2010). In our studies, participants were asked to identify and judge apologetic facial expressions without information about the type or severity of an antecedent transgression. We can therefore only speculate as to whether our findings indicate gender differences in how men and women think about (women’s) apologetic faces, or gender differences in the severity of transgressions we attribute to apologetic faces. Similarly, it is unclear whether participants thought of specific transgressions during either the generation or rating phase. As a next step, it would be interesting to have participants describe a transgression for which the people in the images might be apologizing. Responses coded for severity and comparisons across genders would provide insight into how we think about apologies. Alternatively, one could look at gender differences in generated apologetic faces when the transgression is manipulated.

Participants in the current research were limited to judging faces that varied systematically across genders, however, other social categories experience stereotypes that may also interfere with apologies. For example, when one researcher team asked participants to generate a visual template of a welfare recipient, the resulting face was rated as Black, female, unlikable, incompetent, and lazy relative to the generated image of nonwelfare recipients (Brown-Iannuzzi et al., 2017). If Black women are associated with this idea of a “welfare queen,” it seems reasonable to predict that their apologies may be perceived as insincere and less apologetic than those of others. In addition, each of the base faces in the current study were created using images of White people. Examining groups that face different stereotypes will provide a greater understanding of the prejudices that may bias how we repair and maintain relationships. Deeper understanding may also be achieved by using base faces derived from unaltered photographs rather than digitally morphed faces. Although such an approach would require a large number of replications using multiple different individuals as stimuli, naturalistic images would allow greater generalizability. Similarly, extending this research to include in-person apologies and evaluations of apology sincerity will help us understand if and how mental representations affect apology behavior and judgment. For example, are mental representations used as points of comparison when evaluating aspects of apologies such as sincerity? If so, could the degree of punishment or one’s willingness to forgive be related to the degree of overlap between the two? The current reverse correlation findings create a foundation on which to explore such possibilities.

Conclusion

Apologies help us maintain relationships and fulfill the need to belong following transgressions. Researchers have identified a number of verbal apology components that most effectively repair relationships. In the present research, we set out to determine if apologies could be expressed through facial expressions, and what that would look like. Findings suggest that people have a mental representation of an apologetic face for both women and men. Apologetic visual templates were consistently rated as apologetic, sad, remorseful, and regretful, providing further evidence that these are important features of nonverbal expressions of apology, regardless of apologizer gender. Although gender differences emerged in the image generation phase, it is important to note that women and men did not differ in their ratings of apologetic faces. This finding is promising, as it suggests that men and women might be on the same page when it comes to perceiving apology in someone attempting to repair a relationship. Overall, this work represents a first step toward better understanding nonverbal apologies by incorporating social information such as gender and emotion into face perception.