“The Authors” Make Me Think Equally of Women and Men: Exploring Mixed-Gender Representations in a Visual Categorisation Task

1 Conceptual Background

Text comprehension is generally understood as a process of constructing a coherent mental representation of the text that goes beyond its mere linguistic representation. In this process, the surface meaning of words is merged with prior knowledge from long-term memory to form situation models, as described for example, in Kintsch’s Construction-Integration Model (Kintsch, 1988; Kintsch & Van Dijk, 1978). Concerning the surface meaning of words, person nouns can linguistically refer to the gender of individuals in different ways (Motschenbacher, 2010; Stahlberg et al., 2007): lexically, i.e. in the semantics of nouns such as mother, father, policewoman or policeman, morphologically such as in waitress (female waiter) and widower (male widow) and grammatically, such as in French la _FEM violiniste autrichienne _FEM and le _MASC violiniste autrichien _MASC (the Austrian [female] violinist and the Austrian [male] violinist). The extent to which human referent gender is linguistically encoded varies greatly between languages (Dryer & Haspelmath, 2022). Further, role nouns are often associated with a gender category as part of speakers’ extra-linguistic knowledge, thus biassing word comprehension. This gender bias is especially true for many job titles and has been theorised by researchers to indicate conceptual (Irmen, 2007; Sera et al., 1994), social (Motschenbacher, 2010), typical (Egan & Perry, 2001), or stereotypical gender (Carreiras et al., 1996; Gabriel et al., 2008).

Much of the previous research has examined gender representations of person nouns in the context of discourse processing, where person nouns are inserted into short text passages (recent examples are Körner et al., 2022; Tibblin et al., 2023; Xu & Sturt, 2023). Studies conducted at the word level can be found mainly in the context of the discussion on the “automaticity” of the activation of stereotypical gender information (e.g. Banaji & Hardin, 1996; Cacciari & Padovani, 2007; Fatfouta & Sczesny, 2023; Oakhill et al., 2005) and to investigate the interplay between linguistic (grammatical, morphological) and extralinguistic (stereotypical) information (e.g. Gygax et al., 2012; Gygax & Gabriel, 2008).

In the experimental tasks typically used in this research, participants work through a sequence of trials in which two stimuli are presented either simultaneously or sequentially and to which participants make a dichotomous decision (e.g. yes vs no, same vs different). One type of task requires participants to judge whether or not the two stimuli are associated/linked (explicit association, verification, or matching tasks, e.g. Oakhill et al., 2005; Sato et al., 2016). Participants explicitly compare the stimuli and are thus encouraged to invoke (among others) gender-specific information when evaluating semantic associations.

In another type of task, only one of these stimuli – the target stimulus – must be processed in terms of a basic feature, while the other – the prime stimulus – is task-irrelevant (priming tasks, e.g. Banaji & Hardin, 1996). Such priming tasks can be further differentiated according to whether the target processing task does or does not reflect the assumed relationship between prime and target. Wentura and Degner (2010) accordingly distinguish response priming and semantic priming. Applied to the activation of gender categories by role nouns, the difference lies in whether the (gender) category that is (presumably) activated by the primary stimulus is a relevant feature for the task that is performed on the target stimulus (White et al., 2018). For example, if the primes are lexically feminine or masculine nouns, and the task is to decide whether a target pronoun is feminine or masculine (Banaji & Hardin, 1996, Exp. 1), gender category activation is immediately relevant for the task (response priming). However, if the task is to decide whether a target word is a pronoun or not (Banaji & Hardin, 1996, Exp. 2), gender category activation is not immediately relevant for the task (semantic priming). Response priming effects tend to be more robust than semantic priming effects (e.g. De Houwer, 2003; Wentura & Degner, 2010; for gender categories e.g., Kidder et al., 2018; Müller & Rothermund, 2014; White et al., 2018), which is explained by the priming effects in response priming reflecting the combined influence of memory and response activation processes (Wentura & Rothermund, 2014). Furthermore, while most of the previous research on the gender representations of role nouns exclusively relied on lexical stimuli, some studies have used a Linguistic Visual paradigm (e.g. Irmen & Köhncke, 1996; Lemm et al., 2005; Sato et al., 2016; Zacharski & Ferstl, 2023), in which participants are typically shown written text (lexical stimulus) prior to observing a visual scene (pictorial stimulus).

As we were interested in whether linguistic stimuli spontaneously activate gender categories, we chose a priming design in which participants were not explicitly asked to compare stimuli to complete the task (as in explicit association, verification, or matching tasks). More specifically, we chose a response priming design, in which response competition is expected to produce faster and more accurate responses when the target stimulus and prime activate the same response (response facilitation) and that produces slower and less accurate responses when the target stimulus and prime activate different responses (response interference; Wentura & Rothermund, 2014, p. 53).

Even though cross-modal priming effects are weaker than effects observed with same-modality pairs (e.g. Carr et al., 1982), we decided to use visual target stimuli, namely gendered human faces, to explore whether nouns can simultaneously activate more than one gender category. Using a visual categorisation task allows us to investigate whether gender categories are being activated even when they neither serve to establish textual coherence nor generally facilitate linguistic processing hence informing us about the impact that linguistic choices, such as GFL, have beyond the purely linguistic realm.

We are aware of only two studies that investigated the activation of gender information in a cross-modal response priming design. Lemm et al. (2005, Exp. 1) used gender stereotypical role nouns (e.g. “ballerina’,” “banker,” or “author”) and lexically gendered words (e.g. “chairwoman,” “chairman,” or “chairperson”) to prime female, male, and neutral representations. These primes were followed by target drawings representing either a female or a male person that had to be categorised according to gender. Results showed that participants were quicker to classify targets when their gender was consistent with the preceding prime, though the contrast reached significance for female targets only. Generally, the priming paradigm employed by Lemm et al. (2005) thus demonstrates the adequacy of text-to-image priming to assess the activation of gender representations across modalities.

Discussing the relationship between prime and target stimuli, Kawakami and Dovidio (2001) argue that photographs may provide more ecologically valid stimuli than more abstract category labels, as the photographs represent actual members of the primed category. In their study (Exp. 1), positive and negative gender stereotypical trait words were used as primes (e.g. “fashionable” and “gossipy” for female traits and “brave” and “dominant” for male traits) to a categorisation task in which participants had to decide whether a yearbook photo of a college student represented a woman or a man. Results showed that participants were both quicker and better at categorising the pictures when the gender of the individual in the picture was congruent with the preceding prime, relative to when it was inconsistent.

Building off this previous research, we have designed an experimental set-up using person nouns as the prime stimuli and photographs of faces as the target stimuli. However, to explore the possibility of simultaneous activation of more than one gender category, we use pairs of faces as target stimuli. To match targets and primes in number, the primes also describe people in the plural.

The main prime type we use follows the perspective of GFL as equally activating feminine and masculine beliefs. Specifically, we examine stereotypically gendered pluralised role nouns as a microcosm of this conceptualisation, on the basis that normed non-gender stereotyped role nouns should equally activate both feminine and masculine gender information.

Previous research (Banaji & Hardin, 1996; Müller & Rothermund, 2014) suggests that stereotypically gendered person nouns produce smaller priming effects than linguistically (i.e. lexically, morphologically, or grammatically) gendered person nouns, and although the relevant interaction was not significant in Lemm et al. (2005, Exp. 1), there was evidence that the strength of the priming effects was related to how strongly the primes were perceived to be associated with gender. This is consistent with the notions that linguistically gendered nouns provide categorical gender information, while stereotypically gendered nouns provide probabilistic gender information (Kreiner et al., 2008, for a discussion on differences in gender representations), and that it is easier to overcome cognitive discrepancies/mismatches in gender information between probabilistic primes and categorical targets than between the categorical primes and categorical targets.

Since the sensitivity of our experimental design to detect priming effects has not yet been tested, we decided to also include a word type, which is expected to have a strong gender priming effect, namely gender-typical names. First names have traditionally been selected to indicate a person’s gender. Their use allows us – by combining them into pairs – to maximise the match/mismatch of gender information between prime (name pairs) and target (face pairs), hence serving as a benchmark for the functioning of the design.

An important consideration for priming studies is the stimulus onset asynchrony (SOA; i.e. the time interval between prime onset and target onset). In a recent meta-analysis, Brysbaert (2019) found that across studies conducted in English (N = 27), the average reading speed for non-fiction was 238 words per minute, or 3.9 words per second, and that across studies conducted in Swedish (N = 5), the average reading speed (combined fiction and non-fiction) was 218 words per minute, or 3.6 words per second. As the primes used in this study were a mix of one- and three-word items, this suggests that participants require a longer SOA. Hutchison et al. (2008) conducted a study on associative strength which refers to the level of association between a specific prime and target in both directions: forward strength and backward strength. They demonstrated that forward strength was present in both short SOAs (250 ms; 200 ms prime presentation, 50 ms inter-stimulus period) and long SOAs (1,250 ms; 1,000 ms prime presentation, 250 ms inter-stimulus period); however, backward strength was only present in the long SOA. As this study involves visual targets of linguistic primes, which complicates the estimation of their associative strength, a prolonged SOA that allows for both forward and backward associative strength to inform responses is beneficial. Nevertheless, research has indicated that 2,000 ms SOAs are excessively long, leading to conscious responses that might possibly reverse stereotype effects (e.g. Blair & Banaji, 1996). Following Hutchison et al. (2008), we implemented an SOA of 1,250 ms in our study which encompasses a 1,000 ms prime presentation and a 250 ms inter-stimulus period. This temporal arrangement met all our requirements by providing enough time for participants to read the primes, and for the backwards associative strength to facilitate responses, without inducing conscious responses.

The present study was conducted in Norwegian (Bokmål). Even though Norwegian possesses a grammatical three-gender system (masculine, feminine, and neuter), the language is experiencing the erosion of the correspondence between grammatical and referential gender for personal nouns (Bull & Swan, 2002) where lexically masculine and feminine nouns merge into a common gender (the former masculine). To illustrate, kvinne, the Norwegian word for woman, can be used with both the grammatically feminine determiner (ei _FEM kvinne) and grammatically masculine determiner (en _MASC kvinne). It is possible to morphologically mark femininity via suffixes (e.g. skuespillerinne [actress] instead of skuespiller [actor] to indicate that the actor is female) but these suffixes are no longer productive although a few remnants can still be found (e.g. en venn – en/ei venninne [a friend], en elsker- en/ei elskerinne [a lover], en svoger – en/ei svigerinne [a sibling-in-law]). Personal nouns assigned to the grammatical neuter class such as barn (child) or menneske (human being) remain unchanged, and there is no grammatical gender to designate people who assign themselves to neither the female nor the male gender. As such Norwegian (bokmål) is similar to English in terms of referential gender and is therefore a good choice to overcome the heavy reliance on English in experimental research (e.g. Levisen, 2019) without having to introduce grammatical aspects regarding referential gender.

To summarise, in this study, we utilise a face pair categorisation task (Face Pair Task) in a response priming setup to examine text-to-image gender priming in Norwegian. Specifically, we aim to (a) investigate the existence and reliability of gender priming effects on non-linguistic visual gendered targets and (b) explore whether mixed gender information can both prime and be primed.

2 Overview and Hypotheses

This study comprises two experiments. We first wanted to see if we could replicate previous findings with our triple-choice categorisation task (Face Pair Task). For this purpose, only feminine and masculine primes were used in Experiment 1. Two word-types, role nouns (e.g. “the nurses” [feminine], “the programmers” [masculine]), and name pairs (three words: e.g. “Ann and Catherine” [feminine], “Håvard and Hans” [masculine]) were used as primes. Since the form of these word types differs greatly, they were presented by block. Due to its exploratory nature and in line with Lemm et al. (2005), in Experiment 1, the block order (role nouns before name pairs) was fixed.

Next, we wanted to explore whether mixed-gender targets (i.e. mixed-gender faces) can be successfully primed. To this end, in Experiment 2, gender-balanced primes were included for both role noun primes (role nouns considered suitable roles for both men and women to hold, e.g. “the oceanographers” [non-stereotyped]) and name pairs (pairs of names composed of one female and one male name, e.g. “Ann and Hans” [gender-mixed]). To provide a within-block baseline, control primes (nonsense strings designed to mimic the role noun and name pair primes) were also included. Further, to allow for a direct comparison between role noun and name pair primes, in Experiment 2, the block order was counterbalanced between participants.

Both experiments were composed of five experimental blocks: (1) a button task, (2) the Face Pair Task without primes, (3) and (4) the critical Face Pair Task with primes, and (5) the button task again. Within each block, the trial order was randomised by participants. Responses and response times were collected. The button task was a simple task that allowed for familiarisation with the inputs needed to give responses, while the Face Pair Task without primes allowed for familiarisation with the face pair task specifically. As such, the button task can be seen to provide a measure of baseline motoric costs, while the Face Pair Task without primes provides a measure of both motoric costs and cognitive costs associated with facial perception. Data for both experiments, and code used to explore the data, are available at [https://dataverse.no/privateurl.xhtml?token=1780b24d-1bca-45de-a699-b8c4a2302d05].

In line with previous findings (e.g. Lemm et al., 2005, Exp. 1), in this study, we hypothesise (H1) that a response facilitation effect will be found for the gender-specific face pairs following gender congruent compared to incongruent primes and (H2) that this effect will be stronger for the primes that provide categorical gender information (i.e. gender-matched name pairs) than those that provide probabilistic gender information (i.e. gendered role nouns). Regarding our research question of whether gender-mixed representations can be activated, our assumptions differ for word types. For the role noun primes, the non-stereotyped roles should broadly activate gender such that feminine and masculine information is equally activated, leading to a processing advantage compared to the feminine stereotyped and masculine stereotyped roles when responding to the mixed gender faces. We therefore hypothesise (H3) that, for the role nouns, a response facilitation effect will be found for the mixed-gender faces following gender-balanced compared to gender-specific primes. This is not necessarily the case for the name pair primes, due to match-mismatch bias (e.g. Proctor, 1981), which holds that the decision that two stimuli are similar is, in general, made more quickly and accurately than the decision that they are different. Related to our task, for the gender-specific name pairs, the gendered attributes activated by the first name are reinforced by the second name, while for the mixed-gender name pairs, these attributes are contradicted by the second name. As such, responses to the mixed gender name pairs should have higher error rates and slower response times to an as-yet-unknown degree; this is why we neither theorise for nor against a processing advantage for mixed gender compared to gender-specific face pairs.

3 Experiment 1

The data presented in Experiment 1 were recorded as part of a study exploring the adequacy of different stimulus-response setups (Öttl et al., 2023). For the present study, data from the bimanual set-up from the above study were used. The results of the experimental blocks without priming (i.e. Block 1, Block 2, and Block 5) are presented in Öttl et al. (2023, Bimanual Setup). In contrast, we present the data from the Face Pair Task with priming (Blocks 3 and 4), with the results of the Face Pair Task without primes (Block 2) used as a potential control variable in the data analyses,

3.1 Methods

3.1.1 Participants

Thirty-two native Norwegian-speaking participants (15 male) with a mean age of 23.8 (SD = 3.5) were recruited at the Norwegian University of Science and Technology (NTNU). Since both hands were used to register a response, both left- and right-handed participants were included (5 and 27 participants, respectively).

3.1.2 Apparatus

This experiment was run using a laptop computer in a laboratory setting. E-Prime 2.0.10.356, in association with an eye-tracking system, was used for stimulus presentation and logging responses. Responses were collected using a Chronos response box. For further details see Öttl et al. (2023).

The experimenter was seated facing the participant, with eye contact obstructed by the computer screens placed between them. This arrangement was used to offer the participant as much privacy as possible, with the experimenter still being able to indirectly monitor participants’ progress through access to both their gaze coordinates and a duplication of the stimulus screen. Participants were instructed to rest their fingers (middle and index fingers of both hands) on the response box throughout the experiment’s duration.

3.1.3 Materials

For both the button task and face pair task, responses were given through four buttons, with participants pressing two designated buttons to respond to each response condition (two with either the left or right hand for the matched-response categories or one with each hand for the mixed-response condition, Figure 1). For the button task, stimuli were basic representations of the response options, with the relevant buttons to press for each trial presented on the screen. For a discussion of the adequacy of the response set up, see Öttl et al. (2023).

Figure 1

Bimanual response setup for two spatial arrangements: (i) response set-up 1 (FF left); (ii) response set-up 2 (MM left).

To control for any order effect of stimulus position, the spatial arrangement (left–right) of the stimuli and responses was counterbalanced between participants but kept constant across tasks. Therefore, stimuli were created in two sets, as described below. Participants were pseudo-randomly assigned to the spatial arrangement version to ensure roughly equal numbers of women and men in each version.

3.1.3.1 Face Pair Task

Stimuli were created from pictures of 45 female and 45 male faces taken from the Chicago Face Database (Ma et al., 2015). These were combined into three sets of face pairs: a) 15 image pairs consisting of two female faces, b) 15 pairs consisting of two male faces, and c) 15 pairs consisting of one female and one male face. The faces were pretested to ensure that female and male faces were categorised according to their respective gender category equally rapidly and that comparable pictures were assigned to the matched and mixed gender categories. Further information on this process (image selection-from-database criteria, image post-processing, and image pairing procedure) is described in the project protocol [https://dataverse.no/dataset.xhtml?persistentId=doi:10.18710/WWPPCA#]. The dimensions for each face pair stimulus were 300 × 508 pixels (8.5 cm × 14.3 cm). Examples of the faces used to produce the stimuli are presented in Figure 2. Two sets of face pairs were created to match the response setups: spatial arrangement 1 (female left, male right) and spatial arrangement 2 (male left, female right).

Figure 2

Examples of the Visual Stimuli Used in the Face Pair Task Note: (i) Examples from spatial arrangement 1 (female left); (ii) Examples from spatial arrangement 2 (male left). Faces obtained from the Chicago Face Database (Ma et al., 2015).

3.1.3.2 Primes

Primes (Appendix A, Table A1) were two sets of Norwegian person nouns: definite pluralised role nouns that were stereotypically gendered feminine (e.g. ballettdanserne [the ballet dancers]) or masculine (e.g. gruvearbeiderne [the miners]) and pairs of names that were typically female (e.g. Anna og [and] Marianne) or typically male (e.g. Tommy og [and] Martin).

Stereotypicality of the role nouns was based on explicit ratings from Misersky et al. (2014), with role nouns selected to be equally stereotypical across genders (estimated proportion of female members for feminine role nouns were M = 0.77, SD = 0.05 and for masculine role nouns M = 0.21, SD = 0.04). The definite form was chosen over the indefinite form as it seems more consistent with the specificity of the name pairs. Nevertheless, since definiteness is marked by suffixation in Norwegian, this feature may be less salient with words presented in isolation than would be the case in, e.g. English (e.g. ballettdansere [ballet dancers] vs ballettdanserne [the ballet dancers]).

Names were selected from the list of most common names of persons born in Norway between 1976 and 1996 (Navnestatistikken, 2022). This range of birth years was chosen to exclude names that, although very common, are predominantly associated with very young or very old people, which could lead to a perceived mismatch between face and name. Only names that are used exclusively as unisex names in the cultural context of our study were selected.

3.1.4 Procedure and Research Design

This experiment has a 3 (Face Pair Gender: female vs mixed vs male) by 2 (Prime Gender: feminine vs masculine) by 2 (Prime Type: role nouns vs name pairs) by 2 (Spatial Arrangement: female left vs male left) mixed design. Face Pair Gender, Prime Gender, and Prime Type are all within-participant factors, while Spatial Arrangement is a between-participant factor.

Participants were informed that the purpose of the experiment was to investigate attention switching between different visual tasks, especially between reading and face recognition, and that they would be undertaking five visual categorisation tasks, followed by a memory task. Participants gave written informed consent prior to participation. At experimental onset, they were presented with an information screen that restated that the purpose of the experiment was to learn more about what happens when attention shifts between different tasks, and informed them that, to this end, they would be undertaking two tasks: one consisting of categorising a picture of two faces according to the gender of the faces, and one consisting of reading words and remembering them.

Prior to the first Face Pair Task block with primes (Block 3), participants were told that they would be given the reading task in addition to the categorisation task with which they were already familiar. Specifically, they were told that they would see a word in the centre of the screen before each face pair and that they would have to remember these words as part of a memory task conducted at the end of the experiment, which assessed how many of the words the participant remembered. Prior to the second block with primes (Block 4), participants were told that they would now perform the same tasks, but with pairs of names rather than individual words for the reading task and were informed that the memory task would include both the words and the names that the participant had seen.

Following a 45-trial button block and a 45-trial face pair block, participants completed a 45-trial critical face pair block with role noun primes and a 45-trial critical face pair block with name primes, followed by a 30-trial button block (repetition). All face pair blocks were preceded by six practice trials.

Each trial in the critical blocks began with a gaze-contingent fixation cross being displayed in the centre of the screen. When the participant had fixated on the cross for 500 ms, a prime word appeared for 1,000 ms centrally on the screen, followed by a 250 ms blank screen after which a face pair was presented until the participant gave a response (with no upper time limit). Stimuli remained on the screen for 500 ms after the participant registered their response. Response and response time were collected for each trial.

Finally, participants were asked to complete a post-test during which they were presented with 40 words (20 pluralised role nouns [10 stereotypically feminine, 10 stereotypically masculine], 20 gendered names [10 typically female, 10 typically male]) in randomised order, half of which (five from each group) they had encountered as primes in the experiment. For each item, participants were asked to indicate whether they thought they had read the word during the experiment. This was included to assess whether participants were paying attention during the experiment, as, if they were, they should respond correctly above chance. After completing the post-test, participants were informed of their performance. Individual difference measures, not included in this article, were also collected. The entire experimental session lasted approximately 25–30 min. After the experiment, participants were debriefed and received a cafeteria voucher as compensation for their participation.

3.1.5 Data Preparation

Prior to analyses, both by-participant and item-by-participant data screening were used. The item-by-participant exclusion was based on both response (trials were excluded when more than two button presses were registered) and response time (trials were deselected if they were shorter than 200 ms, Baayen, 2008, or more than three standard deviations above a participant’s mean in each block) and led to the removal of 3.21% of the data. By-participant exclusion was based on individual error (participants who had more than 10% of their responses to the face pair blocks removed during item-by-participant screening were deselected) and on prime recognition (deselection criteria were set at 50% or less of the items in the prime recognition task responded to correctly). No participants were deselected due to these measures.

Error rate (percentage of experimental items responded to incorrectly), and response time (speed of responses to experimental items to which participants responded correctly), were statistically analysed using generalised linear mixed effects modelling (GLM) in R (version 4.0.3; R Core Team, 2018). Data analysis was conducted through the glmer function of Version 1.1-23 of the lme4 package (Bates et al., 2015). To prevent misattribution of order effects as differences between prime types, error rate, and response time analyses were conducted separately for role noun and name pair primes. For all analyses, initial models were designed that captured the maximal random effects structure as justified by the design. These were composed of the experimental factors Prime Gender (feminine vs masculine) and Face Pair Gender (female vs mixed vs male), and their interaction, as well as fixed factor covariates of Participant Gender, Handedness, Age, Trial Number, and Spatial Arrangement, random intercepts of Participant, Image Shown, and Prime Shown, and random slopes of each experimental fixed factor by each random factor.

Further, to account for effects related to the motoric and cognitive costs associated with the different responses in the Face Pair Task (Öttl et al., 2023), responses to the Face Pair Task without primes (Block 2) were also included in the analyses. For each participant, the mean error rates, and response times on the Face Pair Task without primes (Block 2) were calculated per Face Pair Gender and then included in the maximal models and during model fitting as a potential random by-participant slope. To aid with comprehension, the mean value for the Face Pair Task without primes block (Block 2; overall, and per face pair condition) is reported in Table 1. Model refinement to find the model of best fit followed the approach outlined in the project protocol (https://dataverse.no/dataset.xhtml?persistentId=doi:10.18710/WWPPCA#). Comparisons between the maximal models and models of best fit for each analysis can be found in Appendix B, Tables B1 (error rate), and B2 (response time).

Table 1

Mean Error Rates & Response Times per Face Pair Gender for Block 2: Face Pair Task without primes (Experiment 1)

Face Pair Gender	Mean Error Rate (%)	Mean Response Time (ms)
Female	2.1	863
Mixed Gender	3.8	982
Male	1.9	862
Overall Mean	2.6	903

In line with current recommendations (Nakagawa & Schielzeth, 2013), we report conditional R² as an approximation of the global effect size for the best-fitting model and partial eta squared ( η p 2 ) as an approximation of the local effect sizes for significant main effects and interactions. Conditional R² was calculated using the r.squaredGLMM() function of the MuMIn package (Version 1.43.17; Bartoń, 2020), while η p 2 was calculated using the anova_stats function of the sjstats package (Version 0.18.2; Lüdecke, 2022). Tukey- and bias-adjusted post-hoc contrasts were examined through the emmeans() function of the emmeans package (Version 1.5.0; Lenth, 2020). Bias adjustment occurred based on the estimated standard deviation of the best fitting models’ residuals, calculated via the sigma() function of the base R package stats.

The analysis of the response time data is consistent with the analysis plan pre-registered with the Center for Open Science (https://osf.io/njgs4/. This preregistration is based on a larger project, and the relevant plans are described in Part II ("Gender Priming”). The analyses presented here deviate from the pre-registered plan in two respects. First, in addition to an examination of response times, we present results for error rates. Second, while not specified in the analysis plan, participant gender, participant age, and trial number were tested for inclusion as random factors and slopes.

3.2 Results

3.2.1 Error Rate

3.2.1.1 Role Noun Primes

The model of best fit (conditional R² = 0.17, AIC = 324 [16 lower than maximal model]) indicated no significant effects of Prime Gender, Wald X² < 1, or Face Pair Gender, Wald X² = 3.56, p = 0.17, η p 2 = 0.01. Descriptively, participants produced very few errors overall (M _ERROR = 2.4%, SE = 0.4%).

3.2.1.2 Name Pair Primes

The model of best fit (conditional R² = 0.09, AIC = 304 [11 lower than maximal model]) indicated no significant effects of Prime Gender, Wald X² < 1, or Face Pair Gender, Wald X² = 1.34, p = 0.51, η p 2 = 0.01. Participants tended to respond less accurately as the block progressed (Trial Number, Wald X² = 3.09, p = 0.08, η p 2 = 0.01) and produced very few errors overall (M _ERROR = 2.2%, SE = 0.4%).

3.2.2 Response Time

3.2.2.1 Role Noun Primes

The model of best fit (conditional R² = 0.26, AIC = 17,669 [605 lower than maximal model]) indicated a small two-way interaction between Prime Gender and Face Pair Gender, Wald X² = 9.69, p = 0.008, η p 2 = 0.01 (Table 2, Figure 3 [left panel]). Participants tended to respond faster to male and female faces following gender congruent compared to incongruent primes for both male (M _DIFF = 35 ms, z = 1.80, p = 0.04 [one-tailed]) and female faces (M _DIFF = 34 ms, z = 1.75, p = 0.04 [one-tailed]). Participants responded most slowly, but equally quickly, to mixed-gender faces regardless of the gender of the preceding prime, (M _DIFF = 6 ms, z = 0.31, p = 0.76).

Table 2

Estimated Means [and 95% Confidence Intervals] for Response Time (ms) as a Function of Face Pair Gender and Prime Gender for Role Noun Primes and Name Pair Primes (Experiment 1)

Face Pair Gender	Prime Gender
	Feminine	Masculine	M DIFF
Role Noun Primes
Female	792 [721, 870]	826 [752, 906]	−34 [−185, 118]*
Mixed Gender	873 [796, 959]	867 [790, 952]	6 [−156, 169]
Male	841 [763, 926]	806 [732, 888]	35 [−125, 194]*
Name Pair Primes
Female	839 [773, 911]	891 [820, 968]	−52 [−225, 91]**
Mixed Gender	951 [878, 1,031]	918 [847, 996]	33 [−118, 184]
Male	881 [810, 958]	803 [738, 873]	78 [−63, 220]***

*p < 0.05. **p < 0.01. ***p < 0.001.

Figure 3

Estimated Means (and 95% Confidence Intervals) for Response Time (ms) as a Function of Face Pair Gender and Prime Gender for Role Noun Primes and Name Pair Primes (Experiment 1).

3.2.2.2 Name Pair Primes

The model of best fit (conditional R² = 0.28, AIC = 17,716 [612 lower than maximal model]) indicated a small significant main effect of Face Pair Gender, Wald X² = 10.55, p = 0.005, η p 2 = 0.02, which was qualified by a two-way interaction between Prime Gender and Face Pair Gender, Wald X² = 29.49, p < 0.001, η p 2 = 0.02 (Table 2, Figure 3 [right panel]). Participants responded significantly faster to male and female faces following gender congruent compared to incongruent primes. This was more pronounced for male faces (M _DIFF = 78 ms, z = 3.81, p < 0.001 [one-tailed]) than for female faces (M _DIFF = 52 ms, z = 2.52, p = 0.006 [one-tailed]). Participants responded most slowly to mixed-gender faces regardless of the gender of the preceding prime, with participants responding non-significantly faster to male names (M _DIFF = 33 ms, z = 1.46, p = 0.15).

3.2.3 Summary

The results indicated very small error rates overall for both role noun and name pair primes, suggesting a potential floor effect. In line with previous research, for gender-specific face pairs, participants responded faster when the gender of the primes (feminine, masculine) and faces (female, male) were congruent compared to when they were incongruent. Descriptively, the effect was stronger for the name pair primes, compared to the role noun primes, and was for name pair primes stronger when participants responded to male faces. For mixed gender face pairs, participants responded equally quickly to role noun primes regardless of prime gender, but responded faster to masculine compared to feminine name pair primes. Finally, participants responded more slowly to the mixed gender face pairs compared to the matched gender face pairs following both role noun and name pair primes.

4 Experiment 2

The goals for the second experiment were to test whether the results found for the gender-specific primes in Experiment 1 are reliable, to further investigate prime type differences (role nouns vs name pairs), and to explore whether mixed gender targets (i.e. mixed gender faces) can be successfully primed by gender-balanced primes. Therefore, the experiment’s design was changed as follows: Block order (Role noun block first vs Name pair block first) was added as a factor and counterbalanced by participant; prime gender was expanded by gender-balanced primes (non-stereotypical role nouns and gender-mixed name pairs), and to provide a within-block baseline, control primes (nonsense strings) were added.

With the addition of gender-balanced and control primes, the number of trials in all three Face Pair Task blocks was increased from 45 to 60. The results of the pre-test for Experiment 1 did not support enough additional faces as being perceived as feminine or masculine to match the increased number of primes. We decided to reduce the number of faces used, and consequently, participants observed each face pair twice as often (two times per block in Experiment 2, compared to one time per block in Experiment 1). As this additional complexity may alleviate the Error Rate floor effect found in Experiment 1, both Error Rate and Response Time were examined in this experiment.

4.2 Results

4.2.1 Error Rate

The model of best fit (conditional R² = 0.34, AIC = 898 [50 lower than maximal model]) indicated no significant effects of Face Pair Gender, Wald X² = 3.94, p = 0.14, η p 2 = 0.01, Prime Gender, Wald X² = 6.00, p = 0.11, η p 2 = 0.01, or Prime Type, Wald X² = 1.44, p = 0.23, η p 2 < 0.01. Participants tended to respond less accurately as the block progressed (Trial Number, Wald X² = 3.05, p = 0.09, η p 2 < 0.01) and produced very few errors overall (M _ERROR = 2.9%, SE = 0.3%).

4.2.2 Response Time

The model of best fit (conditional R² = 0.24, AIC = 57,684 [390 higher than maximal model, which reached singularity so was ignored]) indicated small significant effects of Face Pair Gender, Wald X² = 54.90, p < 0.001, η p 2 = 0.02, Prime Gender, Wald X² = 53.32, p < 0.001, η p 2 = 0.01, and Prime Type, Wald X² = 33.70, p < 0.001,and η p 2 = 0.01. These were qualified by a small significant interaction between Face Pair Gender and Prime Gender, Wald X² = 43.04, p < 0.001, η p 2 = 0.01, and a very small significant interaction between Prime Gender and Prime Type, Wald X² = 16.16, p = 0.001, η p 2 < 0.01, which were qualified in turn by a very small significant three-way interaction between Face Pair Gender, Prime Gender, and Prime Type, Wald X² = 13.79, p = 0.032, η p 2 < 0.01 (Table 4, Figure 4).

Table 4

Estimated Means [and 95% Confidence Intervals] for Response Time (ms) as a Function of Face Pair Gender, Prime Type, and Prime Gender (Experiment 2)

	Prime Gender
Face Pair Gender	Feminine	Gender-Balanced	Masculine	Nonsense
Role Noun Primes
Female	890[785, 1,010]a	896[790, 1,017]a	891[785, 1,012]a	971[856, 1,102]b
Mixed	989[875, 1,118]ab	963[852, 1,089]a	1,012[895, 1,144]ab	1,031[912, 1,165]b
Male	936[824, 1,065]a	928[816, 1,055]a	857[754, 974]b	918[807, 1,043]a
Name Pair Primes
Female	878[774, 996]a	1,022[901, 1,160]b	956[843, 1,086]c	1,032[910, 1,172]b
Mixed	1,038[918, 1,174]ab	1,039[919, 1,174]ab	984[871, 1,112]a	1,100[972, 1,244]b
Male	930[818, 1,057]ab	980[862, 1,114]a	889[782, 1,010]b	937[824, 1,065]ab

Note. Per row, means sharing a common subscript are not significantly different at α = 0.05 (two-tailed) according to contrast differences.

Figure 4

Estimated Means [and 95% Confidence Intervals] for Response Time (ms) as a Function of Face Pair Gender, Prime Type, and Prime Gender (Experiment 2).

For role noun primes, participants responded equally quickly to female faces following feminine, masculine, and gender-balanced primes – in all cases, this was significantly faster than the nonsense text baseline (feminine M _DIFF = 81 ms, z = 3.44, p = 0.003; masculine M _DIFF = 80 ms, z = 3.34, p = 0.005; gender balanced M _DIFF = 75 ms, z = 3.17, p = 0.008). Participants responded significantly faster to male faces following gender congruent compared to both incongruent primes (feminine M _DIFF = 79 ms, z = 3.51, and p = 0.033; gender balanced M _DIFF = 71 ms, z = 3.19, and p = 0.008) and to the nonsense text baseline (M _DIFF = 61 ms, z = 2.72, p = 0.033). Finally, for the mixed gender faces, participants responded significantly faster following gender-balanced primes compared to the nonsense text baseline (M _DIFF = 62 ms, z = 2.65, p = 0.041) and non-significantly faster following feminine (M _DIFF = 26 ms, z = 1.03, p = 0.731) and masculine (M _DIFF = 49 ms, z = 1.93, p = 0.217) primes.

For name pair primes, participants responded significantly faster to female faces following gender congruent primes compared to incongruent primes (masculine M _DIFF = 78 ms, z = 3.37, p = 0.004; gender balanced M _DIFF = 68 ms, z = 6.03, p < 0.001) and to the nonsense text baseline (M _DIFF = 154 ms, z = 6.46, p < 0.001). Participants responded significantly faster to male faces following gender congruent compared to gender-balanced primes (M _DIFF = 91 ms, z = 3.88, p = 0.001) and tended to respond more quickly to male faces following gender congruent compared to feminine primes (M _DIFF = 44 ms, z = 1.79, p = 0.28) and to the nonsense text baseline (M _DIFF = 48 ms, z = 2.08, p = 0.159). For the mixed gender faces, participants responded non-significantly faster following gender-balanced primes compared to the nonsense text baseline (M _DIFF = 61 ms, z = 2.19, and p = 0.127), similarly fast compared to the feminine primes (M _DIFF = 1 ms, z = 0.03, and p > 0.999), and slower compared to the masculine primes (M _DIFF = 55 ms, z = 2.12, and p = 0.148). Interestingly, for the mixed-gender faces, participants responded significantly faster following masculine compared to the nonsense text baseline (M _DIFF = 116 ms, z = 4.25, and p < 0.001).

4.2.3 Summary

Reproducing the findings from Experiment 1, the results indicated very small error rates overall, suggesting a potential floor effect, even with the additional complexity from this experiment. The response time results were complex. For the role noun primes, participants responded significantly faster to male faces paired with masculine primes than to any of the other prime genders, and the nonsense text baseline, responded significantly faster to mixed gender faces paired with gender-balanced primes compared to the nonsense text baseline, and responded equally quickly to the female faces regardless of the prime gender – in all cases, significantly faster than the nonsense text baseline. For the name pair primes, participants responded faster to the female and male faces when paired with congruently gendered primes than when paired with incongruent primes and the nonsense text baseline. This effect was much stronger for the primes paired with the female compared to male faces. For the mixed gender faces, participants tended to respond more quickly to masculine primes than to the other prime gender categories – and responded significantly faster following masculine primes compared to the nonsense text baseline. In keeping with Experiment 1, the effects were descriptively larger for the name pair primes compared to the role noun primes.

5 General Discussion

This study was aimed at exploring the possibility of simultaneous activation of more than one gender category following the perspective of GFL, which promotes language that seeks to avoid unnecessary references to a specific gender. We employed a text-to-image response priming design utilising a three-alternatives classification task that complements the usual specifically female and specifically male categories with a gender-mixed (female and male) category. Due to the very exploratory nature of this study, the first experiment focused on gender-specific priming effects, and the second incorporated an examination of gender priming for primes that do not correspond to an either-or gender paradigm.

In line with previous research applying a similar design (e.g. Kawakami & Dovidio, 2001; Lemm et al., 2005) and consistent with H1, faster response times were found for the female- and male-specific face pairs following gender congruent compared to incongruent primes (Exp. 1 [all] and Exp. 2 [all except for female face pairs following stereotypical primes]). Also, consistent with H2, the results for both experiments indicated stronger facilitation effects for the primes that provide categorical gender information (i.e. gender-matched name pairs) than those that provide probabilistic gender information (i.e. gendered role nouns). Overall, these results suggest that non-linguistic targets are successfully primed by categorical female and categorical male name pairs and show evidence of successfully being primed by probabilistic female and probabilistic male role nouns (especially, in Exp. 2, compared to the nonsense text baseline).

With reference to the gender-mixed category, in Experiment 1, responses to gender-mixed face pairs were slower than those to gender-matched face pairs – uniformly for role noun primes, but slower for feminine compared to masculine name pair primes. When the gender-balanced primes were incorporated in Experiment 2, consistent with H3, responses to gender-mixed face pairs were faster when preceded by non-stereotypical role noun primes compared to gender-stereotypical role noun primes and the nonsense text strings. When preceded by gender-mixed name pairs, responses to gender-mixed face pairs were faster compared to the nonsense text strings, but similarly fast compared to feminine name pairs and even slower compared to masculine name pairs. This suggests that mixed gender targets are successfully primed by probabilistically gender-balanced information and show some benefit from the presentation of categorical gender-balanced information, but that categorical masculine information is perceived as more relevant.

The lack of a response time priming effect in Experiment 2 for female face pairs following a role noun prime was unexpected. The response times for female face pairs following a stereotypical prime were relatively quick, with feminine, masculine, and gender-balanced primes all resulting in significantly faster responses compared to the nonsense text strings. This contrasts with the results of Experiment 1.

Since far more women than men participated in Exp. 2 while participant gender was balanced in Exp. 1, the difference in findings may indicate an in-group bias, i.e. that participants implicitly consider members of their own gender as more capable of holding any role. However, if this were the case, one could expect participant gender to significantly improve the model of best fit during model fitting, which was not found.

The difference in the findings could also be related to the different number of prime types per experiment. The use of only female and male primes in Experiment 1 may have led to a contrast effect, whereas the addition of gender-balanced primes and nonsense text strings in Experiment 2 may have led to more complex strategies involving other factors such as humanness. In addition, one-third of the prime-target combinations were congruent in Experiment 1, but only one-quarter in Experiment 2, so the role of congruency proportion would need to be explored in future studies. With this in mind, one can speculate whether the finding that women – when gender-balanced primes and nonsense text strings are included – are perceived as capable of holding roles regardless of these roles’ gender stereotypicality can be explained as a combination of historical gender roles and modern approaches to improving gender equality.

A common approach to improving gender equality in role selection (especially occupation selection) is to convey messages (Moser & Branscombe, 2023) aimed at affirming that the role can be held by anyone regardless of gender; e.g. girls can be doctors, and boys can be nurses. Historically, there have been far more “men-only” compared to “women-only” roles, meaning that the majority of gender equality messaging by necessity has reinforced the capability of women to hold any role (reflected in the equal response times between all three occupation gender stereotype categories), seemingly without reaching the same tipping point for the capability of men to hold any role. If this is true, then it can also be seen as an indicator that GFL policies (in this case, the renunciation of morphological female markers in Norwegian) can positively impact how social perceptions change over time and suggests that it would be beneficial to investigate how the perception of men’s abilities to hold counter-stereotyped roles may also be shifted.

Returning to the mixed gender targets for the name pairs in Experiment 2, the finding that participants responded most quickly following masculine name pairs, and equally quickly following feminine and gender-mixed name pairs, with responses to all three being faster than responses to the nonsense strings, is broadly in keeping with the expected match-mismatch bias. In other words, it is possible that the names provided a general processing advantage compared to the nonsense strings, with a second processing advantage for masculine compared to feminine names. It is possible that a third processing advantage of the conceptual match between the gender-mixed names and mixed-gender faces was masked by the match-mismatch effect as participants had to spend more cognitive energy, and therefore time, in correctly interpreting the prime before responding. As such, it would follow that if single words (e.g. gendered personal pronouns) were used as categorical primes instead, thus avoiding the match-mismatch bias, the priming effect of the categorical gender-balanced primes might be much stronger. However, since the number of these forms per category is very limited, each prime would need to be repeated multiple times. This would potentially introduce a repetition bias, would strengthen the bias that each individual prime would have as an outlier, and would severely complicate comparisons of the categorical and stereotypical primes.

In relation to the question of why masculine name pair primes would lead to an improvement in response times for mixed-gender face pairs, some research (e.g. Koivula, 2001) has found that masculine items create far richer cognitive representations than feminine items. As the gender-balanced name pair primes contained one male and one female name, it could be that this provides a comparative processing time advantage when observing ambiguous visual targets.

The priming effects found in both experiments were relatively large compared to similar priming studies (e.g. Kawakami & Dovidio, 2001; Lemm et al., 2005). It is possible that this is related to the length of the SOA, as longer SOAs (such as those chosen by us) generally appear to produce larger effects (e.g. Avneon & Lamy, 2018; Greenwald et al., 1996). It would be interesting to investigate whether a shorter SOA, more in line with conventional response priming studies but, following Hutchison et al. (2008), lacking backwards associative strength, would yield the same or different results. If comparable results were found, they would suggest a strong forward prime-target associative strength – powerful enough to surpass any backward associative strength effects – and would allow for an assessment of semantic feature (e.g. gender) strength for specific primes (e.g. role nouns, first names) via manipulation of SOA. Another possible explanation for these comparatively large priming effects is the experimental design; the experimental setup modelled the face pair task as closely as possible, and stimulus selection and development was also based on tasks using response time measures, which may provide more accurate measures than commonly used approaches such as explicit ratings. A further possible explanation is that this experiment increased in difficulty across experimental tasks, which would allow participants to become familiarised with the task. However, the effect of Trial Number was only strong enough to be added to the models of best fit for the Error Rate analyses, suggesting that any effect of Trial Number on Response Time would be minimal. Another possible explanation is that the inclusion of the data from the Face Pair Task without primes (Block 2) in the analyses may have removed a large amount of noise related to idiosyncratic biases (e.g. between-participant differences in which face gender is easier to recognise) or personal factors (e.g. handedness, factors related to motoric ease of response registration).

As the findings of this study indicate that non-stereotyped role nouns successfully prime mixed-gender targets, but that gender-mixed first name pairs appear less effective than masculine information in priming mixed-gender targets, these results have practical and theoretical implications for gender-fair language policies and open the door for further examinations of how different forms of a) gender-balanced information, and b) truly gender-neutral information affects our perceptions. Research based on linguistic-visual paradigms can thus complement current research on the gender associations of new linguistic forms, which relies heavily on estimation tasks (e.g. Pozniak et al., 2023; Schunack & Binanzer, 2022; Tibblin et al., 2023; Xiao et al., 2023) and reading tasks (e.g. Stetie & Zunino, 2022; Tibblin et al., 2023). Further, as exemplified by Sato et al. (2016), linguistic-visual paradigms are very usable in cross-linguistic experimentation, as the same target stimuli can be used across many different language contexts, allowing for more direct comparisons to be drawn. Similarly, there is the potential for these advantages to be extended to comparisons between spoken and written modes of language comprehension, and to access priming effects across different levels of literacy (e.g. due to differences in age or educational opportunities). As such, these findings indicate that the presented three categorisation-alternatives face pair task is well positioned to be used in exploring the impacts of GFL cross-culturally and cross-lingually.

In conclusion, our results replicate gender-specific priming effects for both gender categorical primes (name pairs) and gender probabilistic primes (gender stereotyped role nouns), with the stronger priming effect found for the categorical primes, as expected. The results also indicated that non-stereotyped role nouns successfully primed responses to mixed-gender faces, although the same was not found for the gender-mixed name pairs. As the non-stereotyped role nouns can be interpreted as a microcosm of GFL focusing on equally activated gender beliefs, these results offer some preliminary support for the idea that GFL can positively impact our perceptions of and interactions with others outside of the linguistic domain and support the use of the face pair task as a method of further exploration. This study provides important insights into how linguistic information can subtly influence our perceptions of the people around us. While these findings cannot be used directly to improve gender-sensitive language, they provide a starting point on which future interventions can build.