In French, two main strategies are used to form GFFs: (a) the re-feminisation strategy that explicitly includes women and men by using double forms, and (b) the neutralisation strategy that implicitly includes all genders as it makes use of forms that lack gender markings corresponding to a person’s gender (Elmiger, 2008; Simon & Vanhal, 2022). Double forms can be either complete, les voisines et voisins ‘the neighbours._FEM and neighbours._MASC’ or contracted by means of different typographical signs, e.g. les voisin·es ‘the neighbours._MASC·FEM’. The neutralisation strategy mainly uses three different types of nouns: collective nouns, epicene nouns, and common gender nouns. Collective nouns, like le voisinage ‘the neighbourhood’, implicitly refer to all members of the group independently of their gender. Epicene and common gender nouns resemble each other but differ in one important aspect, namely, in how they handle grammatical gender (Corbett, 1991). On the one hand, epicene nouns can only take one grammatical gender, as une personne ‘a._FEM person’ or un individu ‘an._MASC individual’. On the other hand, common gender nouns accept both masculine and feminine agreement (one can say une locataire ‘a._FEM tenant’ as well as un locataire ‘a._MASC tenant’), but the form of the noun does not change. As French plural articles are not marked for grammatical gender, common gender nouns can be considered gender-neutral in their plural form, les locataires ‘the tenants’.3

Both strategies have been studied empirically using different experimental methods with the aim of investigating their effects on speakers’ mental representations of gender. As for the re-feminisation strategy, studies have repeatedly found that complete and contracted double forms increase representation of women in German (e.g., Braun et al., 2005; Hansen et al., 2016; Irmen & Roßberg, 2004; Körner et al., 2022; Schunack & Binanzer, 2022) and in French (e.g., Brauer & Landry, 2008; Tibblin et al., 2023; Xiao et al., 2022). The few existing studies on French gender-neutral forms indicate that these forms are successful in attenuating the male bias induced by the masculine form, at least as long as there is no interference with stereotype information (e.g., Kim et al., 2022; Richy & Burnett, 2021; Tibblin et al., 2023).

The studies that have examined German gender-neutral forms are more numerous. However, the results of these studies are mixed, as some studies found gender-neutral forms to resolve the male bias (Sato, Gabriel, et al., 2016; Stahlberg et al., 2001), while other studies found that these forms reproduce the male bias (Braun et al., 2005; Irmen, 2007; Irmen & Roßberg, 2004). There could be several explanations for this discrepancy. First, different methods (e.g., eye-tracking, self-paced reading, sentence evaluation task, questionnaires) and materials have been used. Second, the number of participants and the statistical analyses conducted do vary. This variation could lead to differences in statistical power, and it might therefore be difficult to compare the results of these studies.

Considering the varying results of the studies on gender-neutral forms, one could expect the double forms to be more effective in increasing representation of women compared to the gender-neutral forms. To the best of our knowledge, only three studies (Irmen & Roßberg, 2004; Stahlberg et al., 2001 (Exp. 1); Tibblin et al., 2023) have compared double and gender-neutral forms. While Irmen and Roßberg (2004) found significant differences between complete double forms and gender-neutral forms using common gender nouns in German, the two other studies did not find any statistically significant differences between double and gender-neutral forms, meaning that both GFFs were equally effective in attenuating the male bias induced by the masculine form (Stahlberg et al., 2001 (exp. 1); Tibblin et al., 2023).

When considering anaphoric relationships, a broad and common assumption in text processing is that as we read, we construct mental models of the situation described in the text to maintain local and global coherence (e.g., Garnham, 2001; Graesser et al., 1994; Gygax & Gabriel, 2011). A mental model is a partial representation of a real or imaginary world that readers build while reading (Garnham, 2001). When reading, there are two main sources of information that influence the construction of our mental models: one is grounded in the text itself (bottom-up), and the other in our knowledge of the world4 (top-down) (Gygax & Gabriel, 2011). Over the course of reading, the mental model will be updated as new information is presented in the text. These updates can be based on new information in the text itself, or on inferences made about the world represented in the mental model.

Thus, when constructing a mental model of the nurses in a sentence like The nurses came out of the hospital, both information present in the lexeme nurses (such as its lexical meaning and the plural mark signalling the presence of several nurses) and information drawing on background knowledge contribute to the mental model. Based on their background knowledge, readers might infer that the majority of the nurses are women, given that they live in a society where nurses are – or are believed to be – women more often than men. In a natural gender language like English, the surface form of the role noun nurses will be unlikely to influence the construction of its mental representation with respect to the nurses’ gender, but this is not the case in grammatical gender languages like French. In these languages, the form of the role noun will influence how the group is mentally represented also with respect to gender (Carreiras et al., 1996). Reading infirmières ‘nurses._FEM’ should lead to a model representing only female nurses, while infirmiers ‘nurse._MASC’ could either lead to a model representing only male nurses (specific interpretation) or to a model representing a group of female and male nurses (generic interpretation). When comparing English speakers to French and German speakers, researchers (Gygax et al., 2008) found that, in French and German, the surface form of a role noun had a stronger influence than the stereotype information conveyed by the role noun, a factor that is based on the participants’ knowledge of the world. In other words, when constructing their mental models, English readers relied on their background knowledge – or belief – of nurses often being women, while French and German-speaking readers were more influenced by the masculine form – resulting in a male bias in their mental representations of gender.

To further understand how anaphoric relationships work within mental models, the notational instruments from Reboul and Moeschler (1998) are particularly useful. Although it is generally assumed that mental models do include all elements pertinent to the situations depicted in the text (Garnham, 2001), Reboul and Moeschler (1998) offer tools and a terminology that enable us to make clear predictions as to the smaller entities – such as objects and people – included in readers’ mental representations (MRs). These notions are useful when discussing anaphora resolution and provide a solid ground for understanding how gender information works in anaphoric resolution, which is central in the present paper. As such, two operations discussed by Reboul and Moeschler (1998) are particularly relevant to the current study, namely the grouping and the extraction operations. These operations are exemplified with the following sentences:

When reading NP1 in example (1a), two MRs are created: [@woman]5 and [@man], but when reading the pronoun they, a new MR [@woman&man] is created by the process of grouping the two former ones. Inversely, an extraction process starts with one MR from which two or more new MRs are extracted. As such, after reading NP2 in (1b), the MR [@marbles] is created. But as the reading continues, the reader must distinguish the marbles that were found from the one that was not. Therefore, two new MRs are extracted from the existing one: [@8marbles] and [@1marble]. Both grouping and extraction operations are involved when interpreting anaphoric relations, to which we will turn in the following section.

Since anaphoric expressions take their meaning from an entity previously mentioned in the text, the creation and subsequent manipulations of MRs play an important role when processing and interpreting anaphoric expressions (Garnham, 2001). The investigation of anaphoric expressions, such as she in (2) below, can therefore provide us with useful insights when studying mental representations of gender.

In order to successfully resolve an anaphoric expression, readers must go through three stages. First, they must identify the anaphor, she. Second, they must determine the meaning of the anaphor with the help of linguistic cues about, for example, gender (feminine) and number (singular). After these two stages, readers can match the anaphoric expression to the correct antecedent, Sally (Garnham, 2001).

In psycholinguistic studies, the resolution of a gendered anaphoric expression with a possible antecedent presented in the masculine form has often been studied to investigate how the masculine form is interpreted by speakers of gendered languages. For example, sentence pairs like (3) have been used:

The task for the participants is to evaluate whether sentence (3b) is a possible continuation from sentence (3a) or not. As we will see, this judgment relies on the possibility of readers’ establishing an anaphoric relation between the subjects of the two sentences, a possibility that ultimately relies on the gender included in the MRs that are created during reading.

Figure 1 shows how the two different interpretations of the masculine form would lead to two different MRs and, in turn, to two different outcomes in a sentence evaluation task with sentence pairs such as (3). As we can see, when reading (3), readers should, on the basis of NP3, first create a MR of a group of neighbours. Subsequently, when encountering the anaphoric expression (i.e., NP4) in (3b), they must recall and confront the MR created while reading NP3 in (3a) in order to assess if there is a possible antecedent. Since the two possible interpretations of the masculine form, the specific and the generic, differ in the gender-related information they contain, the interpretation of NP3 selected will have consequences for the possibility of resolving the anaphoric expression in NP4. If the readers have made a specific interpretation of NP3, the MR [@voisins(MASC)] should be activated. Since all the neighbours in this MR are men, it would not be possible to extract [@voisines(FEM)] from it, which would be needed in order to interpret NP4 as an anaphoric expression with NP3 as its antecedent. However, a generic interpretation of NP3 would activate the MR [@voisins(MASC&FEM)]. Since this MR contains both women and men, it should be possible to extract both [@voisines(FEM)] and [@voisins(MASC)] from it. In this case, when the reader encounters NP4 and is forced to create a new MR of the neighbours, it should be possible to make the anaphoric link back to NP3. In brief, a generic interpretation of voisins should allow for a quick and easy extraction of [@voisines(FEM)], requiring little cognitive effort. Inversely, a specific interpretation would not allow for a successful extraction process and, consequently, not allow for an establishment of an anaphoric link back to NP3.

If GFFs are used instead of the masculine form, the anaphora resolution process can be predicted to be rather different. Using complete double forms, NP5 in example (4) should activate the two separate MRs [@voisines(FEM)] and [@voisins(MASC)], analogous to example (1) above. Since one of the MRs created on the basis of NP5 corresponds perfectly to the information presented in the anaphoric expression, no operation needs to be executed on the MR and the anaphoric expression can be resolved effortlessly. As for the contracted double forms, these are morphologically abbreviations of the complete forms, and this is also in line with the use officially recommended in France (Haut Conseil à l’Égalité entre les Femmes et les Hommes, 2022). Therefore, the processing and interpretation of the contracted double forms, NP6 in example (4), should be identical to the interpretation of NP5 just described.

The last type of GFF examined in this study are the gender-neutral forms. These forms included common gender nouns like NP7 in example (4), collective nouns, and epicene nouns (see Table 1 for examples). Since NP7 contains no explicit information regarding the referents’ gender, the MRs created when reading NP7 should include both genders, and thus resemble the generic interpretation of the masculine form described above. Consequently, NP7 should activate the MR [@locataires(MASC&FEM)]. The process of extracting the MR [@locataires(FEM)] should therefore always be possible, similar to the case in which the masculine form is generically interpreted. If epicene or collective nouns are used instead of common gender nouns, the process should be the same, seeing as they all implicitly include referents of all genders.

The sentence evaluation task provides two outcome variables: the participants’ judgments of the second sentence (yes or no) and the response times (RTs). The judgments reflect whether it was at all possible for participants to establish an anaphoric relation between the anaphoric expression and its antecedent. Positive judgment times were analysed (in line with previous research) and give us insight as to how easy the process of establishing the anaphoric relationship was.6

First, we hypothesised that there would be a significant two-way interaction between the variables Form and Anaphor gender, such that when women was used as an anaphor, the GFFs would lead to more positive judgments and faster RTs, compared to the masculine form. When men was used anaphorically, we did not expect any differences between the GFFs and the masculine form (Hypothesis 1).

Second, we hypothesised that when comparing the GFFs to each other, there would be significant differences between them, such that when women was used as an anaphor, the double forms would lead to significantly more positive judgments and faster RTs compared to the gender-neutral forms (as in Irmen & Roßberg, 2004). Again, we did not expect such differences with men as an anaphor (Hypothesis 2).

Finally, we expected a significant three-way interaction between the variables Form, Anaphor gender, and Attitudes towards gender-fair language, such that when the role noun was presented in its masculine form, positive attitudes would lead to fewer positive judgments and slower RTs when the anaphor women was used (as in Braun et al., 2005). We did not expect any effects of attitudes towards gender-fair language in the GFF conditions or when men was used as an anaphor (Hypothesis 3).

In brief, we expected (i) that the GFFs, compared to the masculine form, would increase the chances of resolving the anaphoric expression women and speed up this process, (ii) that the double forms would increase the chances of a successful anaphoric resolution of women as well as accelerate this process compared to the gender-neutral forms, and (iii) that positive attitudes towards gender-fair language would decrease the chances of resolving the anaphoric expression women with a masculine antecedent and slow down this process.

One hundred and sixty-three right-handed native French-speakers participated in the experiment. They were recruited on Prolific, an online participant pool that provides high quality data (Peer et al., 2022), requires fair payments to the participants, and allows researchers to do fine pre-screening. The pre-screening questions, whose exact wording can be found in the pre-registration, concerned the participants’ handedness, minimum age, language disorders and linguistic background. All participants were paid £2.45. Prior to analyses, 10 participants were removed from the sample due to high error rates concerning the filler items (see 4.3 below). The final sample therefore consisted of 153 participants (mean age = 28 years, SD = 9 years; 41% students). 84 of these were women, 64 men, 4 of another gender identity and 1 did not wish to state their gender. The majority (72%) of the sample lived in France, but some lived in the French-speaking regions of Belgium (10%), Switzerland (5%) or Canada (3%). The remaining 10% lived in a non-French-speaking country or region.

The materials consisted of 22 experimental and 44 filler sentence pairs. These were based on the materials used by Gygax et al. (2008) and followed the structure of (3) above. Thirty-three of the filler sentence pairs were incongruent, requiring a negative response, and eleven were congruent, requiring a positive response. All sentence pairs were presented in a randomized order.

The participants’ task was to read and understand the first sentence, then to press the space bar to make the first sentence disappear and the second one appear. Once they had read the second sentence, they should decide, as quickly and accurately as possible, whether it was a possible continuation of the first sentence. It was explicitly stated that the group described in the second sentence was part of the one mentioned in the first sentence. The participants were told to keep their thumbs on the space bar and their index fingers on the keys corresponding to yes and no throughout the experiment. To minimise the risk of including trials where participants might not have been paying attention, the experiment automatically continued if no key was pressed within 10 seconds. Before each sentence pair, a screen asking whether the participants were ready to continue was presented. This screen was presented until the space bar was pressed, and participants were then allowed to take a break if needed. The experiment started after six familiarisation trials.

The 22 role nouns used as antecedents in the experimental sentences were selected on a criterion of non-stereotypicality based on a previous study that collected stereotypicality norms of role nouns in their masculine and GF forms (Tibblin et al., 2023), such as joueurs de tennis ‘tennis players’ or militants écologistes ‘climate activists’. Thus, the average estimated percentage of women in each role noun ranged from 39% (les athlètes ‘the athletes’) to 64% (les personnes qui montent à cheval ‘the people who go horseback riding’) (M = 49%, SD = 13%). When the role nouns were presented in the complete double form, the feminine form always preceded the masculine form (for example les voisines et voisins ‘the neighbours._FEM and the neighbours._MASC’).

We created two lists for each level of the variable Form. In each list, half of the target sentences contained women and the other half men. If a role noun was followed by women in the first list, it was followed by men in the second list, and vice versa. The context-setting content of the sentences did not change between the two lists, and all lists included all 44 filler items. Each participant was randomly assigned to one of the four forms and within each form, to one experimental list.

Upon completing the experiment, participants were presented with questions regarding GFFs, an eight-item scale measuring attitudes towards gender-fair language (Tibblin, 2020; Tibblin et al., 2023) (α = .92, 95% CI[.90, .94]), and a set of socio-biographical questions. Complete lists of both filler and experimental items, details regarding their content, as well as further procedure details, the attitudes scale, and all questions presented to the participants can be found in the pre-registration.

The data were analysed with R (R Core Team, 2021) and all analyses described are in line with the pre-registration unless otherwise stated. The judgments and the RTs were analysed separately, and the preparations are presented accordingly in this section. First, the percentage of correct responses for the two types of filler items (congruent and incongruent) was calculated for each participant. Before doing so, filler responses were coded as incorrect if the RT of the first or second sentence was below the set minimum threshold (300 ms) or hit the maximum time limit (10,000 ms). After this procedure, we removed all trials of the participants (n = 10) who responded incorrectly to more than 50% of the congruent or the incongruent fillers. Following this first data screening, we removed all trials in which the RT of the first or second sentence was above or below the set threshold (i.e., all analysed trials had a RT of the first and second sentence between 300–10,000 ms). This second data screening led to a removal of 0.95% of the data.

Before analysing each outcome variable, we calculated a mean score reflecting each participant’s attitudes towards gender-fair language. Missing values (i.e., No option suits me) were excluded and the scores of items 1–3 and 6–7 were inversed prior to the calculation. This mean score served as a predictor in the upcoming analyses. This variable and the Sentence length in characters variable were centred to their mean (M = 3.4 and M = 70.1, respectively).

Before presenting the results of the inferential analyses, we present descriptive data grouped by the form of the role noun and the anaphoric expressions used. As the leftmost bars in Figure 2 show, the proportion of positive judgments in the masculine condition (les voisins) is lower when women was used as an anaphor compared to when men was used. However, when complete double forms (les voisines et voisins) were used, there is barely any difference between the anaphoric expressions. Interestingly, the contracted double forms (les voisin·es) produced the opposite pattern compared to the masculine form. In fact, the proportion of positive judgments is lower when men was used as an anaphor compared to women. Finally, the gender-neutral forms (le voisinage, les locataires) led to fewer positive judgments overall and to a slightly lower proportion of positive judgments when women was used anaphorically compared to men. In sum, these descriptive results suggest that the GFFs attenuate the male bias induced by the masculine form, but not all in the same way.8

The descriptive results previously described are to a large extent confirmed by the inferential analyses, summarised in Table 2. The intercept in this table refers to the odds ratio, i.e., the exponent of the model’s coefficients, of the following values: men as Anaphor gender, masculine as Form, and the Attitudes variable centred to its mean. An odds ratio of 1 indicates that the odds of a positive response are the same as the odds of a negative response, while a value greater than 1 indicates higher odds of a positive response. Inversely, an odds ratio less than 1 indicates lower odds of a positive response. The model of best fit included a significant interaction between the variables Form and Anaphor gender, as well as a significant main effect of the variable Attitudes towards gender-fair language. Due to convergence errors, the final random-effects structure provided by the analyses described above only included random intercepts of the variable Participant. However, we compared this final model with an identical one including random intercepts of both Participant and Role noun. Since the AIC was lower when two random variables were included (AIC = 1975, compared to 2099), and since it was significantly better according to an analysis of variance (p < .001), we included random intercepts of both Participant and Role noun in the final model. The significant main and interaction effects were the same in both models.

The variable Form was qualified by a small main effect (partial η² = .02). More precisely, the simple effect of Form: contracted double (line 3) shows that the odds of a positive judgment after the anaphor men (the reference value) are significantly lower when the antecedent is presented in a contracted double form (les voisin·es) compared to when it is presented in the masculine form (les voisins), but not when complete double or gender-neutral forms are used (lines 2 and 4). Furthermore, the variable Anaphor was qualified by a close to small main effect (partial η² = .009) and the simple effect of Anaphor: women (line 5) indicates that when the antecedent is presented in the masculine form, the odds of obtaining a positive judgment after the anaphor women are significantly lower than after the anaphor men. The very small main effect (partial η² = .003) of Attitudes towards gender-fair language (line 6) shows that positive attitudes increased the odds of a positive judgment, regardless of the anaphor and the form used. Thus, the more positive a participant’s attitude towards gender-fair language was, the more likely they were to respond yes. This holds for both women and men anaphors, and for all forms of the antecedent. Finally, the effect size of the interaction between Form and Anaphor gender was almost medium-sized (partial η² = .053). In fact, all the GFFs in combination with Anaphor: women led to significant interaction effects (lines 7–9), and all coefficients are larger than 1. Thus, when a GFF is used, the likelihood of a positive judgment following women increases significantly if one of the GFFs is used instead of the masculine form. The predictions made by this model are visualised in Figure 3. The marginal R² value indicates that the fixed effects of this model account for 13% of the variance, while the fixed and random effects together account for 41% of the variance.

When the different forms were compared against each other within each anaphoric expression, contrast analyses revealed that when men was used as an anaphor, the contracted double forms led to a significantly lower probability of a positive judgment compared to all the other forms (ps < .05). When women was used anaphorically, the masculine form led to a significantly lower probability of a positive judgment compared to all GFFs (ps < .01), and both double forms led to a probability that was significantly higher compared to the gender-neutral form (ps < .05).

The comparisons between the anaphoric expressions within each form revealed significant differences between the two anaphoric expressions in the masculine, gender-neutral, and contracted double forms. Within the masculine and gender-neutral forms, the probability of a positive judgment was significantly lower when women was used compared to men (ps < .01). Within the contracted double form, men led to a significantly lower probability compared to women (p < .001).

The mean RTs of the positive judgments, grouped by form and anaphor gender, are presented in Figure 4. These means suggest that when participants were able to establish an anaphoric relation between women and an antecedent presented in the masculine form, it took them a longer time to do so. As this difference is not observed when GFFs are used, GFFs seem effective in neutralising this effect. Interestingly, it seems to require slightly more cognitive effort to establish an anaphoric relationship between men and an antecedent presented in a contracted double form than women. Taken together, these descriptive results reveal patterns similar to the judgment data.

A summary of the final model that best fit the RT data is presented in Table 3. The intercept in this summary refers to the log-transformed RT of the following values: men as the anaphoric expression, masculine as the form, the Attitudes and the Sentence length variables are centred around their means, and trial number is set at 1 (i.e., the first trial of the experiment). The fixed-effects structure of the model of best fit produced by the analyses, described in Section 4.3, contained an interaction between the variables Form and Anaphor gender, and main effects of the variables Attitude, Sentence length and Trial number. It also included Participant and Role noun as random intercepts.

First, none of the simple effects of the GFFs (lines 2–4) were significant, signalling that when men was used anaphorically, it did not take more time to judge the sentence as possible with a GFF antecedent compared to a masculine antecedent. However, the simple effect of the anaphor women was significant (line 5), indicating that in the masculine condition, RTs were slower when women was used as an anaphor compared to when men was used. The small negative main effect of attitudes towards gender-fair language (line 6) indicates that the more positive participants’ attitudes were, the faster their judgments were. No interaction including attitude was found, suggesting that the attitude effect was roughly constant across both anaphoric expressions and all antecedent forms. Furthermore, significant main effects of Sentence length (line 7) and of Trial number (line 8) were found, signalling that RTs increase as the number of characters in the sentence increases and that RTs decrease over the course of the experiment. However, the sizes of these effects were very small (d = 0.015 and –0.012, respectively).

Finally, we found significant interaction effects of all GFFs in combination with the anaphor women (lines 9–11). Thus, the difference between the RTs following the different anaphoric expressions is significantly smaller in the GFF conditions compared to the masculine condition. The interaction effects of the complete double and gender-neutral forms in combination with the anaphor women were small (d = –0.293 and –0.225, respectively), while that of the contracted double forms was close to medium-sized (d = –0.444). The response times predicted by the model are visualised in Figure 5. As indicated by the marginal and the conditional R² values, the fixed effects of this model account for 8% of the variance, while the fixed and random effects together account for 54%.

The contrast analyses between each form within each anaphoric expression revealed that no form led to significantly faster or slower RTs when compared to each of the other forms (ps > .05). However, the comparisons between the anaphoric expressions within each form revealed that within the masculine form, women led to significantly slower RTs compared to men (p < .001). Inversely, within the contracted double forms, men led to significantly slower RTs compared to women (p = .001). No such differences were found within the complete double and the gender-neutral forms (ps > .05).

Finally, we studied the effects of participants’ attitudes towards gender-fair language on their MRs of gender. Hypothesis 3, based on Braun et al. (2005), was that the masculine form would produce a stronger male bias among the participants with positive attitudes compared to those with negative attitudes. In contrast to this hypothesis, we found that positive attitudes activated more gender-balanced MRs, as shown by higher odds of positive judgments and faster RTs regardless of the form of the antecedent or the gender represented in the anaphoric expression. These results stand in contrast to two previous studies that have explored the link between attitudes towards gender-fair language and MRs of gender. This difference might be explained by the methodological differences between our study and theirs in how the variable was analysed. While we kept attitudes as a ratio-scaled variable, Braun et al. (2005) transformed it into a categorical variable by dividing the participants into two groups representing positive or negative attitudes. In addition, the authors investigated the difference between two different forms, while we investigated four different forms, and the statistical power of our analyses is therefore not as strong as theirs. Thus, it is possible that the lack of an interaction in our data is partially due to a lack of statistical power in the present study. In contrast to Braun et al. (2005), Tibblin et al. (2023) found that negative attitudes towards gender-fair language led to an increased perceived general percentage of women when participants were asked to estimate the percentage of women in a set of role nouns. Since most studies that have investigated the effects of attitudes towards gender-fair language on MRs of gender have yielded different results, we can only conclude that the role of attitudes is still not clear, and more research is needed. All studies presented so far have measured attitudes explicitly by means of a questionnaire, and it would therefore be very interesting to measure them implicitly, for example, with an implicit association test (Greenwald et al., 1998), or to conduct in-depth interviews (Gerson, 2020) to gather a deeper understanding of what these attitudes actually reflect.