Glossa Psycholinguistics

2.2 Predictions

If bias in gender inferences is limited to cases when no other direct information about gender is provided, we would expect people to use probabilistic gender information to infer gender when this information is available (First and Full Name conditions), resulting in rates of she responses that match the gender distributions of the first names. A bias to infer characters as male would only appear when a direct cue to gender is not provided, resulting in a bias towards he responses only in the Last Name condition.

Alternatively, if the bias to assume people are male occurs even when probabilistic cues to gender are available, we would expect characters to be more likely to be inferred as male than female in all three conditions, with combined effects of the first name’s gender associations and a bias in gender inferences in the First and Full Name conditions. If this is the case, while she responses will increase as the first names become more feminine, the rate of she responses will be lower than predicted by the gender associations of the first names.

A secondary question was whether introducing a character with the full name (e.g., Jordan Smith) rather than the first name only (e.g., Jordan) would attenuate the influence of the gender information carried by the first name. This question was motivated by the observation that, in English, it is more common to refer to men than women by their last names (Files et al., 2017; Rubin, 1981; Stewart et al., 2003; Takiff et al., 2001; Uscinski & Goren, 2011), and thus the full name may act as a cue to masculinity.

2.3 Results

Responses that used he/him/his pronouns to refer to the named character were categorized together and will be referred to as he responses. Likewise, responses that used she/her/hers pronouns to refer to the named character are categorized as she responses. Responses that did not use a gendered pronoun were coded as other; these responses most commonly repeated the character’s name (e.g., After making coffee…Jordan sat down), but also included responses with no grammatical subject (e.g., …sat down) and responses not referring to the character (e.g., …it started raining). Uses of singular they were infrequent (Supplement §2.3). For the First Name and Full Name conditions, the rates of he and she responses were roughly equal, following the balanced distribution of first names in our stimuli (Table 1). In the Last Name condition, responses overwhelmingly biased towards he. Notably, in the Last Name condition, she responses were slightly less common than other responses that did not gender the character.

Responses were analyzed using logistic mixed-effect regression models with lme4 in R (Bates et al., 2015), predicting the log odds of she responses (coded as 1) as opposed to he and other responses (coded as 0). Other responses were coded as 0 because they were not frequent enough to be placed in a third category. Participant and item were included as random intercepts, with items defined as the unique first, last, and first + last name combinations. Treating the names as the random items meant that the condition manipulations were fully between-participant and between-item, so fitting a random slope model was not possible. The fixed effect of Condition was coded with orthogonal Helmert contrasts, with the first contrast comparing the Last Name condition to the First and Full Name conditions, and the second contrast comparing the First Name condition to the Full Name condition. All models are reported with Bonferroni corrections for multiple comparisons. Overall, participants were less likely to respond she than he and other (β = –1.43, z = –4.65, p < .001). Participants in the First and Full Name conditions were more likely to produce she than participants in the Last Name condition (β = 2.82, z = 4.03, p < .001). The comparison between First and Full Name conditions was not significant (Table 2).

The second model included each first name’s normed Gender Rating as a covariate (Table 3). This analysis included the First and Full Name conditions only, as the Last Name condition did not contain first names. Condition was coded with mean-centered contrasts, comparing the First and Full Name conditions. The Gender Rating for each first name was mean centered, with positive numbers more feminine and negative numbers more masculine.

Figure 1 shows the proportions of he, she, and other responses for the First and Full Name conditions by the Gender Rating of the first name. As the rating of the name became more feminine, she responses increased (β = 1.59, z = 21.97, p < .001). However, inspection of the data shows that the increase in she responses was not symmetrical to the increase in he responses. In the mostly-feminine range of first names (1 to 2 on the X-axis), he responses outnumbered other responses. In the mostly-masculine range (–3 to –2), she responses occurred at similar rates as other responses. Particularly in the First Name condition, she responses did not surpass he responses until the first name in the prompt was biased somewhat feminine, rather than at the midpoint on the scale. With mean-centered fixed effects, the significant intercept term (β = –0.51, z = –4.28, p < .001) reflects overall fewer she than he and other responses, and the effect of Condition (β = 0.53, z = 2.22, p < .05) reflects more she responses in the Full Name than the First Name condition. The interaction between Condition and Gender Rating was not significant, indicating that the effect of Gender Rating was of a similar magnitude in the First and Full Name conditions.³

Figure 1: Experiment 1: Proportions of he (blue), she (red), and other (gray) responses in the First and Full Name conditions by the mean-centered gender rating of the first name. Points indicate means for each of the 21 first names, and solid lines indicate a smooth function on the raw data. On the x-axis, 0 is the mean of the 21 names, and the dashed line indicates the center of the original response scale in the norming study.

An additional analysis examined if including other responses with he responses impacted our findings. The results revealed the same patterns as above, with the following exceptions: In the model testing the effects of Condition and Gender Rating, the intercept (β = –0.22, z = –1.75, p = .08) and the difference between the First and Full Name conditions (β = –0.25, z = –1.57, p = .12) were not significant. A second exploratory analysis added a quadratic effect of Gender Rating to evaluate whether the increase in she responses as a function of Gender Rating was nonlinear. For example, we might expect to see an effect of Gender Rating that is weaker at the endpoints (strongly gendered names) than at the midpoint (androgynous names), with a larger he response bias for androgynous names than for strongly gendered names. The quadratic effect of Gender Rating was not significant, nor did it significantly interact with Condition, inconsistent with this possibility. A final exploratory analysis included participant gender as a covariate, testing if male participants showed a larger he response bias; this analysis revealed no significant effects after Bonferroni corrections for multiple comparisons. These three analyses are discussed in more detail in Supplement §2.3–2.5.

2.4 Discussion

We investigated whether the form of reference—first name, last name, or full name—affected people’s inferences about a character’s gender, measured through the pronouns they used to complete a sentence referring to the character. When participants were not given explicit cues to gender (Last Name condition), participants overwhelmingly used he to refer to the character. Moreover, in the Last Name condition, participants were approximately equally likely to not use a gendered pronoun at all (other responses) as they were to use she. Although probabilistic cues to the referent’s gender did shape inferences, with more she responses when a first name was given (First and Full Name conditions), inspection of the data indicates that the bias towards he responses persisted. A character’s name needed to be more strongly feminine for participants to preferentially refer to them with she. In addition, participants showed a pattern of asymmetry for mostly-masculine and mostly-feminine names. In the First and Full Name conditions, androgynous names that leaned feminine (e.g., Jackie) still elicited he responses, but androgynous names that leaned masculine (e.g., Chris) elicited other responses, rather than she responses. As the first names became more feminine, the rate of she responses remained flat and parallel to the rate of other responses, then increased more sharply, whereas the rate of he responses decreased more gradually. We also hypothesized that introducing a person with a first and last name would attenuate the gender cue from the first name, such that the preference for he responses would be greater in the Full Name condition as compared to the First Name condition. The data were not consistent with this prediction; instead, the preference for he responses was numerically larger in the First Name condition.

The fact that the gender ratings of the first names predicted she responses indicates that participants were willing to produce feminine language forms in this task. The observed bias towards masculine language forms instead points to biased inferences about gender. However, one potential concern with this interpretation is that the pronouns produced in reference to the characters may not entirely match participants’ underlying inferences about the characters’ genders. Instead, it is possible that some he responses in the Last Name condition come from generic masculine usage, with participants producing he in an ostensibly gender-unspecified manner. In the 19^th century, the generic masculine was prescribed as correct, explicitly replacing alternatives like singular they and he or she that had been in use. This was contested by feminists in the 1970s and 1980s, who argued that this language was not inclusive and perpetuated biases of masculine as the default (Bodine, 1975). While this guidance has been replaced in formal language policies by he or she and occasionally singular they constructions (APA, 2019; APA Publication Manual Task Force, 1997; Robertson, 2021), some speakers retain the generic masculine usage. If so, some instances of he responses in the data – particularly those in the Last Name condition, where no direct information about the character’s gender is included – may reflect this generic use. To provide a more direct test of the influence of referential form on gender inferences per se, in Experiment 2 we ask participants to make explicit inferences about the referent’s gender.

3.2 Predictions

The results of Experiment 1 indicate that the bias to infer characters as male was not eliminated when probabilistic information about the gender of a character – given by their first name – was provided. Instead, the findings support a model of gender inference where probabilistic cues about gender are combined with a people=male bias. However, pronouns produced in reference to the character may not necessarily reflect underlying inferences about the character’s gender. In particular, some he responses in the Last Name condition may have been driven by a generic masculine usage, where participants produced he but did not necessarily infer the character as male.

Figure 2: Experiment 2: Procedure and example stimuli.

If the results of Experiment 1, where participants were less likely to produce she than predicted by the gender association of the first names, do reflect a bias to infer characters as male, the pattern of results should be the same when participants are asked directly about the characters’ genders. Characters will be more likely to be recalled as female as the rating of the first name becomes more feminine (First and Full Name conditions). However, a bias to infer characters as male will be present in all three conditions and strongest when direct probabilistic information about gender is not provided (Last Name condition).

Alternatively, the bias to use the pronoun he when describing the characters in Experiment 1, particularly in the Last Name condition, may not have reflected a bias in underlying gender inferences about the character, and instead reflected the use of the generic masculine (Bodine, 1975). If so, in Experiment 2, we would expect to observe no bias to recall characters as male in the First and Full Name conditions, where probabilistic cues to gender are available. In the Last Name condition, we would expect characters to be more likely to be recalled as male than as female, due to the tendency to refer to men using last names more often than women.

3.3 Results

Responses were coded as male (e.g., “m,” “man,” “male”), female (e.g., “f,” “woman,” “female”), or other (e.g., “It wasn’t specified,” “I don’t remember”). As in Experiment 1, the rates of male and female responses were roughly equal in the First and Full Name conditions, following the balanced distribution of the first names, but participants overwhelmingly responded male in the Last Name condition (Table 4). Overall, participants were less likely to respond female than male or other (β = –0.86, z = –5.71, p < .001). Participants in the First and Full Name conditions were more likely to respond female than participants in the Last Name condition (β = 2.00, z = 5.83, p < .001). There was no difference between the First and Full Name conditions.

Next, the effect of Gender Rating was analyzed for the First and Full Name conditions (Table 6), again following the same model specifications as in Experiment 1. The intercept term was significant (β = –0.18, z = –2.99, p < .01), indicating that participants were less likely to respond female in the First and Full Name conditions overall. Female responses became more likely as the names became more feminine (β = 0.78, z = 22.34, p < .001), but did not surpass male responses until the first name in the prompt was biased somewhat feminine, rather than at the mean (Figure 3). The interaction between Gender Rating and Condition was not significant, indicating that the linear effect of Gender Rating was similar in the First and Full Name conditions.

Figure 3: Experiment 2: Proportions of male (blue), female (red), and other (gray) responses in the First Name and Full Name conditions by the mean-centered gender rating of the first name. Points indicate means for each of the 21 first names, and lines indicate a smooth function on the raw data. Here, 0 is the mean of the 21 names, and the dashed line indicates the center of the original response scale in the norming study.

Next, we conducted the same three exploratory analyses as in Experiment 1: exclusion of the other responses, a quadratic effect of Gender Rating, and participant gender effects (Supplement §3.3–3.5). The rate of other responses (3.72%) was lower than in Experiment 1, and excluding other responses did not affect the substantive pattern of results. When adding a quadratic effect of Gender Rating to the Condition and Gender Rating model, neither the quadratic effect nor its interaction with Condition were significant. This finding is inconsistent with the hypothesis that the Condition effect would be larger at the midpoint of Gender Rating compared to the endpoints. Adding participant gender as a covariate revealed a significant interaction between Participant Gender and the Last vs. First + Full contrast (β = –.42, z = –2.93, p < .01), such that male participants were less likely than non-male participants to respond female in the First and Full Name conditions (β = –.26, z = –3.75, p < .001), whereas there was no effect of participant gender in the Last Name condition (β = .15, z = 1.23, p = .22). Across conditions, the effect of Gender Rating was smaller for male participants than for non-male participants (β = –0.16, z = –2.64, p < .01).

3.4 Discussion

Experiment 2 was designed to address the possibility that the preference for he responses in Experiment 1, especially in the Last Name condition, was due to participants using generic masculine forms, rather than a bias in the underlying gender inferences. Our findings were inconsistent with this as the primary explanation of the data. As in Experiment 1, participants’ judgments about the genders of characters introduced in short narratives exhibited a bias to infer characters as male. This bias was strongest in the Last Name condition, compared to the conditions where the character was introduced including a first name (First and Full Name conditions). Participants did not recall the character as female 50% of the time at the midpoint on the name gender rating continuum, but instead when the names were somewhat feminine.

The bias to infer characters as male was present in Experiment 2, but smaller than in Experiment 1 (see comparison between experiments in Figure 8). This difference is primarily due to an attenuated difference between the Last Name and First + Full Name conditions, where participants were 16.84 times more likely to produce a she response in the First + Full Name conditions in Experiment 1, but 7.39 times more likely to respond female in the First + Full Name conditions in Experiment 2. The mismatch between knowledge about the gender associations of first names and inferences about the genders of characters with those names – where characters only began being preferentially inferred as female when first names were somewhat feminine, not at the midpoint – was consistent across the two experiments. In the First and Full Name conditions, the odds ratios were 0.70 for a she response compared to a he or other response and 0.77 for a female response compared to a male or other response (Figure 8).

One reason that the bias to infer characters as male was smaller in Experiment 2, aside from residual uses of generic he in the Last Name condition, is that people may be less likely to assume characters are male by default when the task requires them to think more directly about gender. This would be consistent with prior results, where after writing about a generic person, participants were two and a half times more likely to use masculine names to describe the character, as compared to two times more likely to explicitly label the character as male (Hamilton, 1988). Before considering this further, we first explore whether the people=male bias can be attenuated when people are provided with more information about, and repeated reference to, a character.

4.2 Predictions

The results of Experiments 1 and 2 support a model where gender is inferred based on a combination of probabilistic cues to gender – here in the form of a person’s first name – and an overall people=male bias. However, the characters were only mentioned once in Experiment 1 and twice in Experiment 2, and the Last Name condition contained no cues about the characters’ genders through the names themselves, only more indirect cues from the fact that men are more likely to be referred to by last name. The bias to infer people as male may only be present in the initial inferences established during these brief introductions. If so, additional information about the referent and time to process gender inferences may attenuate or eliminate this tendency. We would then expect to find no bias towards he responses when the probabilistic information about gender carried by the first name is repeatedly presented (First and Full Name conditions).

Figure 4: Experiment 3: Procedure and example stimuli.

Alternatively, if biased gender inferences persist after repeatedly encountering cues to a character’s gender, we would expect to observe a bias to infer characters as male in all conditions, with he responses occurring more frequently than predicted by the gender distributions of the first names. One reason to expect the bias to persist comes from work showing that revising initial inferences about a character’s gender incurs processing costs while reading (Carreiras et al., 1996; Garnham et al., 2002; Kennison & Trofe, 2003; Sturt, 2003).

4.3 Results

Responses were categorized as he, she, and other. The sentence completion prompts were less constrained than Experiment 1, with only 53% of responses beginning with a pronoun (compared to 93% in Experiment 1). As a result, we analyzed pronouns used to refer to the character at any position in the response (69% of responses). Table 7 shows the proportions of responses across conditions, with other responses occurring in about a third of trials, an increase compared to Experiment 1. Responses were analyzed using logistic mixed-effect regression models, as before. Unlike Experiments 1 and 2, the contrasts for Condition were weighted to account for uneven numbers of participants in each condition. Recall that in all conditions, the first of 4 repetitions of the name was always a full name. As a result, we now analyze Gender Rating in all 3 conditions (Table 8).

Figure 5: Experiment 3: Proportions of he (blue), she (red), and other (gray) responses in the First, Last, and Full Name conditions by the mean-centered gender rating of the first name. Points indicate means for each of the 21 first names, and lines indicate a smooth function on the raw data. Here, 0 is the mean of the 21 names, and the dashed line indicates the center of the original response scale in the norming study.

Overall, participants were less likely to produce she responses than he and other responses (β = –1.53, z = –15.09, p < .001). While she responses increased as the first name became more feminine (β = 1.15, z = 19.02, p < .001), she responses only surpassed he responses when names were somewhat feminine, not at the mean (Figure 5). Neither main effect of Condition was significant. The interaction between Gender Rating and the Last vs. First + Full condition contrast was significant (β = 0.12, z = 2.15, p < .05), such that the effect of Gender Rating was larger in the First + Full Name conditions compared to the Last Name condition. This interaction is likely due to the fact that the First + Full conditions had 4 repetitions of the gendered first name, whereas the Last Name condition only had 1 use of the first name. The interaction between Gender Rating and the First vs. Full Name condition contrast was not significant.

We conducted the same three exploratory analyses as in Experiments 1 and 2, following the same model specifications and applying Bonferroni corrections for multiple comparisons. Responses coded as other were similar to the types in Experiment 1 (e.g., repeating the character’s name), but represented a larger proportion of the data (31%). When excluding other responses (Supplement §4.3), the Last vs. First + Full contrast was significant (β = 0.26, z = 2.63, p < .01), such that participants were less likely to produce she in the First and Full Name conditions than in the Last Name condition, similar to the results of the first two experiments. The intercept (β = –0.42, z = –3.42, p < .001) and the interaction between Condition and Gender Rating both remained significant (β = 0.42, z = 5.46, p < .001) in this subset of the data.

Adding a quadratic effect of Gender Rating to the primary model, which included other responses, revealed a significant quadratic effect of Gender Rating (β = –0.11, z = –3.67, p < .001). Inspection of the data suggests that this effect may be due to stronger effects of name rating towards the center of the gender rating scale (androgynous names) than at the end points (strongly-gendered names). The interaction between the quadratic effect of Gender Rating and the Last vs. First + Full contrast was significant (β = –0.10, z = –3.24, p < .001). Probing this interaction indicated that the quadratic effect of Gender Rating was significant in the First and Full Name conditions (β = –0.15, z = –4.28, p < .001), but not in the Last Name condition (β = –0.06, z = –1.67, p = .09), which likely reflects that the Last Name condition only included the gendered first name in 1 out of the 4 repetitions. This analysis also indicated a significant Condition effect for the Last vs. First + Full contrast (β = 0.24, z = 3.00, p < .01), such that participants were more likely to produce she in the First and Full Name conditions compared to the Last Name condition (Supplement §4.4).

Adding participant gender as a covariate revealed that male participants were less likely than non-male participants to produce she responses as compared to he and other responses across all three conditions (β = –.33, z = –3.53, p < .001). Participant Gender did not significantly interact with Condition or Gender Rating (Supplement §4.5). Finally, we conducted exploratory analyses of the Accomplishment, Likeability, and Importance ratings (Supplement §4.6). Because these ratings were near ceiling at the positive ends of the scales, the results were largely nonsignificant, with the exception that more likeable characters were more likely to be referred to with she.

4.4 Discussion

Experiment 3 examined whether the bias to produce he persists with more information about the character and more time to process an inference about their gender. Participants read paragraph-length news stories that mentioned a character four times and described their noteworthy accomplishment. All characters were introduced by their full name, and the following three references carried varying gender cues. In the First and Full Name conditions, the first name was repeated; in the Last Name condition, the first name was not repeated, but the choice of form of reference is an indirect cue to gender, given that men are more likely to be referred to by last name. Despite the fact that all characters were first introduced with their full name, the bias towards producing he persisted. She responses increased as the gender rating of the first names became more feminine, but only overtook he responses when the names were somewhat feminine, not at the midpoint. The effect of gender rating was stronger in the First and Full Name conditions, where the gendered first name was repeated four times, than in the Last Name condition, where the first name only appeared once.

Across conditions, the odds ratio of a she response (vs. a he or other response) was 0.24 in Experiment 1 and 0.22 in Experiment 3 (Figure 8). Note, however, that there was a higher rate of other responses in Experiment 3. It is unclear how much this reflects a greater flexibility in the sentence completion prompts, with the stories in Experiment 3 allowing more felicitous continuations not using a third-person pronoun to refer to the character than the sentences in Experiment 1. Excluding other responses, the odds ratio of a she response was 0.32 in Experiment 1 and 0.65 in Experiment 3. This suggests that the he response bias was attenuated in comparison to Experiment 1, but still present. This difference across studies was most notable in the Last Name condition: In Experiment 1, where the Last Name condition did not contain direct cues to gender, the odds ratio of a she response (vs. a he response, other responses excluded) was 0.04. In Experiment 3, where the first mention was by full name, the odds ratio of a she response (vs. a he response, other excluded) was 0.51. Thus, providing the comprehender with probabilistic information about a character’s gender may attenuate, but cannot completely override, the bias to infer characters as male instilled by reference by a last name.

5.2 Predictions

Thus far, our findings support a model of gender inference where probabilistic cues to gender are combined with a bias to infer that a character is male. In Experiment 4, we test the predictions of this model when the characters are presented in more detail and where the probabilistic cue to gender (in the form of the first name) is more strongly established, but when the sequence of the experiment does not prompt participants to be making inferences about gender when first reading the stories. While the results of Experiment 3 demonstrated a persistent bias to produce he, here we ask if the bias to infer characters to be male persists when directly asked about the character’s gender. If so, the bias towards recalling characters as male will be present even when probabilistic information about gender is repeatedly provided (First and Full Name conditions). Alternatively, if the bias to infer referents as male is attenuated after probabilistic cues about gender are well-established and when gender inferences are asked about directly, we would expect no bias towards recalling characters as male in the First and Full Name conditions.

5.3 Results

As in Experiment 2, responses were coded as male, female, or other (Table 9) and analyzed using logistic mixed-effect regression models predicting the log odds of a female response as opposed to male and other responses (Table 10). Characters were less likely to be recalled as female overall (β = –0.26, z = –3.14, p < .01), and somewhat more likely to be recalled as female in the First + Full Name conditions than in the Last Name condition (β = 0.13, z = 0.94, p < .05). The comparison between First and Full Name conditions was not significant. Participants responded female more frequently as the first names became more feminine (β = 0.76, z = 16.65, p < .001), but female responses did not overtake male responses until the first names were somewhat feminine (Figure 7). The interaction between Gender Rating and Condition was significant for the Last vs. First + Full contrast (β = 0.13, z = 3.81, p < .001), due to a larger effect of Gender Rating in the First + Full Name conditions, where the first name was repeated four times, as compared to the Last Name condition, where the first name was only presented once. An interaction between Gender Rating and the First vs. Full contrast (β = –0.10, z = –2.45, p < .05) was due to a larger effect of Gender Rating in the First Name condition than in the Full Name condition.

Finally, we conducted the same set of supplementary analyses as in prior experiments (Supplement §5.3–5.5). Excluding other responses (2.99% of total responses) revealed a similar pattern of results as the primary analysis. Adding a quadratic effect of Gender Rating revealed no new significant effects after Bonferroni correction for multiple comparisons. Adding Participant Gender as a covariate revealed that male participants were overall less likely than non-male participants to recall the character as female (β = –.20, z = –3.27, p < .001). As in Experiment 3, the Accomplishment, Likeability, and Importance ratings were near ceiling at the positive ends of the scales, and more likeable characters were more likely to be recalled as female. Additionally, interactions between each of the three character ratings and Gender Rating indicated that the effects of Likeability, Accomplishment, and Importance on gender inferences were stronger with more feminine names (Supplement §5.6).

Figure 7: Experiment 4: Proportions of male (blue), female (red), and other (gray) responses in the First, Last, and Full Name conditions by the mean-centered gender rating of the first name. Points indicate means for each of the 21 first names, and lines indicate a smooth function on the raw data. Here, 0 is the mean of the 21 names, and the dashed line indicates the center of the original response scale in the norming study.

5.4 Discussion

Experiment 4 examined whether the bias to explicitly recall characters as male persists when participants see additional information about, and repeated reference to, the character in a narrative, and when they have more time to develop inferences about the character but are less prompted to do so intentionally by the structure of the experiment. Although all characters were first introduced with their full name, which provided probabilistic information about their gender, participants were overall less likely to infer the character as female than to infer them as male or not indicate a gender inference. Characters needed to have first names that were rated somewhat feminine before they were more likely to be recalled as female, while characters with androgynous first names were more likely to recalled as male. The effect of the first name’s gender rating was strongest in the First Name condition, where the first name appeared in all 4 references to the character, and weakest in the Last Name condition, where the first name only appeared once. The bias to infer characters as male was smaller in Experiment 4 as compared to Experiment 2, but still not eliminated: the odds ratio of a female (vs. male or other response) across conditions was 0.42 in Experiment 2, and 0.77 in Experiment 4 (Figure 8).

6.1 Overview of findings

The present studies investigated how choices in how we refer to a person affect inferences about that person’s gender. Specifically, we examined how referring to a character by first name, last name, or full name impacted two measures of the readers’ gender inferences: pronoun use in a sentence completion task and responses to an explicit question about gender. We considered two competing hypotheses about the gender inference process. One hypothesis was that people only show a bias to assume referents are male when few cues about gender are available (e.g., you only know the person’s last name). Alternatively, we hypothesized that gender inferences might be shaped by a combination of the people=male bias (Silveira, 1980), along with other probabilistic cues to gender.

Across four experiments, using both short and long character introductions and two measures of gender inference, we observed that inferences about gender were shaped by a persistent people=male bias, along with clear use of probabilistic cues to gender. The results of Experiment 1 showed that characters who were referred to by last name only were overwhelmingly referred to with he. While providing more direct cues to gender through a first name attenuated this bias, it did not eliminate it. Instead, a character’s first name had to be at least somewhat feminine before she responses became more common than he responses. The increases in she and he responses were asymmetric, with androgynous names that leaned feminine still eliciting he responses, but androgynous names that leaned masculine eliciting responses that did not use a pronoun instead of she responses. An alternative interpretation of these findings, however, is that bias to use he, particularly in the Last Name condition, was due to participants using the generic masculine, and not due to biased inferences about gender. To address this possibility, Experiment 2 asked participants about their gender inferences directly. The results of Experiment 2 revealed a persistent, if smaller, overall bias to recall the characters as male. In addition, when cues to gender were provided through the use of the character’s first name, the name had to be somewhat feminine before the character was preferentially recalled as female. This mismatch between knowledge about the gender distributions of the first names and inferences about the characters from those names remained similar between Experiments 1 and 2.

In the first two experiments, participants read sentence-length descriptions of the characters. One question, then, is whether the observed bias in gender inferences would persist when a character is referred to multiple times and more is known about them. If the people=male bias attenuates as more information about the person has accrued and more time has been spent thinking about them, it may be attenuated in longer texts. To address this question, Experiments 3 & 4 examined gender inferences in paragraph-length news stories, where characters were introduced by full name and subsequently referred to three more times (manipulated by condition). In comparison to Experiments 1 & 2, these stories provided repeated cues to the character’s gender and more time for the reader to process inferences about the character, as well as more closely resembling a way we might read about a new person in everyday life. While the people=male bias was numerically smaller in Experiments 3 & 4 compared to Experiments 1 & 2, characters were still less likely to be referred to with she and less likely to be recalled as female than the distributions of the first names would predict. The strongest biases were observed when no direct information about gender was provided (Last Name condition in Experiments 1 & 2), where about 80% of characters were referred to with he or recalled as male. It is worth noting the magnitude of the people=male bias is roughly the same as results from 20–30 years ago (Davis Merritt & Kok, 1995; Hamilton, 1988), despite continuing social advances in gender equality.

Figure 8: Comparing Experiments 1–4: Odds ratios of she vs. he + other and female vs. male + other responses averaged across all condititions, in the Last Name condition only, in the First + Full Name conditions only, and in the Last Name compared to the First + Full Name conditions. Values less than 1 indicate being less likely to produce a she/female response; values greater than 1 indicate being more likely to produce a she/female response, with a discontinuous X axis to include the 2 largest values. Odds ratios correspond to the exponentiated beta estimates reported in the models.

6.2 Choosing referential forms

A primary focus of research on pronoun production in English has been on when speakers use pronouns instead of other referential expressions. A common assumption is that there is a causal link between a referent’s status in the discourse (e.g., whether it was the sentence topic, how often it was mentioned, and in what syntactic position, among others) and the form of reference the speaker selects. As a referent becomes more focused, salient, or prominent, the forms of reference used generally become more reduced, and pronouns are more likely (e.g., Gundel et al., 2012; Rodhe & Kehler, 2014; see Arnold & Zerkle, 2019, for discussion). For example, Schmitt et al.’s (1999) model of lexical access in pronoun production, tested in German, assumes that if a lexical concept is activated and sufficiently “in focus” in the discourse, the speaker will produce a pronoun instead of a full noun phrase. In this model, activating the lexical concept also activates the corresponding grammatical gender node (masculine, feminine, neuter), and if the speaker uses a pronoun, the gender node is selected in order to produce the correct pronoun (Jescheniak & Levelt, 1994; Levelt et al., 1999; Roelofs, 1992). Lexical access models generally agree that grammatical gender is represented as a separate lexical-syntactic feature in the mental lexicon (Wang & Schiller, 2019), but models differ in the structure of, and time course with which, this feature is connected to other linguistic representations (serial, unidirectional connections, e.g., Jescheniak & Levelt, 1994; Levelt et al., 1999; Roelofs, 1992; or bidirectional connections, e.g., Dell, 1986, 1988, 1999; Dell & O’Seaghdha, 1992). When producing gender-marked pronouns, as well as determiners, competition between different forms can arise from the grammatical gender features of other lexical concepts that are also activated (Schiller & Caramazza, 2003). Generally, models of grammatical gender selection do not include competition between multiple gender features activated by the same lexical concept. The closest analogue is languages where the singular and plural forms of determiners vary for the same grammatical gender, and one approach argues that the singular form is activated by default, and can interfere with activating the plural form (Jescheniak et al., 2014; Schriefers et al., 2002).

Complexities arise, however, when we consider that gendered language talking about people reflects a social construct negotiated between speakers, not a discrete grammatical or semantic feature (Ackerman, 2019; Conrod, 2020; McConnell-Ginet, 2014). In contrast to grammatical gender, information about social gender carried by names is probabilistic. We know from experience that most people named Mary are referred to with she, most people named Brian are referred to with he, and people named Jordan are commonly referred to with he or she. But without knowing more about a particular person, speakers may be unsure of which pronouns are appropriate. Additionally, there are many contexts in which multiple choices of pronouns are available, such as using singular they instead of he or she to leave a referent’s gender unspecified (e.g., My friend_i sent me a picture of their_i cat) and using singular they instead of he or she for people who use they/them pronouns. This means that models of pronoun production that include reference to people need to account for speakers’ decisions about which pronouns to produce, in addition to decisions about when to produce pronouns.

Our findings offer insights into the mechanisms guiding inferences about gender and the processes by which people choose pronouns to refer to a person. We propose that the ways in which we refer to people are influenced by multiple factors, including speaker knowledge of gender distributions, speaker inference about a referent’s gender, and speaker pronoun choice, in turn influencing the comprehender’s inference about referent gender (Figure 9). To explain our findings, we discuss potential locations for bias to infer referents as male in this process. In particular, we focus on contexts like those in our experimental stimuli, where speakers have cues about gender in the form of names, but need to make an inference about which pronouns, if any, are appropriate to produce.

Figure 9: Factors influencing personal pronoun choice.

First, we assume that, based on world experience, speakers form and store estimates of gender distributions across contexts, including, at the most basic level, the knowledge that close to half of people are women or girls ([A] in Figure 9). Speakers also have information about the proportions of women in specific contexts: Estimates about the relative rates of women in various jobs showed a strong correlation with employment data, and when actual and estimated data diverged, participants were more likely to overestimate the proportion of men in a given occupation than to overestimate the proportion of women (Garnham et al., 2015; Misersky et al., 2014). Similarly, our norming study found a strong positive correlation between how feminine a first name was rated and the proportion at which it was given to children assigned female at birth. These findings indicate that people have reasonably well-calibrated knowledge of gender distributions in occupational contexts and based on first names, and that biases, when present, more frequently underestimate the proportion of women.

Speakers also have knowledge about a particular referent [B], and in many contexts, this includes information about what names, pronouns, titles, and other forms of reference are appropriate for them. In the present experiments, we focus on contexts where relatively little information is provided about the referent, requiring speakers to make inferences about their gender [C] and what language to use to refer to them [D]. Context-specific knowledge about gender distributions [A] contributes to the speaker’s inferences about a referent’s gender [C]. Previous findings show that this process [red] underestimates the prevalence of women, since people are biased to infer gender-unspecified referents as male (Davis Merritt & Kok, 1995; Davis Merritt & Wells Harrison, 2006; Silveira, 1980). In the present research, when no direct cues to gender were provided and participants were asked about their gender inferences (Last Name condition in Experiment 2), approximately 80% of responses inferred the referent to be male. This bias persisted, albeit attenuated, when some gender information was given (First and Full Name conditions) and after repeated reference (Experiments 3 & 4).

The speaker’s knowledge of gender distributions in general [A] is one contributor to their choice of what referring language to use [D], though speakers tend to use feminine language forms less frequently than their probabilistic knowledge about gender distributions would predict [blue] (Boyce et al., 2019; Hamilton, 1988; von der Malsburg et al., 2020). In the present research, participants were less likely to use she to refer to characters than the gender distribution of the first names would predict, after only one reference to the character and after repeated reference. One explanation is that the criterion for inferring referents as male is lower than the criterion for inferring them as other genders. This is one implication of the people=male hypothesis (Silveira, 1980): if the generic person is a man, then producing he (the unmarked category) might require a lower threshold of evidence than producing she (the marked category). Since pronouns are typically reserved to refer to the most salient, accessible, or in focus character (Arnold & Zerkle, 2019), a related possibility is that characters inferred as male are seen as more salient, and thus more likely to be referenced using a pronoun. This may explain why, in the First and Full Name conditions, he responses were dominant across the masculine half of the scale, whereas she and other responses (which typically did not use a pronoun) were both common in the feminine half of the scale.

Another factor in the choice of referring language [D] is the speaker’s inference about that specific referent’s gender [C]. A disconnect between the two is another potential source of bias [purple]: Recall that while around half of participants believed Hillary Clinton would win the 2016 US election, only 10% of responses used she to refer to the next president (von der Malsburg et al., 2020). When asked for both explicit and implicit measures about a character’s gender, participants described generic referents as female at higher rates than they chose feminine names to refer to them (Hamilton, 1988).

These findings suggest that inferences about referent gender and choices about gendered language are distinct processes. Mappings between gender inferences and language choice also vary by dialect, such as uses of generic he and singular they. Moreover, these mappings are not always symmetric. This was clearest in Experiment 1, where participants still used he when the character had a feminine-leaning androgynous name, but were equally likely to use she or no pronouns for characters with masculine-leaning androgynous names. This suggests that a speaker’s inference about a referent’s gender may need to be more certain to prompt the use of she than to prompt the use of he.

It is important to note that the discrete choices involved in gendered language production do not preclude the underlying inference about a referent’s gender being probabilistic. When participants in Experiments 1 & 3 used he or she, this did not necessarily reflect certainty about a gender inference. This response pattern may be more common for speakers of dialects where forms that directly encode uncertainty, such as using singular they for a referent with an unknown or unspecified gender, are not available. Future work could explore how the same language produced may reflect underlying levels of confidence. The experiments here cannot distinguish between the contributions of distributional knowledge [blue] and inferences about a specific referent’s gender [purple], only conclude that the resulting choices show a bias against using feminine language forms.

Comprehenders use speakers’ gendered language choices [D], as well as their knowledge of gender distributions [A], to form their own inference about a new referent’s gender [E]. One possibility is that comprehenders know that speaker pronoun choice is biased towards he and correct for this bias [weighting yellow over orange]. Although the experiments here do not address this question, Boyce et al. (2019) suggest that this is not the case. Participants read stories that included two repetitions of a role noun and one gendered pronoun. When asked to recall the character’s gender, participants did not correct for masculine bias in pronoun use, and instead continued to recall the referents as feminine at lower rates than the normed gender distribution of the role nouns.

Another aspect at play here is comprehenders’ knowledge of who and what speakers discuss. While the present experiments have focused on the probabilistic information about gender carried by names, the contexts in which a person is mentioned can also provide gender cues. This is particularly relevant in Experiments 3 & 4, where the stimuli were news stories highlighting a person’s accomplishment, instead of sentences describing a person performing everyday activities. While a comprehender may know that people named Jordan are about equally likely to be male or female, they may also know that a person mentioned for their recent career accomplishment is more likely to be male. Several of the stories described more stereotypically-feminine accomplishments (i.e., charity work), but most were stereotypically more masculine (i.e., politics, sports). The gender stereotypes of the stories were counter-balanced within the experiment by having each story paired with masculine, feminine, and androgynous first names across lists. However, the present experiments did not attempt to measure or experimentally manipulate the fact that people may be making additional inferences about gender based on the fact that the character accomplished something and that their accomplishment was judged as newsworthy. The results here show a bias to infer people are male and to use masculine language forms when making inferences about a character in a brief story, leaving open questions about how these biases interact with speakers’ original choices of who to discuss.

Finally, it is likely that inferences about gender in comprehension [E] influence underlying beliefs about gender distributions [A]. As such, speaker’s choices about how to refer to entities in the world arguably drive patterns in language comprehension (MacDonald, 2013). Thus, if speakers consistently underuse feminine forms of reference and comprehenders do not correct for this bias, beliefs about gender distributions in general and in specific contexts may then become biased to underestimate women [green].

6.3 Implications for talk about women

These results have potential implications for how we talk about women. When we refer to people, we choose between different combinations of forms, including pronouns, first names, last names, gendered titles (Mr., Mrs., Ms.), and nominally ungendered titles (Doctor, Professor). If certain forms of reference make feminine referents less likely to be inferred as feminine, should this influence which forms we choose? On the one hand, prior work suggests that people who are referred to with masculine-coded terms are judged more competent and successful (Atir & Ferguson, 2018; Rubin, 1981; Stewart et al., 2003; Takiff et al., 2001). Given these observations, a strategic speaker or writer could refer to a woman using masculine-coded forms to encourage a more masculine interpretation of the referent. This could mean reaping potential advantages (e.g., in perceived “eminence”), but potentially at the cost of having someone’s femininity be diminished or unacknowledged, and of perpetuating language production patterns that in turn may shape biases in comprehension. Alternatively, it may be preferable to work to change the underlying tendency to underestimate the presence of women, by choosing language forms that make it more difficult to assume a person is a man by default, especially in contexts where women are less visible. An open question, then, is if using gendered language to refer to women in contexts where their presence is systematically underestimated (e.g., doctors) would change perceptions about the contributions of women to that sphere of life.

6.4 Conclusion

Across a series of four experiments, we asked if alternative forms of reference to individuals guide inferences about the gender of a referent introduced in a sentence or brief story. Using measures of personal pronoun choice (Experiments 1 & 3) and explicit queries about gender (Experiments 2 & 4), a persistent bias to assume that the referent was male was observed across all four studies. This bias was strongest when the character was introduced by last name alone, but persisted even when the character was referred to by their first name. We argue that the observed male bias in gender inference likely results from multiple processes, including biases in knowledge of gender distributions, inferences about a referent’s gender, and pronoun choice.