Intersectional Male-Centric and White-Centric Biases in Collective Concepts

Abstract

In principle, the fundamental concepts person, woman, and man should apply equally to people of different genders and races/ethnicities. In reality, these concepts might prioritize certain groups over others. Based on interdisciplinary theories of androcentrism, we hypothesized that (a) person is more associated with men than women (person = man) and (b) woman is more associated with women than man is with men (i.e., women are more gendered: gender = woman). We applied natural language processing tools (specifically, word embeddings) to the linguistic output of millions of individuals (specifically, the Common Crawl corpus). We found the hypothesized person = man / gender = woman bias. This bias was stronger about Hispanic and White (vs. Asian) women and men. We also uncovered parallel biases favoring White individuals in the concepts person, woman, and man. Western society prioritizes men and White individuals as people and “others” women as people with gender, with implications for equity across policy- and decision-making contexts.

Keywords

androcentrism ethnocentrism intersectionality language word embeddings

Recent advances in natural language processing (NLP) have enabled researchers to use the linguistic output of millions of individuals to characterize human concepts at a massive, previously inconceivable scale. Here, we used NLP tools (specifically, static word embeddings trained on the Common Crawl corpus) to investigate collective understanding of three fundamental concepts: person, woman, and man.¹ These three concepts organize much of social life and scaffold legal and policy decision-making in modern societies. For instance, the “reasonable person” legal standard is central to multiple areas of U.S. law, including personal injury, negligence, and criminal law (Miller & Perry, 2012). This standard assigns culpability to a defendant based on the degree to which their behavior departs from what a “reasonable person” would have done in the same circumstances. More broadly, evaluations of the morality of an action are often argued to be based on principles that no “reasonable person” could reject (Scanlon, 2000). If the collective concept person is male- and White-biased, then legal and moral evaluations might likewise favor White men (Hegarty et al., 2013). In the present research, we investigated whether person, woman, and man prioritize certain groups over others.

The concept person should apply equally to people of all genders, but interdisciplinary theories of androcentrism argue that in reality society treats (a) men as generic people and (b) women as people with gender (Bem, 1993; de Beauvoir, 1949/2010). Here, we investigated both of these claims, as well as—crucially—whether the magnitude of this androcentric bias in collective concepts varies about people of different races/ethnicities. The latter is an important contribution because previous research on androcentrism, and research more generally (Field et al., 2021; Roberts & Mortenson, 2023), has tended to focus on White people, leaving open the possibility that androcentrism might be more or less pronounced about people other racial/ethnic groups.

A secondary goal of the present research was to provide a preliminary exploration of White-centric biases in collective concepts. In principle, person, woman, and man should apply equally to people of all races/ethnicities, but in reality, these collective concepts might prioritize White people (i.e., person = white, woman = white, and man = white), especially in U.S. and other Western contexts (Hegarty, 2017; Morehouse et al., 2023). Whether collective concepts exhibit such a White-centric bias is unknown.

Investigating Collective Concepts Using NLP Tools

According to classic arguments in linguistics, aspects of the meaning and function of a word—and the content of the concept denoted—can be captured by observing the word’s linguistic context (Harris, 1954; Lenci, 2018). This is known as the distributional semantics hypothesis. For instance, in the sentence, “Each morning, Joe boiled water in the balak for tea,” a naive reader might guess that balak means something similar to kettle, because the other words alongside balak also frequently co-occur with kettle (McDonald & Ramscar, 2001). Inspired by this idea, vector-based language models distill information about a word’s co-occurrences with other words in a corpus into a word embedding—a multidimensional vector that carries information about the word’s context, and thus its meaning (Mikolov et al., 2013, 2018; Pennington et al., 2014). These NLP tools thus create vector-based representations of word meaning based on information about a word’s linguistic contexts (syntactic and semantic) in a language corpus.

Because the vector embedding for a word is constructed based on information about which other words it co-occurs with (i.e., its typical linguistic context), the similarity between any two words’ embeddings captures the similarity between the two words’ linguistic contexts—and thus, according to the distributional semantics hypothesis, the similarity in their meanings. Notably, two words can have similar word embeddings even if they rarely co-occur with each other. For instance, consider the words scientist and researcher. These two words might rarely co-occur in the span of a single sentence (because they are redundant with each other, so speakers tend to use one or the other), but they nevertheless have similar word embeddings (similarity score = 0.77) because they often occur in similar contexts (e.g., alongside words such as lab, data, and rigorous; Mikolov et al., 2018). Word embeddings can thus capture complex semantic relationships with high precision—for example, that brother is to sister as grandson is to granddaughter (Mikolov et al., 2013).

When word embeddings are trained on large-scale corpora of text written by millions of individuals, they begin to capture collective concepts (see also collective representations; Durkheim, 1915; Moscovici, 1963). Collective concepts reflect consensus, widely shared views in a community (Momennejad, 2022). These ideas are “in the air” and familiar to people in a community regardless of whether they personally endorse them. The fact that collective concepts can persist regardless of any individual’s personal endorsement makes them resistant to change, and thus particularly important to understand.

Recently, researchers have begun to use word embeddings to investigate social biases in collective concepts (Caliskan et al., 2017; Charlesworth et al., 2021; Goodhew et al., 2022; Lewis & Lupyan, 2020). Some of this NLP work draws on the now classic implicit association test (IAT); for instance, IAT research consistently finds that science and math concepts are more associated with men in individuals’ minds and arts and humanities concepts are more associated with women—that is, science = man / arts = woman (Greenwald et al., 1998; Kurdi et al., 2019; Nosek et al., 2007²). The recently developed Word-Embedding Association Test (WEAT) has suggested that these conclusions “scale up” to the level of collective concepts as well: Based on text written by millions of individuals, the WEAT revealed that words denoting science concepts appear in more similar linguistic contexts to words for men and words denoting arts concepts appear in more similar contexts to words for women than the reverse (Caliskan et al., 2017). Word embeddings are thus a powerful tool because they distill information from language use in a community into valuable insights regarding the collective concepts of that community, including any biases therein.

Theory and Prior Evidence on Androcentrism

So far, research using NLP tools has tended to focus on well-studied social biases, in part to establish the validity of this approach. Here, we investigated a less well-studied, but perhaps more fundamental, social bias: androcentrism. Interdisciplinary theories of androcentrism argue that societies are androcentric (or male-centric³), in that men are collectively understood to be a gender-neutral standard whereas women are regarded as a gendered exception to this norm (Bem, 1993; de Beauvoir, 1949/2010; Gilman, 1911; P. Martin & Papadelos, 2017). In terms of psychological processes, theories of androcentrism propose that men are (proto)typical exemplars of the generic, superordinate person concept, whereas women tend to be thought of at a subordinate level of categorization, in terms of their gender (Bailey et al., 2019). Androcentrism also marginalizes people of other gender identities because they disrupt status quo (binary) conceptualizations of gender and gender hierarchy (Dembroff, 2019; Morgenroth & Ryan, 2021).

Empirical studies using traditional social-psychological methods have found evidence for androcentrism, including a tendency to associate (a) men with generic person concepts (person = man) and (b) women with their gender more than men with theirs (gender = woman). For instance, participants were more likely to spontaneously describe men using generic person words (e.g., a person) and women using gender-specific words (e.g., a woman; Hamilton, 1991). Participants were also more likely to list men than women as examples of generic person concepts (Bailey & LaFrance, 2016; Gastil, 1990) and to classify women according to their gender faster than they classified men (Zarate & Smith, 1990). Finally, an IAT investigation found that participants associated (a) words for person with men’s names and (b) words for woman with women’s names more than the reverse (Bailey et al., 2020).

Androcentrism is less commonly studied compared with other forms of gender bias (e.g., science = man / arts = woman) and has been examined almost exclusively using traditional social-psychological methods in small-scale studies involving dozens to hundreds of participants (for a review, see Bailey et al., 2019). Beginning to fill this gap, Bailey et al. (2022) recently investigated androcentrism in large-scale text corpora using word embeddings. They found that words for generic person concepts, generic human attributes, and generic human actions were used in more similar linguistic contexts to words for men than to words for women. These results provided initial evidence of a person = man association in collective concepts.

The first goal of the present investigation was to extend this prior work to assess both aspects of androcentrism at a broad, societal-level scale of analysis: that is, both (a) the tendency to associate person with men more than women (person = man, as in Bailey et al., 2022) and (b) the tendency to associate women with the concept woman more than men with the concept man (gender = woman, new to this work). Investigating both aspects of androcentrism is a key contribution because these aspects are theorized to be distinct and mutually reinforcing: Part of what maintains the bias to associate person with men more than women is the “othering” that results from associating women with their gender identity more than men with theirs (i.e., regarding women as “the sex” or “the gender”; de Beauvoir, 1949/2010). Thus, in the present research, we extended prior work to investigate both theorized aspects of androcentric bias in collective concepts while also—crucially—investigating possible intersections with race/ethnicity, addressed next.

Intersectional Approaches to Investigating Gender and Racial/Ethnic Biases

A chief goal of the present work was to investigate whether and how androcentrism intersects with race/ethnicity. Biases and beliefs about gender and about race/ethnicity tend to be investigated separately, an approach that can isolate systematic differences between beliefs about, say, women in general and men in general. However, intersectionality theories highlight the drawbacks inherent in investigating each identity dimension in isolation. These theories note that individuals’ gender, their race/ethnicity, and their other social identities cannot be separated but operate together to co-create social meaning (Bowleg, 2008; Cole, 2009; Crenshaw, 1990; Goff & Kahn, 2013; Lorde, 1984). Thus, each individual is an amalgamation of multiple overlapping identities that mutually influence each other. For instance, what it means to be a woman differs for a queer White woman from the Midwest in her early 40s, etc., compared with a straight Hispanic woman in Los Angeles in her 70s, etc. (Lorde, 1984). Although it is not feasible to investigate all possible overlapping identities, gender and race/ethnicity are two primary dimensions of identity (Fiske & Neuberg, 1990; Lei et al., 2023). Gender and race/ethnicity are perceived within a few hundred milliseconds of seeing a face (Amodio et al., 2014), and they automatically elicit basic reactions such as positive/negative attitudes and attributions of humanity (this is true especially for gender; Connor et al., 2022; A. Martin & Mason, 2022). It is thus particularly important to investigate possible intersections between androcentric gender bias and biases based on race/ethnicity.

It is possible that androcentrism varies in magnitude across collective concepts of different racial/ethnic groups (Eagly & Kite, 1987). Much of the research on androcentrism—and psychological research more generally—tends to directly or indirectly investigate representations of White people more than people of other races/ethnicities (Cundiff, 2012; McGorray et al., 2023; Roberts & Mortenson, 2023). For instance, the androcentrism IAT described previously used names for women and men in general (e.g., John; Bailey et al., 2020), but there is reason to suspect that these names were more likely to be perceived as belonging to White people than to people of other races/ethnicities (Cotton et al., 2014; Devos & Banaji, 2005; Hegarty, 2017). This raises the possibility that much of what is known about androcentrism is limited to beliefs about White people specifically. Perhaps it is not the case that the collective concept person is associated with men more than women but instead that person is associated with White men specifically more than White women. Thus, it is important to test whether androcentric gender bias prioritizes men across different races/ethnicities or White men per se.

Theory and research on intersectional gender and racial/ethnic biases provide insights into how androcentrism might differ in collective concepts of different races/ethnicities. One prominent theory, intersectional invisibility (Purdie-Vaughns & Eibach, 2008), argues that the tendencies to regard men and White people as more prototypical combine such that women of color are especially overlooked or “invisible” (Coles & Pasek, 2020; Lei et al., 2023; Phills et al., 2018; Sesko & Biernat, 2010). Consistent with this claim, when asked to select three examples of the concept person from a set of images, participants were most likely to select a White man, a Black man, and a White woman (Bailey & LaFrance, 2017). This result suggests that men and White people are more central to the person concept, at least as investigated with traditional social-psychological methods.

Although research on intersectionality initially focused on a Black-White race contrast,⁴ newer works on this topic (Lei et al., 2023; Zou & Cheryan, 2017) encompass other groups and often make different predictions about perceptions of Asian people by incorporating insights from gendered race theory. Gendered race theory argues that traits associated with the Asian group overlap more with stereotypes about women (resulting in “feminization”), whereas traits associated with the Black and Hispanic groups overlap more with stereotypes about men (resulting in “masculinization”; Galinsky et al., 2013; Johnson et al., 2012). For instance, U.S. participants imagine a Black person and a White person as a man more often than a woman but imagine an Asian person as a man or a woman at chance (Schug et al., 2015). Similarly, traits associated with Americans of different races/ethnicities overlap more with traits associated with the men of those groups for Black, Latino, Middle Eastern, and White Americans but not for Asian Americans (Ghavami & Peplau, 2013).

Extrapolating insights from intersectionality theories and research to the present work, we thus might expect to find person = man biases about a range of racial/ethnic groups, with the possible exception of Asian people. Because the Asian group is relatively “feminized” from a U.S. and Western perspective (Johnson et al., 2012), Asian men may be marginalized instead of women, causing person = man biases about Asian people to be less pronounced or to reverse. Conversely, person = man biases might be heightened about Black and Hispanic people. We might also expect to find stronger gender = woman biases about the Asian group because Asian women are hyper-feminized. In contrast, we might find weaker gender = woman biases about the Black and Hispanic groups because it is the men in these groups that appear to be more associated with their gender and hyper-masculinized. In summary, prior theory and research on intersectionality highlights the importance of incorporating race/ethnicity into investigations of androcentrism and licenses several predictions about how androcentrism might differ across racial/ethnic groups.

White-Centric Bias

In addition to our primary goal of investigating androcentrism and its intersection with race/ethnicity, the present research provides a preliminary investigation of White-centric biases in collective concepts. In principle, the generic concepts of person, woman, and man apply equally to people of all races/ethnicities, but in reality, these concepts might prioritize White people over people of other groups, at least in U.S. and other Western societies. That is, we might observe person = white, woman = white, and man = white biases in collective concepts.

Research using traditional social-psychological methods suggests that people are more likely to think of a generic person as being White than other races/ethnicities. For instance, as described previously, when U.S. participants were asked to select images to represent humanity, they selected more images of White than Black people (Bailey & LaFrance, 2017; for a non-U.S. investigation, see Sibley & Liu, 2007). Participants also associate the idea of an American and other generic human concepts (celebrity) with White people more than with people of other races/ethnicities (Danbold & Huo, 2015; Devos & Banaji, 2005; Hegarty, 2017). Similarly, U.S. participants appear to perceive a human face as White by default (Uddenberg & Scholl, 2018). Finally, a recent large-scale IAT investigation found evidence for a human = white / animal = people of color association among White participants (Morehouse et al., 2023). Although our investigation of collective concepts is not U.S.-specific, we still might find that the concepts person, woman, and man share more overlap with representations of White people than people of other races/ethnicities.

The Present Research

The primary goal of the present research was to provide an investigation of androcentric bias in collective concepts (including both its person = man and gender = woman components) and the intersection of this bias with race/ethnicity. We first addressed this goal by adapting the word embedding association test (WEAT; Caliskan et al., 2017), which allowed us to test the two components of androcentric bias jointly (Study 1). We then isolated each component to investigate (a) whether the person = man association is present (Study 2) and, separately, (b) whether women are more associated with woman than men are with man (gender = woman; Study 3). Thus, in terms of contributions, we sought to replicate prior work on androcentrism in collective concepts by investigating the person = man component (Bailey et al., 2022) and then extend this work by investigating the gender = woman component (new to this work).

Within both of these approaches to examining androcentric bias—with its components considered both jointly and separately—we investigated whether the strength of androcentric bias differs across collective concepts of different racial/ethnic groups. This is a crucial contribution because prior work has often indirectly or directly investigated androcentrism about White people. We expected to find androcentrism across racial/ethnic groups, but we also expected this bias to be somewhat attenuated in collective concepts of Asian people. With respect to specific components, we expected to find (a) weaker person = man bias about Asian people, moderate levels about White people, and possibly more about Black and Hispanic people (Ghavami & Peplau, 2013; Johnson et al., 2012; Schug et al., 2015). Regarding the second component, we expected to find (b) stronger gender = woman bias about Asian people, moderate levels about White people, and possibly less about Black and Hispanic people (Johnson et al., 2012).

Third and finally, we provided an initial investigation of possible White-centric biases in collective concepts—testing whether the person (Study 2), woman (Study 3), and man (Study 3) concepts are more associated with White people, White women, and White men, respectively, compared with people of other races/ethnicities. Based on prior research using traditional social-psychological methods (Morehouse et al., 2023) with predominantly small-scale, U.S. participant samples (Bailey & LaFrance, 2017; Danbold & Huo, 2015; Devos & Banaji, 2005; Hegarty, 2017), we expected to find evidence for all three of these White-centric biases: person = white, woman = white, and man = white. Although our approach was not designed to provide a comprehensive investigation of these biases, we hope that it will pave the way for future studies of this important topic.

General Method

To investigate intersectional androcentric and White-centric biases in collective concepts, we compared three different sets of words: words denoting person, words denoting woman and man, and individuals’ first names (see Table 1 for examples and supplementary information for complete lists [SI.9]). For each word on these word lists, we extracted word embeddings (fastText, 300 dimensions; Mikolov et al., 2018) trained on the Common Crawl corpus [CC-MAIN-2017-22]. These word embeddings capture associations among words based on a large cross-section of text on the internet—2.96 billion web pages with over 630 billion words from government websites, news outlets, personal websites, blogs, etc. (Dodge et al., 2021; Mehmood et al., 2017). To measure similarity among concepts, we computed the overlap in the linguistic context of the relevant words by computing the cosine similarity of the corresponding word embeddings in vector space. Cosine similarity is a common measure of relative similarity between any two word embeddings.

Table 1.

Lists of person Concept Words, Words for woman and man Concepts, and First Names

Set of words		Examples	Counts
person words		human, humanity, person, they, them	30
woman words		woman, women, girl, she, her	38
man words		man, men, boy, he, him	36
First names	All names		4,061
	Asian woman names	Asha, Fan, Hoa, Jia, Priya	257
	Asian man names	Abdul, Dai, Jin, Qiang, Yang	364
	Hispanic woman names	Alma, Ana, Josefina, Leonor, Yasmin	190
	Hispanic man names	Alonso, Bernardo, Demetrio, Jose, Marco	233
	White woman names	Abby, Bianca, Clare, Ellen, Gloria	1,696
	White man names	Aaron, Dmitri, Edward, Lars, Philip	1,255

Note. Analyses that do not involve the race/ethnicity variables include the full list of 4,061 names (3,995 names have race/ethnicity designations). We also conducted supplementary analyses using continuous probabilities instead of discretizing names into gender and racial/ethnic groups; conclusions do not change (SI.8). Complete lists of words are reported in supplementary materials (SI.9).

The training corpus (Common Crawl) is a snapshot of the internet and thus includes multiple languages, but it overrepresents English-language text (e.g., in 2017, over 60% of the web pages in the corpus were exclusively in English; Mehmood et al., 2017). Because of this and because our lists of words were all in English and/or came from U.S. and Canadian sources (described next), it is likely that our investigation captures collective concepts from a predominantly English-language, U.S. and Canadian perspective. Next, we describe each of our three lists of words.

Words for person

To examine biases in the collective concept person, we used 30 words that denote this concept (e.g., person, individual; SI.9). These words have been used in prior work and have been previously validated by a set of coders as capturing the ostensibly inclusive concept person (Bailey et al., 2022).

Words for woman and man

To examine biases in collective gender concepts, we used 78 words: 38 words denoting woman (e.g., female, she) and 36 words denoting man (e.g., male, he; SI.9). These words have also been used in prior work and previously validated by a set of coders as capturing the gendered concepts woman and man (Bailey et al., 2022).

First Names

We used a large list of first names (Tzioumis, 2018; SI.9) because first names can simultaneously capture gender and race/ethnicity information, allowing us to test intersectional androcentric and White-centric biases. This list came from proprietary mortgage lender applications from the United States and Canada, which included over 2 million individuals’ self-identified race/ethnicity information. Although the proprietary data are not publicly available, the list of names has been processed and shared in prior work (Tzioumis, 2018).

Additional Processing and Validation Steps

For each name, we computed the probability that it refers to a woman or a man (Hannak et al., 2023) and to an Asian, Black, Hispanic, or White individual based on self-identifications (Tzioumis, 2018). We undertook additional processing and validating steps to ensure that the list of first names corresponded to the expected gender and race/ethnicity concepts (summarized in Figure 1 and described in full in SI.1 and SI.2). As part of this process, we removed the Black names from the main analyses because these names were not unambiguously associated with Black people more than White people in the training corpus. Thus, these names were not a valid measure of collective concepts of Black people. However, analyses with the Black names are reported in supplementary materials for completeness (SI.3).

Figure 1.

Summary of the Steps Undertaken to Process and Validate the First Names.

Final List of First Names

Our final list consisted of 4,061 names (see Table 1 for examples and SI.9 for complete lists). White names were overrepresented on this list (73%). This is likely because White people are the majority racial/ethnic group in the source countries (e.g., 62% in the United States and 70% in Canada; Jones et al., 2021; Statistics Canada, 2022), and because White people are disproportionately represented among those seeking mortgage applications (e.g., 82% of the people in the mortgage names database self-identified as White; see also Aladangady & Forde, 2021). However, because of the large size of the original list of names, there were also substantial numbers of names that had a high probability of belonging to other races/ethnicities. Notably, although our main analyses discretize each name into a single gender and racial/ethnic group, we also conducted supplementary analyses using continuous probabilities instead (SI.8); conclusions do not change.

A strength of our approach is that the probabilities that names were associated with a race/ethnicity or gender were calculated based on reports by actual people with those identities, rather than being based on stereotypic beliefs about which names “sound White,” etc. Also notable, other possible lists of names we could have used for this purpose were either much smaller (e.g., 5 names per group) and based on beliefs about stereotypical names for races/ethnicities (Caliskan et al., 2017; Greenwald et al., 1998), or they were lists of last names (Jones et al., 2021). Last names were not viable because they do not convey information about gender.

Transparency and Openness

Datafiles, analysis scripts, and the preregistration are available online https://osf.io/myfsg/?view_only=ed08b5c8671b41caa36e693974f2d342. We conducted preregistered replication studies using word embeddings trained on the same corpus (the Common Crawl) but via a different algorithm (GloVe, 300 dimensions; Pennington et al., 2014). GloVe has a different training architecture than fastText: It uses the global co-occurrence statistics of a word within a corpus to calculate its embedding, whereas in fastText a word’s embedding consists of the weights of a shallow neural network that was trained to encode information about the context of that word. Aside from using a different type of word embedding (GloVe vs. fastText), these replication studies were identical to main studies reported here. The findings from the replication studies are summarized throughout, and the results are reported in full in the supplementary materials (SI.5).

Study 1: Testing Androcentrism in a Word Embedding Association Test

To investigate androcentrism in collective concepts of different racial/ethnic groups, we conducted a WEAT (Caliskan et al., 2017). The WEAT was developed to be analogous to the influential IAT (Greenwald et al., 1998). The androcentrism WEAT we conducted here was modeled on the androcentrism IAT (Bailey et al., 2020) and consisted of two components: (a) the relative associations of the generic person concept with men’s first names versus women’s first names (person = man), and (b) the relative associations of the generic gender concepts woman and man with women’s first names versus men’s first names, respectively (gender = woman). The WEAT provides a single effect size estimate jointly quantifying the person = man / gender = woman association in collective concepts.

Analyses

We conducted the androcentrism WEAT across all 4,061 names and separately for each racial/ethnic group using formulas adapted from prior work (Caliskan et al., 2017). p values were obtained using permutation tests. Bias-corrected and accelerated bootstraped 95% confidence intervals (CIs) were calculated to facilitate comparisons among groups.

To compute the androcentrism WEAT, we first calculated the mean similarity of the person words to each first name (e.g., John). We then calculated the mean similarity of the matching gender concept words to each first name—for instance, the mean similarity of man words (male, etc.) to each of the first names identified with high probability as being a man’s first name (e.g., John). At this stage, we thus had two averages for each first name, an average similarity to person words and an average similarity to matching gender concept words. Next, we subtracted the average similarity to matching gender words (man or woman words, depending on the first name’s gender classification) from the average similarity to person words. At this stage, we thus had a difference score for each name that captured its similarity to person relative to its similarity to its matching gender concept, separately for high probability men’s first names and women’s first names. To compute the WEAT effect size d, we then (a) averaged these difference scores for women’s first names, (b) subtracted this average from the analogous average of difference scores for men’s first names, and then (c) divided this difference of averages by the standard deviation of all difference scores for individual first names.

Formally, let M represents our set of men’s first names and W represents our set of women’s first names. Let P and G represent our set of person concept words and corresponding gender concept words, respectively. Let cos $(\vec{n,} \vec{w})$ represents the cosine of the angle between the word embeddings for a given first name and a given person word or gender concept (man, woman) word. The WEAT effect size is,

d = \frac{m e a n_{m ϵ M} s (m, P, G) - m e a n_{w ϵ W} s (w, P, G)}{s t d_d e v_{n ϵ M \cup W} s (n, P, G)}

where for each man or woman first name (n),

s (n, P, G) = m e a n_{p ϵ P} c o s (\vec{n,} \vec{p}) - m e a n_{g ϵ G} c o s (\vec{n,} \vec{g})

Results

We found evidence for androcentric person = man / gender = woman associations across all names, d = 1.072, p < .001, 95% CI = [1.026, 1.114], and in collective concepts of each race/ethnicity group: Asian, d = 0.556, p < .001, 95% CI = [0.391, 0.701]; Hispanic, d = 1.285, p < .001, 95% CI = [1.157, 1.385]; and White, d = 1.121, p < .001, 95% CI = [1.072, 1.171]. Androcentric bias was significantly stronger for Hispanic names compared with White and Asian names and for White names compared with Asian names (Figure 2).

Figure 2.

Androcentrism WEAT Capturing person=man / gender = woman Associations, Overall and Separately for Each Racial/Ethnic Group.

East vs. South Asian Names

Asian sub-groups vary substantially in their experiences and in how they are stereotyped (Goh & McCue, 2021), even though they are often lumped together in research conducted from a Western perspective. Thus, in a supplementary analysis, we re-classified the Asian names into two subgroups, East Asian and South Asian, based on ratings given by trained, culturally competent coders. We found evidence for reduced androcentrism only in collective concepts of East Asian people (SI.4).

Preregistered Replication Study

A preregistered replication study using GloVe word embeddings (Pennington et al., 2014) replicated all findings in direction and significance (see SI.5).

Discussion

We examined word embeddings for approximately 4,000 first names belonging to millions of individuals to assess androcentric bias in collective concepts. We jointly examined the two components of androcentrism: person = man and gender = woman. We found a substantial person = man / gender = woman association (d = 1.07) that was similar in magnitude to well-studied gender-stereotypical associations (e.g., math = man / arts = woman, d = 1.06; Caliskan et al., 2017). This bias was observed across a range of racial/ethnic groups. We also found it was slightly (but significantly) higher about Hispanic people than White people and substantially higher for Hispanic and White people compared with Asian people (in particular, East Asian people, as found in the supplementary analysis).

Study 2: Testing person = man

As with all traditional IATs and WEATs, the androcentrism WEAT from Study 1 does not differentiate between two possible sources of bias: here, person = man and gender = woman. Thus, in Study 2, we tested the person = man component on its own. We expected to find person = man bias about all races/ethnicities investigated. Based on prior intersectionality theory and research (e.g., Schug et al., 2015), we also expected this bias to be weaker in collective concepts of Asian people, moderate in collective concepts of White people, and possibly stronger in collective concepts of Hispanic people.

To investigate whether person = man, we used an analytic approach that retained more of the nuance and complexity in the data than the WEAT. That is, we computed cross-classified mixed-effects models with the data at the level of each individual first name rather than using difference scores or averaging across sets of words. By not relying on a difference score between associations with generic person and gender concepts, the mixed-effects approach can isolate the person = man component of the WEAT and reveal other biases in the person concept (e.g., person = white), if present. In addition, by not averaging across the names associated with a particular gender or racial/ethnic group, the mixed-effects approach can incorporate additional information about each name. Key for our purposes, we used analytic weights based on the names’ frequency in the training corpus and in the database of first names.

Analyses

We computed two mixed-effects models. First, we computed a model with the names’ similarity to person words predicted by name gender (dummy coded: men, women) and name race/ethnicity (dummy coded: Asian, Hispanic, and White) with crossed random intercepts for name and person concept word. Second, we computed this model again but added the interaction between the first names’ gender and race/ethnicity. We used this two-step approach to simplify estimating the main effects of name gender and name race/ethnicity. Fitted means and standard errors are reported.

In these two models, we also included weights based on frequency information. Including these weights meant that our analyses relied more on names that were more frequent in the corpus—and thus had more accurate word embeddings (Mu & Viswanath, 2018)—and names that were more frequent in the database for people of that self-identified race/ethnicity, which meant that their race/ethnicity classifications were based on more information. For instance, the name Jun was weighed more than Xuejun (respective weights = 6.069 and 1.494) despite both being high-probability Asian names (respective probabilities = 0.966 and 1.000) because Jun was more frequent in the training corpus and belonged to 262 compared with 15 self-identified Asian people in the names database. For details about how we computed these weights, see the supplementary materials (SI.6).

Results

Words for person were more similar in meaning to men’s first names (M = 0.047, SE = 0.007) than women’s first names (M = 0.045, SE = 0.007). This finding conceptually replicates previous research on person = man bias in collective concepts (Bailey et al., 2022), except with a large list of first names (e.g., Qi) to denote gender. This result suggests that the concept person is more associated with men than women (see Table 2, Model 1⁵).

Table 2.

Androcentric Bias, White-Centric Bias, and Intersectionality Bias in Collective person Concepts

Model	Term	B (SE)
1	Woman (= 1) vs. man names (= 0)	−0.041 (0.010)***
	Asian (= 1) vs. White names (= 0)	−0.007 (0.014)
	Hispanic (= 1) vs. White names (= 0)	−0.079 (0.017)***
	Hispanic (= 1) vs. Asian names (= 0)	−0.072 (0.020)***
2	Woman (= 1) vs. man (= 0) × Asian (= 1) vs. White names (= 0)	0.074 (0.029)*
	Woman (= 1) vs. man (= 0) × Hispanic (= 1) vs. White names (= 0)	0.033 (0.034)
	Woman (= 1) vs. man (= 0) × Hispanic (= 1) vs. Asian names (= 0)	−0.041 (0.041)
	Simple slope of woman (= 1) vs. man names (= 0) for Asian names	0.018 (0.026)
	Simple slope of woman (= 1) vs. man names (= 0) for Hispanic names	−0.022 (0.032)
	Simple slope of woman (= 1) vs. man names (= 0) for White names	−0.056 (0.012)***

Note. This table shows two different models: Similarity to person is predicted by name gender and name race/ethnicity without the name Gender × Name Race/Ethnicity interaction terms (Model 1) and with the interaction terms included (Model 2). The coefficients are from a model in which the outcome was standardized and can thus be interpreted as a measure of effect size analogous to Cohen’s d. Each model was re-computed 2 to 3 times, each time changing the reference group to facilitate interpretation. For Model 2, the simple slopes for race/ethnicity are not reported here for brevity but were included when computing the model.

p < .05. ***p < .001.

In the same model, words for person were also more similar to names typical of White individuals (M = 0.047, SE = 0.007) and Asian individuals (M = 0.046, SE = 0.007) compared with Hispanic individuals (M = 0.042, SE = 0.007). This result suggests the presence of person = white and person = asian associations in collective concepts, new to this research (see Table 2, Model 1). This finding is partially in line with the hypothesized White-centric bias.

To examine intersectionality in androcentrism, we tested whether the person = man bias differs in magnitude across racial/ethnic groups by adding the name Gender × Name Race/Ethnicity interaction terms. We found that the person = man association was stronger for White than Asian first names: Generic person words were more associated with White man names (M = 0.048, SE = 0.007) than White woman names (M = 0.045, SE = 0.007; see simple slopes in Table 2, Model 2). In contrast, generic person words were non-significantly less associated with Asian man names (M = 0.046, SE = 0.007) than Asian woman names (M = 0.047, SE = 0.007). Androcentric bias was not significant but in the expected direction for Hispanic names: Generic person words were non-significantly more associated with Hispanic man names (M = 0.043, SE = 0.007) than Hispanic woman names (M = 0.041, SE = 0.007).

Finally, we conducted a preregistered replication study using GloVe embeddings rather than fastText (see Table S2 in SI.5). With respect to androcentrism, we found an overall person = man bias (across all names) with the GloVe embeddings, just as we had found with fastText embeddings. With respect to racial/ethnic bias, we again found that person is more associated with White and Asian people than Hispanic people, replicating the results above. With respect to intersectionality in androcentrism, we again found greater person = man bias in collective concepts of White than Asian people. However, the GloVe embeddings also revealed a significant person = man bias in collective concepts of Hispanic people; this bias was in the predicted direction but not significant with fastText embeddings reported above.

Discussion

In Study 2, we uncovered a person = man bias in collective concepts, replicating prior work with a different approach (Bailey et al., 2022). The more distinctive contribution of this study was to reveal that, consistent with theories of intersectionality, person = man bias is stronger in collective concepts of White people than Asian people. Finally, with respect to racial bias, we found that person was more associated with White and Asian people than with Hispanic people. The difference between White and Hispanic names is consistent with the hypothesized White-centric bias in collective concepts. The difference between Asian and Hispanic names was not predicted but could be connected to relative differences in status between Asian and Hispanic individuals in U.S. and other Western societies (Kuo et al., 2020; Zou & Cheryan, 2017).

Study 3: Testing gender = woman

Using a similar approach to Study 2, in Study 3, we isolated the second component of the androcentrism WEAT. We tested the prediction that women are more associated with their matching gender concept (woman) than men are with theirs (man). That is, we tested whether Jane, Juana, and other high-probability women’s first names are more similar in meaning to generic woman words (e.g., female) than John, Juan, and other high-probability men’s names are to generic man words (e.g., male).

We expected to find a gender = woman bias across races/ethnicities but possibly more about Asian people and less about Hispanic people compared with White people (e.g., Johnson et al., 2012). We also expected to find additional evidence for White-centric biases, with White women’s and men’s first names being more similar in meaning to the corresponding collective gender concepts (woman and man, respectively) than the first names of women and men from other racial/ethnic groups (e.g., Hegarty, 2017).

Analyses

As in the preceding study, we computed two mixed-effects models. First, we computed a model with the first names’ similarity to woman or man words, depending on how the first name was itself classified (e.g., Jane–woman words, John–man words), as an outcome. The predictors included name gender (dummy coded: men, women) and name race/ethnicity (dummy coded: Asian, Hispanic, and White) with crossed random intercepts for first name and generic gender concept word. Second, we computed this model again but added the Name Gender × Name Race/Ethnicity interaction term. We again included frequency weights.

Results

As expected, women’s first names were more similar in meaning to woman words (M = 0.129, SE = 0.010) than men’s first names were to man words (M = 0.096, SE = 0.010). Women were thus associated with their matching gender more than men were with theirs, in support of the gender = woman component of androcentrism (Table 3, Model 1).

Table 3.

Androcentric Bias, White-Centric Bias, and Intersectionality Bias in Collective Gender Concepts

Model	Term	B (SE)
1	Woman (= 1) vs. man names (= 0)	0.409 (0.170)*
	Asian (= 1) vs. White names (= 0)	−0.225 (0.016)***
	Hispanic (= 1) vs. White names (= 0)	−0.189 (0.019)***
	Hispanic (= 1) vs. Asian names (= 0)	0.036 (0.023)
2	Woman (= 1) vs. man (= 0) × Asian (= 1) vs. White names (= 0)	−0.202 (0.033)***
	Woman (= 1) vs. man (= 0) × Hispanic (= 1) vs. White names (= 0)	0.106 (0.038)**
	Woman (= 1) vs. man (= 0) × Hispanic (= 1) vs. Asian names (= 0)	0.308 (0.047)***
	Simple slope of woman (= 1) vs. man names (= 0) for Asian names	0.228 (0.172)
	Simple slope of woman (= 1) vs. man names (= 0) for Hispanic names	0.535 (0.173)**
	Simple slope of woman (= 1) vs. man names (= 0) for White names	0.429 (0.170)*

Note. This table shows two different models: Similarity to matching gender concepts is predicted by name gender and name race/ethnicity without the name Gender × Name Race/Ethnicity interaction terms (Model 1) and with the interaction terms included (Model 2). The coefficients are from a model with a standardized outcome variable and can thus be interpreted as a measure of effect size analogous to Cohen’s d. Each model was re-computed 2 to 3 times, each time changing the reference group to facilitate interpretation. For Model 2, the simple slopes for race/ethnicity are not reported here for brevity but were included when computing the model.

p < .05. **p < .01. ***p < .001.

In the same model, woman and man words were more similar to first names for White women and men (M = 0.118, SE = 0.007), respectively, compared with names for Hispanic women and men (M = 0.103, SE = 0.007) and names for Asian women and men (M = 0.100, SE = 0.007; Table 3, Model 1). This finding provides evidence for White-centric biases in collective gender concepts, new to this research: The collective gender concepts woman and man are more similar to concepts of White women and White men, respectively, than they are to concepts of Hispanic and Asian women and men (i.e., woman = white and man = white).

In terms of intersectionality, we tested whether the gender = woman bias differed in magnitude across different racial/ethnic groups by adding the Name Gender × Name Race/Ethnicity interaction terms to the model. Contrary to our predictions, we found that the gender = woman bias was stronger for Hispanic than White and Asian names and for White names than Asian names. Words denoting generic woman concepts were more similar in meaning to Hispanic women’s first names (M = 0.123, SE = 0.010) than words denoting generic man concepts were to Hispanic men’s first names (M = 0.080, SE = 0.010), and the same was true for White first names (M = 0.134, SE = 0.010 for women vs. M = 0.099, SE = 0.010 for men). For Asian first names, however, this difference was not significant (M = 0.106, SE = 0.010 for women vs. M = 0.088, SE = 0.010 for men; see the simple slopes in Table 3, Model 2).

Finally, we conducted another preregistered replication study using GloVe embeddings rather than fastText (see Table S3 in SI.5). With respect to androcentrism, the preregistered replication study found an overall gender = woman bias with GloVe embeddings as well. With respect to racial/ethnic bias, we again found evidence that woman and man concepts are more associated with White women and men, respectively, than with women and men of other races/ethnicities. With respect to intersectionality in androcentrism, the gender = woman bias was again strongest for Hispanic names, just as it was for fastText embeddings. However, in the GloVe replication study, the gender = woman bias was now smallest for White names rather than Asian names, a deviation from the findings with fastText.

Discussion

Study 3 provides evidence for a gender = woman bias in collective concepts. That is, collective concepts portray women as more gendered than men. Notably, the magnitude of this component of androcentrism (d = 0.41; Table 3) was substantially larger than that of the person = man component (d = 0.04; Table 2)—an important finding given that the gender = woman component of androcentrism has received less empirical attention. With regard to intersectionality, we unexpectedly found greater gender = woman bias about Hispanic people than White and Asian people. Finally, we found evidence that generic woman and man concepts were more associated with White women and men, respectively, than with Asian and Hispanic women and men. This is evidence of White-centric bias in the collective gender concepts.

General Discussion

We investigated three collective concepts that are foundational to human psychology and societal functioning: person, woman, and man. Our primary goal was to provide an investigation of androcentrism—including both its person = man and gender = woman components—and intersections between androcentrism and race/ethnicity. Using the WEAT to measure both components jointly (Study 1), we found evidence for androcentric bias of substantial magnitude (d = 1.07). We found androcentrism across all racial/ethnic groups investigated, with larger biases about Hispanic and White people (ds > 1.10) compared with Asian people (d = 0.56). We then tested and found evidence for, the separate person = man (Study 2) and gender = woman (Study 3) components of androcentric bias. Consistent with the results of the WEAT, we also observed that the person = man bias was stronger about White people than Asian people and the gender = woman bias was stronger about Hispanic people than White and Asian people. Building confidence in these results, all of the findings highlighted in this summary replicated in pre-registered replication studies using a different type of word embeddings (GloVe).

As a secondary goal, we also conducted an initial investigation of three possible White-centric biases in collective concepts: person = white (Study 2), woman = white (Study 3), and man = white (Study 3). We found consistent evidence for all three of these biases in our original analyses and replication studies. Unexpectedly, however, we also found some evidence for a person = asian bias, discussed more below.

In summary, the present research reveals pervasive androcentric and White-centric biases in collective concepts extracted from over 600 billion words written by millions of individuals. This work contributes nuanced, theory-driven findings to interdisciplinary literatures on androcentrism and intersectionality, and begins to address analogous race-based White-centric biases. Next, we enumerate and expand on the contributions of the present work.

Androcentrism in Collective Concepts

First, we found evidence for androcentrism in collective concepts. We found a greater person=man / gender = woman association than the reverse in the WEAT. This WEAT finding jointly captures the two components of androcentrism and mirrors findings obtained with traditional social-psychological methods, including the androcentrism IAT (Bailey et al., 2020), but at a societal scale. The size of the androcentrism effect (d = 1.07) was large and similar in magnitude to that found for well-established gender biases in collective concepts: for instance, science = man / arts = woman and math = man / arts = woman (ds = 1.24 and 1.06, respectively; Caliskan et al., 2017). This is particularly striking given that, in principle, (a) the person concept applies equally to women and men and (b) the man concept applies to individual men just as much as the woman concept applies to individual women.

Our analyses also confirmed that the two components of androcentrism made independent contributions to the overall effect. Study 2 identified a person = man bias, replicating prior work on collective concepts that used different word lists to investigate the same bias (Bailey et al., 2022). New here, we found that the gender = woman component of androcentrism is also present in collective concepts—that is, women’s names were more associated with woman than men’s names were with man. This is a key contribution because theories of androcentrism portray these two components as separate but interlocking: Part of what causes person to be associated with man more than woman is that women are “othered” as especially gendered. Notably, even work using traditional social-psychological methods has focused mostly on the person = man component of androcentrism. The reason for this emphasis is unclear, given that both aspects of androcentrism are described in early theorizing (Bem, 1993). Speculatively, this emphasis might reflect the very bias this literature seeks to address, in that it privileges investigation of the collective concept man over the collective concept woman. The present work provides the first evidence for the complementary gender = woman bias in collective concepts. It is also worth noting that the size of the gender = woman component was much larger (d = 0.41) than the person = man component (d = 0.04), highlighting the importance of investigating both.⁶

Androcentrism Intersects With Race/Ethnicity

Second, we tested whether androcentrism would emerge and/or differ in magnitude about people of different races/ethnicities. We found evidence for androcentrism in collective concepts of all three races/ethnicities investigated: Asian, Hispanic, and White. This is a critical contribution because prior work on this topic has typically focused either directly or indirectly on White people, leaving open the possibility that androcentrism might not emerge about other race/ethnicity groups. In the present work, we found slightly stronger androcentrism about Hispanic people than White people and substantially stronger androcentrism about Hispanic and White people compared with Asian people (especially East Asian people, Study 1).

The analyses on the separate person = man and gender = woman components clarified the nature of these intersectional differences. The person = man association was reduced about Asian names (Study 2), meaning that the reduced androcentrism WEAT difference observed for Asian names (d = 0.56) compared with the other groups was observed at least in part because person is not associated with Asian men more than Asian women. This finding is consistent with prior social-psychological theory and research on intersectionality which suggests that in Western contexts, Asian women are more—or at least equally—prototypical of their group compared with Asian men (Schug et al., 2015). The present finding supports these theoretical claims using a very different methodology and at a much larger scale of analysis.

The gender = woman bias also differed in magnitude across racial/ethnic groups (Study 3). The only difference that emerged consistently across the main and replication studies was that gender = woman bias was greater for Hispanic people than White and Asian people. Thus, the slightly stronger androcentrism WEAT for Hispanic names (d = 1.29, vs. White names d = 1.12) was observed at least in part because Hispanic women are more gendered than Hispanic men (relative to the analogous differences for other races/ethnicities). Notably, this pattern of results is not consistent with gendered race theory, which instead predicts that Hispanic men (vs. women) and Asian women (vs. men) should be viewed in terms of their gender, given that they are typically perceived to be, respectively, hyper-masculine and hyper-feminine (Galinsky et al., 2013; Johnson et al., 2012).

White-Centric Bias in Collective Concepts

Third and finally, we examined potential White-centric biases in collective concepts, testing whether the concepts of a person, woman, and man are more associated with White people, White women, and White men, respectively, compared with people, women, and men of other races/ethnicities. In Study 2, we found a bias to associate the concept person with White and Asian people more than Hispanic people (i.e., person = white and person = asian). In Study 3, we also found a bias to associate the gender concepts woman and man with White women and White men, respectively, more than with women and men of other races/ethnicities. These results were mostly in line with the hypothesized White-centric bias (Hegarty, 2017; Morehouse et al., 2023), although the finding that person was also associated with Asian people was not predicted. A possible explanation is that power and status shape the collective concept person such that this concept prioritizes high-status groups and marginalizes low-status groups. This might explain why person was as closely associated with White individuals, who are high-status in U.S. and other Western contexts, as it was with Asian individuals, who are often lauded as high-status “model minorities” (Kuo et al., 2020; Zou & Cheryan, 2017). This possibility awaits further research.

Although investigating White-centric bias was a secondary research goal in the present investigation, this contribution is no less important, both theoretically and societally. White people are prioritized in individuals’ concepts (Hegarty, 2017) as well as in research practices (McGorray et al., 2023; Roberts & Mortenson, 2023) and scientific writing (Cundiff, 2012). Here, we show a similar pattern in collective concepts. Because collective concepts reflect the shared worldview of a community, demonstrating White-centric bias at this level of analysis is particularly troubling.

Broader Implications

The present work has implications for researchers in psychology and artificial intelligence (AI), as well as for society at large. For psychologists, NLP methods applied to large corpora of naturalistic text have unique benefits: They can bypass the thorny problems of demand characteristics and observer effects. In traditional psychological studies, participants often try to guess the researcher’s hypothesis and change their responses accordingly; this tendency is likely heightened when the study concerns sensitive topics, such as gender and race/ethnicity. Even when such task demands are avoided, a study’s measures can themselves influence how participants respond in ways that cannot be fully anticipated or accounted for by researchers. For instance, a prior study of people = man bias instructed participants to list an example of a person, asked about this person’s gender, and found male-centric bias (Bailey et al., 2020). However, it is possible that participants did not consider gender at all until asked about it, which would mean that the resulting bias was in part a byproduct of the question itself. NLP methods can bypass these issues when they are applied to (large quantities of) preexisting text that individuals produce spontaneously rather than in response to an experimenter’s query.

For researchers working on AI, the present work also adds to a growing literature showing that word embeddings and other AI tools encode social biases, including complex and nuanced intersectional biases (e.g., Caliskan, 2023; Ghai et al., 2021; Guo & Caliskan, 2021; Jiang & Fellbaum, 2020; Lalor et al., 2022; Nelson, 2021; Sabbaghi et al., 2023; Tan & Celis, 2019). This means that even if individual-level biases were hypothetically eliminated, society would still be inequitable because these biases are embedded in active AI tools. Our work thus underscores the importance of rigorously auditing and “debiasing” word embeddings to increase fairness in AI applications. Word embeddings provide a valuable window into where social biases persist in machine learning models. By identifying bias “blind spots” at the intersection of key collective concepts such as gender and race/ethnicity, the present research highlights priority areas for intervention. Concrete next steps should include constructing debiasing techniques tailored to the intersections revealed here and incentivizing commercial providers of AI tools to allow independent scrutiny of their models, including the source code. In addition, because the biases uncovered here stem in part from the underlying training data itself (i.e., the text being analyzed, which reflects systemic prejudices in society), effective mitigation requires an interdisciplinary approach that incorporates expertise from the humanities and social sciences, as well as greater diversity among AI researchers and greater collaboration with stakeholders from marginalized groups. Fostering a culture of transparency and inclusion will help the field make progress toward fairer AI.

Our work also has broader implications for society. Throughout, we have argued that collective concepts are important to understand. Why? These concepts reflect widely shared ideas that are “in the air” for people in a community (see related theorizing on the bias of crowds; Payne et al., 2017). These collective ideas exist independently of personal endorsement. For instance, any given individual might disagree that “person = man” if asked, but this idea is nevertheless part of the cultural “air” they breathe and might unwittingly influence their thoughts and behaviors. In fact, the very nature of communication requires individuals to contend with ideas they themselves might not endorse (Grice, 1975), further perpetuating these ideas in the process. For this reason, collective concepts—and the biases therein—can be remarkably enduring, even over extended timeframes (Payne et al., 2019).

This is troubling because the collective concepts we focused on— person, woman, and man—are foundational to policy- and decision-making across domains. Because these concepts prioritize men and White people, any policies or decisions based on them might advantage these groups. In the legal domain, decisions based on the “reasonable person” standard might advantage White male defendants over defendants from other groups. In global ethics, the notion of “human rights” might be applied unevenly across different groups. In the consumer domain, products designed with “people” in mind (e.g., vehicles) might be a better fit and safer for men to use than women (Epker, 2023). Hair products designed with “women” in mind might work better for White women than women of color. In the medical domain, research samples of “patients” tend to overrepresent White people and men (Cundiff, 2012; Shansky & Murphy, 2021). Patient symptom profiles used by both human and AI decision-makers better represent White people and men than other groups, with serious consequences for the health of women and people of color, including death (Obermeyer et al., 2019; Woodward, 2019).

Can anything be done to change (biases in) collective concepts? The fact that these concepts are encoded and perpetuated through language and other cultural artifacts suggests a path forward. For instance, gendered language can be used as a flexible tool to remind people that women exist and are people too (e.g., “A person . . . she”; Atir, 2022). Laws and policies might also be able to shape collective concepts. For instance, legalizing marriage equality produced substantial changes in large-scale attitudes toward sexual minorities (Ofosu et al., 2019). Policy can be leveraged to address the effects of androcentric and White-centric biases as well, not just the biases themselves. For instance, several funders now require medical researchers seeking grants to include racial minorities and women (as well as female rats) in their patient samples, although enforcement is still an outstanding question (Shansky & Murphy, 2021). Our work underscores the need for such systemic solutions to mitigate (the effects of) androcentric, White-centric, and intersectional biases.

Collective Concepts About Other Groups

One of the limitations of the present work is the absence of the Black names from the main analyses, despite the fact that much of the original theory (e.g., Crenshaw, 1990; Purdie-Vaughns & Eibach, 2008) and research (e.g., Coles & Pasek, 2020) on intersectionality focuses on Black women’s experiences and biases about this group. We omitted these names because the results from an initial validation study (see SI.2) indicated that they would not yield interpretable results: Black names were not a valid measure of collective concepts of Black people; they seemed to track collective concepts of White people instead. (Analyses with the Black names are reported in SI.3 for completeness)

One reason why the Black names on our list did not sufficiently differentiate collective concepts of Black and White people could be that the names came from mortgage databases and thus belonged to disproportionately high-income Black individuals. There is evidence that high-income Black families are more likely to choose less distinctively “Black” names and instead choose names that are also common among White people (Fryer & Levitt, 2004). Another reason why the Black names were less distinct could be that, historically, both the United States and Canada have been dominated by White, English-speaking people of European descent. Black people differ from the Asian and Hispanic groups in how recently and how fully they have joined—or been forced to join—this English-language context (Cox & Tamir, 2022). These points do not undermine our findings about high-probability Asian, Hispanic, and White names, but they do suggest that future text-based research should rely on markers of race/ethnicity other than names, particularly in cultural contexts where the racial/ethnic groups in question have high degrees of linguistic overlap.

Another limitation of the present work is that we only investigated representations of binary gender groups: women and men. We also think it is important for future work to spotlight representations of non-binary people and other gender minorities, who might be particularly marginalized in the person concept (Tebbe & Budge, 2022). However, first names might also not be the best tool for capturing ideas about non-binary people in future text-based research. Almost all of the 4,000+ names in the present research were classified as highly probable of either women or men; only 4 names were ambiguous. However, other word lists could be used, such as neopronouns (e.g., ze) or terms for the identities themselves (e.g., non-binary).

Androcentrism Across Contexts

In the present research, we used word embeddings extracted from large-scale text corpora that collapsed across a wide range of different contexts and authors (e.g., news outlets, personal blogs, instructional manuals, and government websites; Dodge et al., 2021; Mehmood et al., 2017). This approach has both advantages and disadvantages. An important advantage is that a 600+ billion word corpus that consists of language from millions of individuals discussing a huge range of topics enables us to capture widely shared, truly collective understandings of what it means to be a person, a woman, and a man. A disadvantage of the present approach, however, is that we are not in a position to describe variation across contexts, authors, or sub-communities. Theories of intersectionality note that the dimensions of identity that are most salient (e.g., gender vs. race/ethnicity) vary across individuals, contexts, and communities (Petsko et al., 2022). In addition, the specific attributes attached to gender groups can also vary based on the local context and comparison sets (A. Martin, 2022). One way to characterize such variation would be to use word embeddings trained on different, more specialized text (Charlesworth et al., 2021). However, once a certain threshold for the size of the text is reached, the resulting embeddings can be quite similar, even with respect to social biases, regardless of whether they are trained on general-purpose text written by a wide range of authors (e.g., text on the internet) or on more specialized text written by a narrower set of authors (e.g., biomedical text or child-directed text; Bailey et al., 2022; Charlesworth et al., 2021).

It is also noteworthy that collective concepts likely reflect the views of high-power, majority-group members, whose views are more likely to be reflected in public language and institutions. Thus, the intersectional biases we found here likely reflect biases about marginalized groups from a White-majority perspective rather than biases coming from these subcommunities themselves. Future research on androcentrism and intersecting racial/ethnic biases should strive to use methods besides those we used here, specifically methods that allow for greater control, to precisely identify boundary conditions, moderating contexts, and the role of author identity.

Although we investigated collective concepts by using language output across a range of contexts and authors, in terms of cultural context the present investigation is likely tapping a Western perspective. Although the training corpus (the Common Crawl) does include text from multiple cultures and in multiple languages, it nevertheless overrepresents English-language text (Dodge et al., 2021; Mehmood et al., 2017). Our lists of words were also all in English and/or came from U.S. and Canadian sources. Little is known about androcentrism in other cultural contexts. Future research should investigate androcentric gender bias and its intersection with race/ethnicity-based biases in multiple different linguistic and cultural contexts around the world.

Conclusion

We investigated androcentric, White-centric, and intersectional biases in the collective understanding of three fundamental concepts (person, woman, and man) and found that these concepts prioritize men and White individuals as people and “other” women as people with gender. Because these concepts scaffold many important decisions made by human—and also increasingly by artificial intelligence—agents, it is critical that they are inclusive and fair.

Supplemental Material

sj-docx-1-psp-10.1177_01461672241232114 – Supplemental material for Intersectional Male-Centric and White-Centric Biases in Collective Concepts

Supplemental material, sj-docx-1-psp-10.1177_01461672241232114 for Intersectional Male-Centric and White-Centric Biases in Collective Concepts by April H. Bailey, Adina Williams, Aashna Poddar and Andrei Cimpian in Personality and Social Psychology Bulletin

Footnotes

Acknowledgements

The authors thank Nicholas DiMaggio for assistance with manuscript preparation.

Author Contributions

A.H.B.: conceptualization, project administration, methodology, investigation, formal analysis, data curation, visualization, writing—original draft. A.W.: conceptualization, methodology, investigation, software. A.P.: conceptualization, methodology, investigation, writing—review and editing. A.C.: conceptualization, methodology, supervision, and writing—review and editing.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

April H. Bailey

Andrei Cimpian

Data and Materials Availability

Datafiles, analysis scripts, and the preregistration for the replication studies are available on the Open Science Framework (OSF; ).

Supplemental Material

Supplemental material is available online with this article.

Notes

References

Aladangady

Forde

(2021). Wealth inequality and the racial wealth gap. Board of Governors of the Federal Reserve System.

Amodio

Bartholow

Ito

(2014). Tracking the dynamics of the social brain: ERP approaches for social cognitive and affective neuroscience. Social Cognitive and Affective Neuroscience, 9(3), 385–393.

Atir

(2022). Girlboss? Highlighting versus downplaying gender through language. Trends in Cognitive Sciences, 26(8), 623–625.

Bailey

LaFrance

(2016). Anonymously male: Social media avatar icons are implicitly male and resistant to change. Cyberpsychology: Journal of Psychosocial Research on Cyberspace, 10, Article 8.

Bailey

LaFrance

(2017). Who counts as human? Antecedents to androcentric behavior. Sex Roles, 76, 682–693.

Bailey

LaFrance

Dovidio

(2019). Is man the measure of all things? A social cognitive account of androcentrism. Personality and Social Psychology Review, 23(4), 307–331.

Bailey

LaFrance

Dovidio

(2020). Implicit androcentrism: Men are human, women are gendered. Journal of Experimental Social Psychology, 89, 103980.

Bailey

Williams

Cimpian

(2022). Based on billions of words on the internet, people=men. Science Advances, 8, Article eabm2463.

Bem

(1993). The lenses of gender: Transforming the debate on sexual inequality. Yale University Press.

10.

Bowleg

(2008). When Black+ lesbian+ woman≠ Black lesbian woman: The methodological challenges of qualitative and quantitative intersectionality research. Sex Roles, 59, 312–325.

11.

Caliskan

(2023). Artificial intelligence, bias, and ethics. Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, 6, 7007–7013. https://doi.org/10.24963/ijcai.2023/799

12.

Caliskan

Bryson

Narayanan

(2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356, 183–186.

13.

Charlesworth

Yang

Mann

Kurdi

Banaji

(2021). Gender stereotypes in natural language: Word embeddings show robust consistency across child and adult language corpora of more than 65 million words. Psychological Science, 32(2), 218–240.

14.

Cole

(2009). Intersectionality and research in psychology. American Psychologist, 64(3), 170–180.

15.

Coles

S. M.

Pasek

(2020). Intersectional invisibility revisited: How group prototypes lead to the erasure and exclusion of Black women. Translational Issues in Psychological Science, 6(4), 314–324.

16.

Connor

Weeks

Glaser

Chen

Keltner

(2022). Intersectional implicit bias: Evidence for asymmetrically compounding bias and the predominance of target gender. Journal of Personality and Social Psychology, 124(1), 22–48.

17.

Cotton

O’Neill

Griffin

(2014). Whiteness of a name: Is “White” the baseline? Journal of Managerial Psychology, 29(4), 405–422.

18.

Cox

Tamir

(2022). Race is central to identity for Black Americans and affects how they connect with each other: 2 Family history, slavery and knowledge of Black history. Pew Research Center.

19.

Crenshaw

(1990). Mapping the margins: Intersectionality, identity politics, and violence against women of color. Stanford Law Review, 43, 1241–1299.

20.

Cundiff

(2012). Is mainstream psychological research “womanless” and “raceless”? An updated analysis. Sex Roles, 67, 158–173.

21.

Danbold

Huo

(2015). No longer “all-American”? Whites’ defensive reactions to their numerical decline. Social Psychological and Personality Science, 6(2), 210–218.

22.

de Beauvoir

. (2010). The second sex. Knopf. (Original work published 1949)

23.

Dembroff

(2019). Beyond binary: Genderqueer as critical gender kind. Philosopher’s Imprint.

24.

Devos

Banaji

(2005). American=White? Journal of Personality and Social Psychology, 88(3), 447–466.

25.

Dodge

Sap

Marasović

Agnew

Ilharco

Groeneveld

. . . Gardner

(2021). Documenting Large Webtext Corpora: A case study on the Colossal Clean Crawled Corpus. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 1286–1305).

26.

Durkheim

(1915). The elementary forms of religious life. Oxford University Press.

27.

Eagly

Kite

(1987). Are stereotypes of nationalities applied to both women and men? Journal of Personality and Social Psychology, 53(3), 451–462.

28.

Epker

(2023, September 13). Fasten your seatbelts: A female car crash test dummy represents average women for the first time in 60+ years. Forbes. https://www.forbes.com/sites/evaepker/2023/09/12/fasten-your-seatbelts-a-female-car-crash-test-dummy-represents-average-women-for-the-first-time-in-60-years/?sh=28c722bd55ba

29.

Field

Blodgett

Waseem

Tsvetkov

Y. A.

(2021). Survey of race, racism, and Anti-racism in NLP. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 1, 1905–1925.

30.

Fiske

S. T.

Neuberg

S. L.

(1990). A continuum of impression formation, from category-based to individuating processes: Influences of information and motivation on attention and interpretation. In Zanna

M. P.

(Ed.), Advances in experimental social psychology (Vol. 23, pp. 1–74). Academic Press.

31.

Fryer

Jr. Levitt

(2004). The causes and consequences of distinctively Black names. The Quarterly Journal of Economics, 119(3), 767–805.

32.

Galinsky

Hall

Cuddy

(2013). Gendered races: Implications for interracial marriage, leadership selection, and athletic participation. Psychological Science, 24(4), 498–506.

33.

Gastil

(1990). Generic pronouns and sexist language: The oxymoronic character of masculine generics. Sex Roles, 23, 629–643.

34.

Ghai

Hoque

M. N.

Mueller

(2021). WordBias: An interactive visual tool for discovering intersectional biases encoded in word embeddings. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems (pp. 1–7). Association for Computing Machinery. https://doi.org/10.1145/3411763.3451587

35.

Ghavami

Peplau

(2013). An intersectional analysis of gender and ethnic stereotypes: Testing three hypotheses. Psychology of Women Quarterly, 37(1), 113–127.

36.

Gilman

(1911). The man-made world: Or, our androcentric culture (3rd ed.). Charlotte Company.

37.

Goff

Kahn

(2013). How psychological science impedes intersectional thinking. Du Bois Review: Social Science Research on Race, 10(2), 365–384.

38.

Goh

McCue

(2021). Perceived prototypicality of Asian subgroups in the United States and the United Kingdom. Journal of Experimental Social Psychology, 97, 104201.

39.

Goodhew

Reynolds

Edwards

Kidd

(2022). The content of gender stereotypes embedded in language use. Journal of Language and Social Psychology, 41(2), 219–231.

40.

Greenwald

McGhee

Schwartz

(1998). Measuring individual differences in implicit cognition: The implicit association test. Journal of Personality and Social Psychology, 74(6), 1464–1480.

41.

Grice

H. P.

(1975). Logic and conversation. In Cole

Morgan

(Eds.), Syntax and semantics, Vol. 3: Speech acts (pp. 41–58). Academic Press.

42.

Guo

Caliskan

(2021). Detecting emergent intersectional biases: Contextualized word embeddings contain a distribution of human-like biases. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (pp. 122–133). Association for Computing Machinery. https://doi.org/10.1145/3461702.3462536

43.

Hamilton

(1991). Masculine bias in the attribution of personhood. Psychology of Women Quarterly, 15(3), 393–402.

44.

Hannak

Joseph

Larremore

D. B.

Cimpian

(2023). Field-specific ability beliefs as an explanation for gender differences in academics’ career trajectories: Evidence from public profiles on ORCID.Org. Journal of Personality and Social Psychology, 125(4), 681–698.

45.

Harris

(1954). Distributional structure. Word, 10(2–3), 146–162.

46.

Hegarty

(2017). On the failure to notice that White people are White: Generating and testing hypotheses in the celebrity guessing game. Journal of Experimental Psychology: General, 146(1), 41–62.

47.

Hegarty

Parslow

Ansara

Y. G.

Quick

(2013). Androcentrism: Changing the landscape without leveling the playing field. In Ryan

M. K.

Branscombe

N. R.

(Eds.), The Sage handbook of gender and psychology (pp. 29–44). Sage.

48.

Jiang

Fellbaum

(2020). Interdependencies of gender and race in contextualized word embeddings. In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing (pp. 17–25). Association for Computational Linguistics. https://aclanthology.org/2020.gebnlp-1.2

49.

Johnson

K. L.

Freeman

J. B.

Pauker

(2012). Race is gendered: How covarying phenotypes and stereotypes bias sex categorization. Journal of Personality and Social Psychology, 102(1), 116–131.

50.

Jones

Marks

Ramirez

Rios-Vargas

(2021). 2020 census illuminates racial and ethnic composition of the country. United States Census Bureau. https://www.census.gov/library/stories/2021/08/improved-race-ethnicity-measures-reveal-united-states-population-much-more-multiracial.html

51.

Kuo

Kraus

Richeson

(2020). High-status exemplars and the misperception of the Asian-White wealth gap. Social Psychological and Personality Science, 11(3), 397–405.

52.

Kurdi

Morehouse

K. N.

Dunham

(2023). How do explicit and implicit evaluations shift? A preregistered meta-analysis of the effects of co-occurrence and relational information. Journal of Personality and Social Psychology, 124(6), 1174–1202.

53.

Kurdi

Seitchik

Axt

Carroll

Karapetyan

Kaushik

Tomezsko

Greenwald

A. G.

Banaji

(2019). Relationship between the Implicit Association Test and intergroup behavior: A meta-analysis. American Psychologist, 74(5), 569–586.

54.

Lalor

Yang

Smith

Forsgren

Abbasi

(2022). Benchmarking intersectional biases in NLP. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 3598–3609). Association for Computing Machinery. https://doi.org/10.18653/v1/2022.naacl-main.263

55.

Lei

R. F.

Foster-Hanson

Goh

J. X.

(2023). A sociohistorical model of intersectional social category prototypes. Nature Reviews Psychology, 2(5), 297–308.

56.

Lenci

(2018). Distributional models of word meaning. Annual Review of Linguistics, 4, 151–171.

57.

Lewis

Lupyan

(2020). Gender stereotypes are reflected in the distributional structure of 25 languages. Nature Human Behaviour, 4(10), 1021–1028.

58.

Lorde

(1984). Sister outsider: Essays and speeches. Crossing Press.

59.

Martin

(2022). Gender relativism: How context shapes what is seen as male and female. Journal of Experimental Psychology: General, 152(2), 322–345.

60.

Martin

Mason

(2022). What does it mean to be (seen as) human? The importance of gender in humanization. Journal of Personality and Social Psychology, 123(2), 292–315.

61.

Martin

Papadelos

(2017). Who stands for the norm? The place of metonymy in androcentric language. Social Semiotics, 27(1), 39–58.

62.

McDonald

Ramscar

(2001). Testing the distributional hypothesis: The influence of context on judgments of semantic similarity. Proceedings of the Annual Meeting of the Cognitive Science Society, 23, 1–6.

63.

McGorray

Emery

Garr-Schultz

Finkel

(2023). “Mostly White, heterosexual couples”: Examining demographic diversity and reporting practices in relationship science research samples. Journal of Personality and Social Psychology, 125(2), 316–344.

64.

Mehmood

Shafiq

Waheed

(2017). Understanding regional context of World Wide Web using common crawl corpus. In IEEE 13th Malaysia International Conference on Communications (MICC) (pp. 164–169). IEEE.

65.

Mikolov

Chen

Corrado

Dean

(2013). Efficient estimation of word representations in vector space [Preprint]. https://arxiv.org/abs/1301.3781

66.

Mikolov

Grave

É.

Bojanowski

Puhrsch

Joulin

(2018, May). Advances in pre-training distributed word representations [Conference session]. Proceedings of the Eleventh International Conference on Language Resources and Evaluation, Miyazaki, Japan.

67.

Miller

Perry

(2012). The reasonable person. New York University Law Review, 87, 323–392.

68.

Momennejad

(2022). Collective minds: Social network topology shapes collective cognition. Philosophical Transactions of the Royal Society B, 377(1843), 20200315.

69.

Morehouse

Maddox

Banaji

(2023). All human social groups are human, but some are more human than others: A comprehensive investigation of the implicit association of “Human” to US racial/ethnic groups. Proceedings of the National Academy of Sciences, 120, Article e2300995120.

70.

Morgenroth

Ryan

(2021). The effects of gender trouble: An integrative theoretical framework of the perpetuation and disruption of the gender/sex binary. Perspectives on Psychological Science, 16(6), 1113–1142.

71.

Moscovici

(1963). Attitudes and opinions. Annual Review of Psychology, 14(1), 231–260.

72.

Viswanath

(2018). All-but-the-top: Simple and effective post-processing for word representations [Conference session]. 6th International Conference on Learning Representations, Vancouver, British Columbia, Canada. April and May 2018.

73.

Nelson

L. K.

(2021). Leveraging the alignment between machine learning and intersectionality: Using word embeddings to measure intersectional experiences of the nineteenth century U.S. South. Poetics, 88, 101539. https://doi.org/10.1016/j.poetic.2021.101539

74.

Nosek

Smyth

Hansen

Devos

Lindner

Ranganath

. . . Banaji

(2007). Pervasiveness and correlates of implicit attitudes and stereotypes. European Review of Social Psychology, 18(1), 36–88.

75.

Obermeyer

Powers

Vogeli

Mullainathan

(2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453.

76.

Ofosu

E. K.

Chambers

M. K.

Chen

J. M.

Hehman

(2019). Same-sex marriage legalization associated with reduced implicit and explicit antigay bias. Proceedings of the National Academy of Sciences, 116(18), 8846–8851.

77.

Payne

B. K.

Vuletich

H. A.

Brown-Iannuzzi

J. L.

(2019). Historical roots of implicit bias in slavery. Proceedings of the National Academy of Sciences, 116(24), 11693–11698.

78.

Payne

B. K.

Vuletich

H. A.

Lundberg

K. B.

(2017). The bias of crowds: How implicit bias bridges personal and systemic prejudice. Psychological Inquiry, 28(4), 233–248.

79.

Pennington

Socher

Manning

(2014). Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (pp. 1532–1543). Association for Computational Linguistics.

80.

Petsko

C. D.

Rosette

A. S.

Bodenhausen

G. V.

(2022). Through the looking glass: A lens-based account of intersectional stereotyping. Journal of Personality and Social Psychology, 123(4), 763–787.

81.

Phills

Williams

Wolff

Smith

Arnold

Felegy

Kuenzig

(2018). Intersecting race and gender stereotypes: Implications for group-level attitudes. Group Processes & Intergroup Relations, 21(8), 1172–1184.

82.

Purdie-Vaughns

Eibach

(2008). Intersectional invisibility: The distinctive advantages and disadvantages of multiple subordinate-group identities. Sex Roles, 59, 377–391.

83.

Roberts

Mortenson

(2023). Challenging the White=neutral framework in psychology. Perspectives on Psychological Science, 18(3), 597–606.

84.

Sabbaghi

S. O.

Wolfe

Caliskan

(2023). Evaluating biased attitude associations of language models in an intersectional context. https://doi.org/10.1145/3600211.3604666

85.

Scanlon

T. M.

(2000). What we owe to each other. Harvard University Press.

86.

Schug

Alt

Klauer

(2015). Gendered race prototypes: Evidence for the non-prototypicality of Asian men and Black women. Journal of Experimental Social Psychology, 56, 121–125.

87.

Sesko

Biernat

(2010). Prototypes of race and gender: The invisibility of Black women. Journal of Experimental Social Psychology, 46(2), 356–360.

88.

Shansky

R. M.

Murphy

A. Z.

(2021). Considering sex as a biological variable will require a global shift in science culture. Nature Neuroscience, 24(4), 457–464.

89.

Sibley

Liu

(2007). New Zealand=bicultural? Implicit and explicit associations between ethnicity and nationhood in the New Zealand context. European Journal of Social Psychology, 37(6), 1222–1243.

90.

Statistics Canada. (2022). The Canadian census: A rich portrait of the country’s religious and ethnocultural diversity. https://www150.statcan.gc.ca/n1/daily-quotidien/221026/dq221026b-eng.htm

91.

Tan

Y. C.

Celis

L. E.

(2019). Assessing social and intersectional biases in contextualized word representations. Advances in Neural Information Processing Systems, 32. https://proceedings.neurips.cc/paper/2019/hash/201d546992726352471cfea6b0df0a48-Abstract.html

92.

Tebbe

E. A.

Budge

S. L.

(2022). Factors that drive mental health disparities and promote well-being in transgender and nonbinary people. Nature Reviews Psychology, 1(12), 694–707.

93.

Tuchman

(1978). Hearth and Home: Images of women in the mass media. Oxford University Press.

94.

Tzioumis

(2018). Demographic aspects of first names. Scientific Data, 5, 1–9.

95.

Uddenberg

Scholl

B. J.

(2018). Teleface: Serial reproduction of faces reveals a whiteward bias in race memory. Journal of Experimental Psychology: General, 147(10), 1466–1487.

96.

Woodward

(2019). Cardiovascular disease and the female disadvantage. International Journal of Environmental Research and Public Health, 16(7), 1165.

97.

Zarate

Smith

(1990). Person categorization and stereotyping. Social Cognition, 8(2), 161–185.

98.

Zou

Cheryan

(2017). Two axes of subordination: A new model of racial position. Journal of Personality and Social Psychology, 112(5), 696–717.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

33.85 MB