Introduction

Diversity is a fundamental part of the advancement of science. Evidence shows that the current lack of social diversity, including gender, race, and ethnicity, in academia represents a highly inefficient equilibrium1,2,3. Limiting the diversity of perspectives not only hinders the scope of inquiry but also reduces the potential for innovative solutions, underscoring the importance of inclusivity in fostering a more robust and dynamic scientific community4,5. For instance, gender equity is listed as one of the 17 goals of the United Nations 2030 Agenda6.

The lack of representation and discrimination against women in academia is a reality that has been widely recognized. Women publish fewer first-authored articles7,8,9, receive smaller grants10,11 and start-up funding12,13, are paid less14, are less invited to talks15, are promoted with reduced frequency, and hold fewer positions of power or influence16,17, such as being reviewers in scientific publications and grants18 or in the editorial board of scientific journals19 (but see ref. 20). All of this contributes to the well-known phenomenon of the “leaky pipeline” of women’s representation in science, i.e., women tend to leave the academic career path earlier than men11,21.

Recent policies have been enacted to tackle the “leaky pipeline” phenomenon and increase the presence of women in university committees, journal editorial boards, scientific events, and organizations22. While these measures primarily focus on enhancing female representation, gender-science stereotypes, which are entrenched and overly simplistic views about gender roles, continue to challenge these efforts by significantly shaping perceptions and behaviors23. Such stereotypes persist as a major source of gender bias in academia, with pervasive cultural effects against equity24,25,26. These stereotypes typically present scientists as male25,27, creating an academic environment that diminishes the visibility and recognition of women’s contributions. This reduced recognition leads to lower prestige for female scientists, perpetuating a vicious cycle that keeps them in a disadvantaged position within academia28. Such dynamics illustrate the complex interplay between affirmative actions aimed at increasing representation and the deep-rooted biases and stereotypes that continue to impede true gender equity.

Using the audience in talks of a seminar series in Ecology, Evolution, and Conservation Biology, we evaluate whether affirmative actions focused on increasing women’s representation as speakers affected their visibility and recognition in science, measured by audience size, as an indirect outcome. To do so, we first evaluated (i) the representation of females as speakers through academic levels and the effect of affirmative actions. This is a necessary step to further understand any possible indirect effect of the affirmative actions on the audience. Then, we analyzed (ii) whether audience size depends on the speaker’s gender, academic level, and affirmative actions for women’s representativeness. As audience size can be influenced by speakers’ attributes other than gender, we additionally evaluated (iii) if differences in the audience of male and female professors reflected differences in the speaker’s career length and productivity. In addition, we considered (iv) whether the research topics covered in the talks might differ between male and female speakers (e.g., ref. 29). We hypothesized that such differences, if present, could contribute to explaining audience size.

We rely on the analysis of decadal-scale data (2008–2019) on women’s representation among speakers, audiences, and topics of the talks in an ecological seminar series at one of the main Latin American universities, the University of São Paulo, Brazil. Such events are fruitful occasions to catalyze learning, discuss ideas, contribute to further developing the speaker’s research, and expand collaboration networks. They are pillars for promoting individual and social changes within scientific communities locally and globally.

Results

From the 327 talks analyzed in 12 years, 184 were given by men (56%) and 143 by women (44%). When separated by academic level (n = 320, excluding non-academic speakers), women gave fewer talks than men at higher academic levels, from 52% of the students’ and 43% of the postdocs’ talks to 24% of the professors’ talks (Fig. 1a). Before 2018, men were the majority of speakers in 7 out of 10 years (Fig. 1b). In 2018 and 2019, after the affirmative actions began, the gender balance among speakers was 52% and 50% women in each respective year.

Fig. 1: Gender representation in speakers at the seminar series over 12 years.
figure 1

a Total number of speakers by gender (females in purple and males in yellow) and academic level for all talks in 12 years of the EcoEcontros seminar series. b Number of talks by gender for each year. The dashed vertical line indicates the beginning of affirmative action to increase women’s representation. Percentages in both figures are the proportion of female researchers within each academic level in (a) and year in (b). A similar figure with only the Graduate Program community is presented in Fig. S2.

Full size image

Female speakers across academic levels

Two models were equally plausible for the proportion of female speakers (Table 1a). Both models included academic level as a predictor, with the difference that the best-fitted model includes affirmative actions and the interaction between them (conditional R2 = 0.15, marginal R2 = 0.12, Fig. 2). Before the start of affirmative action, we found a decrease in the proportion of female speakers through academic levels, with female speakers being only 21% of the professors’ speakers (Fig. 2, gold lines). After implementing affirmative action, the proportion of females in all academic levels became more balanced and did not differ from 50% (Fig. 2, green lines). If we consider the second most plausible model, the proportion of female speakers also decreased with academic level, being smaller than 50% only for female professors (26%, Fig. S3).

Fig. 2: Predicted proportions of female speakers by academic level and affirmative actions.
figure 2

Predictions were made by the best-fitted model (Table 1a), with academic levels on the x-axis and affirmative actions represented in color (before 2018 in gold and after in green). Vertical line ranges mean 95% confidence intervals for the estimated proportions. The size of the circles is proportional to the number of talks given by males (y-axis 0) and females (y-axis 1) in each category, ranging from 3 (smallest circle—male postdocs after affirmative actions) to 69 (largest circle—male professors before the affirmative action).

Full size image
Table 1 Model selection results for speaker representation and seminar audience
Full size table

When considering the subset data for the Graduate Program academic community, we found that the proportion of female speakers followed that of female academics within each academic level (best-fitting model, Fig. 3), suggesting no inherent gender bias in speaker selection within the academic community. However, there was high uncertainty in the model selection with all models being equally plausible (ΔAIC < 2), except the null (Table S1), probably due to a smaller (44% of the original dataset) and imbalanced data between academic levels (99 students, 24 postdocs, 13 professors) and affirmative actions (109 before, 27 after). The marginal R2 of the best-fitted model was 0.07.

Fig. 3: Proportion of female speakers according to the proportion of female academics at the Graduate Program in Ecology (PPGE-USP) population.
figure 3

The solid black line is the predicted relationship from the best-fitting model (Table S1), and the shaded area indicates a 95% confidence interval of the estimates. The dashed line indicates the 1:1 relationship between the proportion of female academics and the proportion of female speakers per year and academic level. Dots represent the 136 talks given by females (1) and males (0), and the proportion of female academics according to the speaker's academic level (colors), which varies by year (years not shown). We created a small displacement in the y-axis around zeros and ones to better show the data for each academic level, and in the x-axis to display the overlaid data points. The R2 of the best-fitting model was 0.07.

Full size image

Speaker gender differences in the seminar’s audience

We found that male professors had the largest audience on average for their talks (Fig. 4a, Table S2). The two equally plausible models for the audience (Table 1b) included gender, academic level, and affirmative actions as predictors, with the difference that the best-fitted model included an interaction between gender and academic level (conditional R2 = 0.22, marginal R2 = 0.18, Fig. 4a and Fig. S4). For both models, (1) male speakers had, on average, a larger audience than female speakers, (2) the higher the academic level, the larger the audience, and (3) affirmative actions increased the audience of the seminars. According to the best-fit model, male professors’ talks had, on average, 1.4 times the audience size of female professors’ talks (predicted values from the model: before affirmative actions—27 and 19 attendees, respectively; after affirmative actions—34 and 24 attendees, respectively).

Fig. 4: Audience size and speaker productivity in relation to gender and affirmative actions.
figure 4

a Audience (number of attendants) in seminars according to gender, academic level, and affirmative actions (before and after 2018) with the prediction (black contour circles) and confidence intervals (vertical black lines) from the best-fitted model for the audience (Table 1b). b Principal component analysis (PCA) for the productivity metrics for professors and institutions (N = 87); for variable code, see Table S3. c The professor’s audience analysis is based on the gender and productivity index (PCA first axis). Lines and shaded areas represent marginal predictions and 95% confidence intervals for the estimates of the best-fitted model with additive effects of productivity index, gender, and affirmative actions. We fixed the affirmative action to ‘before’ to display the predictions because most data come from this period (N = 67).

Full size image

For the subsequent analysis of professors’ talks (N = 87), the PCA results (Fig. 4b) show that career length and productivity metrics for professors were highly correlated with the first axis (52% of variance explained), while the institution indexes composed the second PCA axis (21% of variation explained). In general, male and female professors did not show multivariate differences in career length and productivity metrics.

To explain the professor’s audience, we used the first PCA axis as a proxy for productivity (Fig. 4b). As expected, the professor’s audience increased with productivity for both equally plausible models (Table 1c). However, male professors still had, on average, an audience 1.4 times larger than that of female professors, regardless of the productivity index (Fig. 4c). The marginal R2 of the best-fitted model was 0.28.

Gender differences in topics of research presentation

The frequencies of the most used words by male and female speakers were highly correlated (all data rp = 0.87; professors rp = 0.66), indicating that there is no clear distinction between the words used by male and female speakers in their titles and abstracts (Fig. 5, all speakers, Fig. S5, only professors). Moreover, we found no difference in topics between male and female talks in general (Chi-square = 0.28, df = 1, p-value = 0.59, Fig. S6), nor for professors (Chi-square = 0.50, df = 1, p-value = 0.48, Fig. S7).

Fig. 5: Gender differences in word usage from seminar titles and abstracts.
figure 5

Frequency plot of the most used words in the titles and abstracts of the seminars given by female (y-axis) and male (x-axis) speakers. Both axes are on a logarithm 10 scale. The color scale indicates the absolute percentage differences between male and female speakers. Words with the exact same frequency were randomly assigned to display. The dashed line indicates the slope of 1; words closer to it have similar frequencies in both sets of texts. The Pearson correlation between word frequencies was 0.87 for all talks (this plot) and 0.66 for professors only (Figure S5).

Full size image

Discussion

Our results revealed a smaller audience for women professors’ talks, suggesting a persistent lower visibility and recognition of women in an academic seminar series. Although the affirmative actions successfully increased the representation of female speakers across all academic levels as expected, they did not produce a proportional increase in the recognition of women speakers (estimated through changes in audience size). The fact that female professors attract smaller audiences, even when presenting on similar topics and having comparable productivity to male professors, suggests that there may be underlying biases or cultural factors at play that we can partially attribute to the gender-science stereotype that is pervasive in both academic and non-academic communities.

We found an underrepresentation of women giving talks, especially at higher academic levels. However, our results cannot distinguish between two interconnected but distinct dimensions of gender inequity in academia. First, the gender imbalance within the academic community, characterized by a small proportion of female academics, would consequently result in a small proportion of female speakers18,22, a well-known phenomenon in science21,30,31. We had some evidence of this effect when analyzing the subset of talks from the Graduate program and comparing it with the population's gender rates. Second, there may be a gender bias in the proportion of female speakers despite the gender balance in the academic community; that is, women give disproportionately fewer talks than men in relation to their representation in the academic community. Although affirmative actions can successfully increase female presence in academic spaces (our study, ref. 18), the second dimension raises the question of whether simply having more women in academia will be sufficient to close all representation gaps. Nevertheless, our findings support the idea that tackling numerical imbalances is only part of the broader challenge32.

To the best of our knowledge, this is the first decadal-scale study evaluating audience gender bias in a seminar series covering themes in Ecology, Evolution, and Conservation. Studies from different disciplines found contrasting results. For example, the audience size for female speakers was smaller in Philosophy33, similar in Biology and Psychology33, and higher in Economics34. However, unlike what we did, these studies did not investigate further reasons for the observed differences. Nevertheless, our study complements the findings of many other studies on gender bias in seminar and conference talks (e.g., refs. 3, 35, 36), showing that the culture of seminars is not gender-neutral and the audience is not blind to gender34. Women speakers are usually treated differently, receiving more questions in general35 (but see ref. 36) and even harsher and more patronizing questions34. It seems unlikely that the fact that female speakers attracted smaller audiences could reflect any explicit decision by seminar attendees to treat women differently. Instead, our results may indicate a systemic bias favoring male scientists24,25. In this regard, the male-scientist stereotype25,27, rooted in our male-dominated culture37 and especially stronger for college-educated people25, provides the best hypothesis to explain the academic’s willingness to attend a seminar based on the speaker’s gender. Our study presents another layer of evidence of how gender-biased stereotypes still influence the visibility and recognition of women in science.

Seminars and talks are a way for academics to get feedback, disseminate their work, and expand their professional networks3,36. Similar to what happens in many other instances, the academic community’s gender bias in attending talks given by women may decrease the visibility of research carried out by them, potentially impacting professional development and restricting the reach of the research. In the long run, smaller visibility and recognition of women in science perpetuates the gender productivity gap18 if they do not force women to evaluate whether they have chosen the right career34. Therefore, it is utterly important to address the underlying cultural and systemic factors that may be contributing to the gender bias in academic speaking opportunities and audience attendance. Our results highlight the need for continued efforts to promote gender diversity and to challenge gender stereotypes at all levels of academia, while at the same time providing support and resources to women academics to succeed in their careers.

On the one hand, we found that the problem of gender bias in the audience of female speakers seems harder to address with the most common affirmative actions toward representativeness38,39, in our case, those supporting and encouraging female speakers. On the other hand, we found that even simple changes in how committees motivate women to participate were successful in the short term. This highlights the importance of communities taking action to promote equal opportunities for women in science, regardless of its form38,40. We argue that since female scientists provide positive role models for women37, attending seminars presented by a woman not only increases the scientist’s visibility but may help reduce the implicit stereotype that science is masculine in the culture-at-large37. Although this positive feedback may seem hard and slow to achieve, it is crucial to increase awareness of the commonly ignored biases26. Addressing gender disparities in scientific events demands a more comprehensive and sustained approach.

While our study provides valuable insights into gender bias in academic seminars, it has limitations, such as focusing on a specific seminar series at one institution, the indirect nature of the affirmative actions implemented, and its timeframe. Moreover, a two-year range (after affirmative actions) might be too short to assess any indirect effects of affirmative actions, focusing on women’s representation in the audience. Our findings, however, provide a starting point to ignite discussions and more studies. The patterns we show point to the importance of rethinking how recognition is distributed in academic spaces4,5,18, in which future studies could look into whether less hierarchical and more collaborative seminar formats make a difference in how speakers are perceived. Future research could also expand the scope to encompass a broader range of institutions and disciplines, shedding light on whether the phenomenon of a smaller audience for female academics is widespread or specific to some disciplines in science. Exploring the intersectionality of gender with other factors such as race, ethnicity, and geographic origin is also necessary to address ways to improve diversity in academia36,41,42. Since our study is observational, we also encourage experimental approaches, such as Bertrand & Mullainathan43 for racial discrimination in the labor market and Moss-Racusin et al.44 for gender discrimination in academic science. Future experimental studies could, for instance, assess the willingness to attend talks depending on the features of the speaker. By addressing these gaps, academia can continue to work towards creating a more equitable and inclusive scientific community where all voices are valued and represented.

Many different levels of affirmative actions to promote community engagement and to support inclusive, socially aware, and diverse sciences26,41 are necessary to speed up the time to achieve equity and ban the skewed societal tendency to perceive scientists as an elder white man25,27. For instance, our institute organized a webinar with experts in social research to explore stereotypes, visibility, and recognition in light of our findings. We invited our community to reflect on why we put more effort into attending certain talks and not others, and to pay attention to whether there may be any unnoticed bias regarding the characteristics of the speaker in this decision. We, as academics, should be able to ask ourselves the following question: If the same seminar were given by a prestigious male professor, would I attend?

Methods

Seminar series in Ecology

The EcoEncontros is a seminar series of weekly talks at the Ecology Graduate Program of the University of São Paulo (PPGE-USP), Brazil. EcoEncontros began in 2008 and is organized by a committee composed mainly of graduate students (master’s and doctorate), in which females comprised around 70% of the organizing committee members until 2019. The committee primarily operates with open calls for volunteer speakers. In the seminars, speakers present their research at any stage of development: as a project, preliminary results, published papers, or any other topics of interest. Although it is a graduate program seminar series, almost 20% of the speakers between 2008 and 2019 were affiliated with foreign institutions.

Affirmative action can take various forms to promote equal opportunities for women in science38,40. In 2018, the EcoEncontros organizing committee became aware of gender imbalance in their seminar talks. Hence, it began pursuing ways to improve it in response to ongoing discussions about gender disparity in Science. However, these efforts aimed to preserve the seminars’ decentralized, horizontal, and voluntary nature, which relies on open calls for volunteer speakers rather than direct invitations. The initiatives (henceforth affirmative actions) aimed to create a more inclusive environment and focused on reinforcing calls for women to encourage greater female participation and engagement. Ultimately, when multiple volunteers expressed interest in presenting a seminar on a given date, preference was given to women. However, if no women volunteered, the slot was assigned to a male volunteer to ensure continuity in the schedule.

Data collection

We retrieved recorded information from all talks between 2008 and 2019 from the EcoEncontros committee attendance list archives (N = 344 talks). We retrieved data about the speaker (gender, academic level, and affiliation) and the seminar (date, title, abstract, and audience size). We inferred the speaker’s gender by name and photo (always present on the seminars’ posters). Even though we are aware that the binary classification underrepresents gender diversity and may not reflect the self-declared gender of the speaker, we believe that any possible bias by the audience in attending the talks is also led by the same information.

We classified the speaker’s academic level into three categories: student (bachelor’s, master’s, or doctoral degrees), postdoctoral researcher, and professor (assistant, associate, full, or lecturer). Senior researchers at non-university scientific institutions were also included in the professor category. We assessed audience size based on the attendance list of the seminar, where all attendees signed their names and affiliations. We excluded special seminars such as round tables and talks unrelated to the speaker’s research, totaling 327 talks for the analyses. We classified talks in terms of whether they were presented before or after the start of the organizing committee’s affirmative actions (2018): 256 talks (78%) were given before and 71(22%) after.

Data analyses

Female speakers across academic levels

To investigate the representation of female speakers across academic levels and the effect of affirmative actions, we modeled the proportion of female speakers as a function of their academic level and whether the talk occurred before or after affirmative actions. We excluded talks from non-academic professionals, totaling 320 talks used in this analysis.

We used generalized linear mixed-effects models with a Binomial distribution (response variable: 0 for male; 1 for female) and set up models based on the combination of academic level and before-after affirmative actions (Table 1a). We included the year of the talk as a random intercept to account for differences in the proportion of female speakers through the years. We used model selection based on the Akaike Information Criterion (AIC) to infer the models that best fit our data (lower AIC). We used the criterion of equality, plausible models for those with a difference in AIC lower than 2.

Additionally, to differentiate gender bias in talks from the possible effect of gender imbalance in the Graduate Program community (PPGE), we performed a similar analysis with a subset of data for speakers from the PPGE (136 talks, 44% of the original dataset). The proportion of female academics in the PPGE community was calculated for each academic level and year (Fig. S1) and used as a predictor variable in all competing models to represent the speaker’s pool. That is, for each talk, this variable was the proportion of female academics in the program according to the year of the talk and the academic level of the speaker. Competing models were set up based on the combination of academic level and affirmative actions in additive models (Table S1). This way, we evaluate if the proportion of female speakers follows the gender ratio of the PPGE community or if it is more or less biased towards male speakers at the different academic levels, as well as whether these proportions changed before and after affirmative actions.

Speaker gender differences in seminar audiences and affirmative action effects

To evaluate whether audience size depends on the speaker’s gender, academic level, and the effects of affirmative actions, we modeled audience (number of attendees) as a function of the speaker’s gender, academic level, and whether the talk occurred before or after the affirmative actions. We excluded talks from non-academic professionals and seminars when more than one speaker presented on the same day, totaling 298 talks for this analysis (see Table S2 for the descriptive summary). Similarly to the previous analysis, we modeled the year as a random intercept to account for possible differences in audience through time. Given the considerable variation in the audience (ranging from 4 to 101), we used generalized linear models with the Negative binomial distribution. We set up models using the same procedure as previously explained (Table 1b).

To investigate whether gender differences in the audience of professors reflected differences in the speaker’s career length and productivity, we collected information on the professor’s productivity, career length, and institutional prestige rank. We collected the following information on each professor’s Google Scholar profile: (1) career length, measured as the number of years from the first cited publication until the year of the talk; (2) i10-index, which measures the number of papers with at least ten citations; (3) H-index, which counts the number of papers with at least the same number of citations; (4) total number of citations; (5) cumulative number of citations until the year of the talk; (6) citations of the most cited paper. To measure the professor’s institution’s rank, we used two Nature Indices45: count and share. A count of one is to an institution or country if one or more authors of the research article are from that institution or country, regardless of how many co-authors are from outside that institution or country45. A fractional count (also called “share”) considers the percentage of authors from that institution and the number of affiliated institutions per article. We performed a PCA with all metrics and used the first axis as the predictor variable for the productivity index. We analyzed 87 professors’ talks since we could not get productivity information for nine professors.

Gender differences in seminar topics

To investigate possible gender differences in the topics of the talks, which could explain part of the gender differences in the previous questions, we performed a text analysis with the titles and abstracts of the talks. We recovered talk titles from 320 talks (140 for females, 180 for males) and abstracts from 234 talks (99 for females, 135 for males). Titles and abstracts written in Portuguese or Spanish were translated into English. We compared the frequency of words used by male and female speakers using Pearson correlation. Given the small sample size for text analysis, we did not compare it by academic level. However, we also analyzed the data separately for professors, comprising 96 titles (24 for females, 72 for males) and 77 abstracts (20 for females, 57 for males).

To investigate the differences in research topics of talks given by male and female speakers, we performed a topic modeling analysis, an unsupervised machine learning model that identifies groups of similar words (i.e., topics) within a body of text. We used Latent Dirichlet Allocation (LDA), following Silge & Robinson46, which treats each document (abstracts and titles of the talks) as a mixture of topics and each topic as a mixture of words. We compared LDA models with different numbers of topics (k = 2, 3, 4, 5, 10, 20) using AIC model selection. After classifying the talks within topics, we compared the frequency of topics between male and female speakers with a Chi-squared test.

All data analysis was performed in R (version 4.3, R Core Team, 202247), using the main packages: glmmTMB48, DHARMa49, bbmle50, performance51, ggeffects52 for modeling; tidytext53, topicmodels54, tm55, and quanteda56 for text analysis. The complete list of packages, together with all code and data, is openly available57 on the Zenodo repository (https://doi.org/10.5281/zenodo.11237445).