Polarization of public opinions on feminism in China

Gender equality, as one key element of the sustainable development goals (SDGs), is crucial for achieving social justice, economic prosperity, and a fair and inclusive society for women and girls (Sachs et al. 2019). To achieve this goal, feminist movements have recently harnessed the powerful tool of social media to fight against gender inequality (Arruzza et al. 2023; Delmar, 2018). A notable case is the #MeToo movement, which gained global attention through social media, contributing to reducing gender-based discrimination and violence (Hillstrom, 2018). However, recent years have also witnessed a surge in conservatism and patriarchy, the advocates of which criticize feminists for their pursuit of excessive equality (Flood et al. 2021; Yang et al. 2023). While these intense disputes over equality between men and women are seen globally, they have become a particularly complex issue in today’s China. For example, a Chinese comedian’s recent remarks on feminism have sparked backlash on social media and triggered intense debates, involving a variety of radicals, feminists, conservatives, and patriarchists (Wang, 2024). Similar examples are not rare, as seen in phenomena like Ding Zhen’s popularity (Xing et al. 2023) and “Pastoral Feminism”, a derogatory term of Chinese feminism (Li, 2018). Overall, all these examples (Flood et al. 2021; Hillstrom, 2018; Li, 2018; Wang, 2024; Xing et al. 2023; Yang et al. 2023) hint at an alarming polarization of opinions on feminism and women’s equality in the general public.

Indeed, the complexity underlying the intense disputes over equality between men and women in China has deep roots. Chinese opinions on women’s equality are shaped not only by traditional Chinese culture and values (Croll, 2011) but also by domestic political dynamics (Barlow, 2004; Zheng, 2005). Specifically, Chinese culture, particularly Confucianism, along with traditional values like collectivism, have bred the conservative camp (Croll, 2011). For example, Confucianism embodies patriarchal values, asserting that women have a lower natural status than men and that women should be subordinate to men both in the family and in society (Legge et al. 2022). With modernity and globalization, feminism has gradually grown in China, confronting such traditional patriarchal values (Hershatter, 2007; Leung, 2003). Emerging during the Mao era and flourishing after the Opening and Reform in 1978, Chinese feminism increasingly incorporated ideas from Western feminism, gradually shifting from Marxist and socialist feminism toward liberal feminism (Mohajan, 2022; Vogel, 2013; Zheng and Zhang, 2010). Consequently, in the context of China, the opposites of conservative traditions and feminist values are deeply intertwined, creating a unique complex environment that could potentially sustain the polarization of public opinions. However, we know little about whether this polarization exists in this real-world context and, if so, what its characteristics are.

Polarization has become a widespread global trend, permeating into aspects of our society (Bail et al. 2018; Bakshy et al. 2015; Cinelli et al. 2021; Falkenberg et al. 2022; Flamino et al. 2023; Lorenz-Spreen et al. 2023; Murphy et al. 2021; Nyhan et al. 2023; Sunstein, 1999, 2009, 2018). Prior studies have mainly centered on traditional social media platforms, e.g., Facebook and Twitter (Bail et al. 2018; Bakshy et al. 2015; Cinelli et al. 2021; Flamino et al. 2023; González-Bailón et al. 2023; Nyhan et al. 2023). The studied platforms have built their information ecosystems mainly around social networks, which are organized into echo chambers that foster the formation of polarization (Bakshy et al. 2015; Cinelli et al. 2021; Sunstein, 2006). However, the rapid growth of AI has transformed the landscape of online social media, gradually replacing traditional platforms with emerging short video platforms (Dean, 2023; Gao et al. 2023b; Piao et al. 2023; Wang and Wu, 2021). For example, in 2023, TikTok reported over 50 million daily active users in the U.S., with adults averaging 33 minutes daily on the platform (Dean, 2023). Moreover, the number of short video users in China has reached 1.05 billion, accounting for 96.4% of all internet users (CNNIC, 2024). These short video platforms have shifted from reliance on social networks to AI-driven recommendation algorithms for disseminating information. This shift means that people are no longer directly connected but gathered through algorithmic recommendations, challenging the very existence of echo chambers, i.e., the traditionally recognized foundation of polarization. Given the importance of these emerging AI-driven social media platforms, understanding whether and how polarization emerges on these platforms is a crucial question.

To investigate whether and how public opinions are polarized on feminism and women’s equality in China, particularly under an AI-dependent background, we conduct the first large-scale empirical study on one of China’s largest short video platforms. We collect a fine-grained dataset from this platform, involving over 62 thousand active users and their 67 million behavioral records over two years. Leveraging the power of the fine-tuned large language model BERT (Kenton and Toutanova, 2019), we accurately classify opinions on feminism into five categories, including radical feminism, liberal feminism, neutral, conservatism, and patriarchy. We observe a severe polarization of opinions on feminism, where the majority (88.7%) are asymmetrically polarized in two opposing camps—54.3% in the conservative camp and 34.4% in the liberal camp. Over the two years, this observed severe polarization has significantly intensified over two years, with the largest increase of 30.8% in polarized users. We reveal that the AI-driven recommendation algorithm, which is the core of short video platforms, contributes to two primary drivers responsible for polarization on feminism: (i) echo chambers, where the algorithmic aggregation of like-minded users reinforces their shared similar opinions, and (ii) battlefields, where users’ confrontations strengthen their pre-existing opinions. More interestingly, these two polarization drivers are asymmetrically applied to two camps. Specifically, users in the conservative camp are more likely to be recommended to and engage in echo chambers while those in the liberal camp are more inclined to battlefields. Our findings not only reveal polarization on feminism on the short video platform but also offer practical implications for reducing polarization on increasingly AI-dependent social media.

The remainder of this work is organized as follows. We first review prior works about in the “Literature review” section. We then introduce the studied platform and dataset, the developed spectrum of opinions on feminism and women’s equality, and the built opinion classifier in the “Methods” section. In the “Results” section, we present and interpret our main results about polarization of public opinions on feminism and women’s equality. Based on the results, we conclude this work by summarizing key findings, discussing their implications, highlighting limitations, and suggesting directions for future research in the “Discussion” section.

Feminism: concept, history, and categories

Feminism is generally defined as a social and political movement and ideologies pursuing gender equality (Delaney, 2005). It aims at “reducing gender inequality” (Walby, 2011), and achieving gender equality in social, political, and economic dimensions (Hundleby, 2011; Ritzer and Ryan, 2011). It is rooted in historical social efforts to counter a patriarchal society, fight against the oppression and exploitation of women, reconstruct gender roles, and advocate for women’s interests (Mohajan, 2022).

The history of the feminist movement can be roughly divided into three waves of development (Humm, 2021; Krolokke and Sorensen, 2006). The first wave was from the mid-19th century to the mid-20th century. During this period, feminists criticized the legal systems of capitalist countries that segregated women from their rights to education, employment, and suffrage, and strove for women’s universal suffrage (Freedman, 2007). Next, the second wave of development was from the 1960s to the 1970s, during which feminist thought and the feminist movement gained unprecedented momentum, both in terms of depth and breadth (Whelehan, 1995). Research on feminism in this period was dominated by radical feminism, demanding the breaking down of male-centered ideology and advocating a feminine perspective on social life. In addition, the feminist movement of this period is different from the first moderate feminism movement: it required not only changes in the legal system, but also a revolution in all aspects of women’s political participation, economic independence, and the breaking down of traditional values of male supremacy in culture. Then, the third wave started in the 1980s and 1990s with a focus on individuality and diversity, during which more interdisciplinary scholars studied feminism, and the views became more and more diverse (Krolokke and Sorensen, 2006). However, the form of feminism is not only a movement and ideology, it is also changing over time: feminism is more a project or program than a movement in some countries, including different stakeholders and practices (Walby, 2011). Feminism becomes an approach and tool to counter a patriarchal society (Ratna, 2004).

As feminism proceeded and expanded globally, feminism has developed to characterize oppression and exploitation faced by women in different times or different social contexts (Brunell and Burkett, 2019) and therefore diverged into different subgroups. These subgroups include radical feminism, liberal feminism, Marxist feminism, socialist feminism, conservative feminism, ecofeminism, cultural feminism, black feminism, and postmodern feminism (Bart, 1991; Jaggar, 1983; Jaggar and Rothenberg, 1993). As of now, feminism has become “a diverse, rival and often opposing collection of social theories, political movements, and moral philosophies” (Mohajan, 2022). The ultimate goal of every feminist branch is, however, to achieve gender equality (Tong, 2009). Most of the branches of feminism advocate eliminating misconceptions, gender inequality, restrictions (e.g., economic, political, and social), and oppression (Bressler, 2007; Bryson, 2007).

Feminism and women’s equality in China

Chinese opinions on feminism are shaped by traditional Chinese culture (Croll, 2011) and are also deeply intertwined with domestic political dynamics (Pan, 2001; Zheng, 2005). Traditional Chinese culture, particularly Confucianism and Taoism, along with traditional values like collectivism, have bred conservative opinions (Croll, 2011). For example, Confucianism embodies patriarchal values, asserting that women have a lower natural status than men and that women should be subordinate to men both in the family and in society (Legge et al. 2022). Moreover, Confucianism also advocates a division of labor, with men managing external affairs and women focusing on household duties (Legge et al. 2022). This ideology limits women’s social and political participation (Legge et al. 2022). Furthermore, some patriarchists emphasize traditional gender roles of women, expecting women to be “righteous” and to cultivate virtues and skills in domestic tasks (Legge et al. 2022).

On the other hand, feminism in China started late and emerged in the early 20th century with the modernization of China and the introduction of Western ideas. Initially, feminism was defined as obtaining women’s rights (Croll, 2011)—intellectual women in the middle and upper class with new education were the first to take up the fight for the rights and interests of the women’s movement (Pan, 2001). Through the foundation of journals and the organization of group activities, they focused on individual rights, such as the abolition of feudal rituals and women’s freedom of marriage, educational rights, and political participation (Pan, 2001). During the May Fourth Movement, Chinese feminists began to criticize traditional gender roles and family concepts and demanded women’s emancipation. However, along with the development of the anti-imperialist and anti-colonial national independence movement, the idea of women’s liberation was soon overshadowed by nationalist thinking. The concept of “women” for human liberation was gradually replaced by the concept of “women” for social liberation—women stood together with men against the feudal social system (Song, 2013).

Then in the 1920s, feminism in China developed and referred to the women’s movement, intertwining with wider revolutionary movements, it turned to the transformation of fundamental social structures The term “feminism” in Chinese discourse has been reinterpreted over time (Croll, 2011). It attempted to transcend and overcome the limitations of liberal feminism, which emphasized only individual rights, and to combine the women’s movement with the overall transformation of society (Song, 2013). The Chinese Communist Party organized numerous women’s labor movements and related public discourse debates and propaganda, resulting in the increasing visibility of the working women’s movement (Federation, 2011). From the 1930s onward, Chinese feminism was gradually integrated into the political ideology of the Communist Party.

Next, feminism in the Mao era was also known as the women’s liberation movement (Federation, 2011). During this period, the Party, the state, and the government became the main actors in promoting the women’s liberation movement, and therefore, it was also called state feminism (Zheng, 2005). State feminism absorbed ideas from Marxist and socialist feminism, acknowledging that the root of women’s oppression was private ownership and part of class oppression (Mohajan, 2022; Vogel, 2013); claimed women’s participation in socialist and communist revolutions was the only correct path to complete emancipation; and the socialist state implemented the basic state policy of equality between men and women (Zheng, 2005). The new China’s legal policies of gender equality, equal pay for equal work, and marital autonomy narrow the alienation of gender roles for women, who at this time is not a dependent, subordinate, or auxiliary role to male roles such as father, husband, or son, but is the embodiment of a collective-national group. However, the egalitarianism of men and women in the public sphere of this period did not unfold in the private sphere, meaning that the patriarchy in family relations was never shaken (Zheng, 2005).

Finally, after the Opening and Reform in 1978, the relationship between the feminist movement and the state underwent a profound change when mainland China began to reform its market economy and open to the outside world (Li, 1995). Feminism in this period is called market feminism, and the idea of a “sexualized person” is proposed to break the degendered women’s studies of the Mao era (Li, 2000). Additionally, The UN Fourth World Conference on Women, held in Beijing in 1995, was an important historical turn in the development of the Chinese women’s movement. This conference directly brought about the legitimization of non-governmental organizations (NGOs) in China, and concepts such as “gender equality”, “female empowerment”, and “gender mainstreaming” provided Chinese feminists with new discursive resources (Federation, 2011). Moreover, numerous theoretical works from American feminism were translated and published during the 1980s and 1990s, enriching the theoretical discussion of feminism in China.

To conclude, the feminist movement in China has gone through a process of development from personal liberation to political movements to theoretical construction. Except for the consensus of pursuing gender equality and women’s rights, feminism in China exhibits complex characteristics. Chinese feminism has been deeply influenced by Chinese culture and traditional values (Croll, 2011). Due to the differences in history and social structures between China and the Western world, feminism in China highlights the harmonious coexistence based on collectivism while Western feminism leans more towards individualistic and liberal values.

Platform and dataset

With the boost of AI-driven recommendation algorithms, short video platforms have transformed the landscape of online social media (Dean, 2023; Gao et al. 2023b; Piao et al. 2023; Wang and Wu, 2021), challenging the dominant position of traditional social media giants like Facebook and Twitter (Dean, 2023; Gao et al. 2023b; Wang and Wu, 2021). Therefore, we focus our study on the emerging form of social media, short video platforms. In particular, we launch our study on one of China’s largest short video platforms. The platform, as China’s first short video platform, was established in 2012 and gained significant market traction after 2015. Data show that in the fourth quarter of 2023, the daily active user (DAU) had reached 382.5 million, up 4.5% year-on-year, and the average monthly active user (MAU) grew 9.4% year-on-year to 700.4 million. On average, these active users spent 124.5 min on the platform each day. The widespread adoption and great popularity make it a valuable case for understanding public opinions in today’s China.

We collect a large-scale fine-grained dataset from one of the largest short video platforms in China. The dataset involves over 29.6 billion videos from 2018 to the present as well as 62K active users’ 67 million behavioral records over two years. Specifically, we first filter over 15.7 million feminism-related videos from the overall pool of 29.6 billion videos, using the keyword search method. All the keywords are curated by a team of four researchers from research backgrounds in feminism, gender studies, and public policy (see details in Section S2 in SI). The list, covering 50 keywords, is shown in SI Fig. S2. It is worth mentioning that despite careful curation, some terms could inevitably introduce some noisy data. Therefore, to minimize the impact of such noise on opinion measurement, we have developed an accurate classifier to filter out video comments irrelevant to feminism. This method naturally excludes videos with no feminism-related comments, which are most likely noise, from the analysis. We then filter users who have posted more than 20 valid comments on these videos from April 1, 2021 to April 1, 2023, totaling 62,224 users. We collect their recommendation and comment records over the two years. Each recommendation record contains the user identifier (ID), the recommended video ID, and the timestamp. Each comment record contains the user ID, the video ID, the comment content, and the timestamp. Table 1 summarizes the basic statistics of the collected dataset.

As discussed above, unlike traditional social media, short video platforms depend mainly on AI-driven recommendation algorithms for information dissemination. In this way, users are no longer directly connected by their social relationships (Fig. 1a); instead, they are gathered together by recommendation algorithms (Fig. 1b). In particular, the AI-driven recommendation algorithm first recommends appropriate videos to users (Ricci et al. 2022). Given the well-known black-box nature of AI, the recommendation mechanisms that these algorithms follow remain largely intractable. However, prior observations and practices (Gao et al. 2023a; Piao et al. 2023; Ricci et al. 2022; Zhang et al. 2019) have shown that recommendation results are influenced by several factors. For the purpose of personalization, these algorithms filter out items similar to those users have previously liked, which is also known as the similarity-based matching mechanism (Gao et al. 2023a; Piao et al. 2023; Ricci et al. 2022; Zhang et al. 2019). In addition, these algorithms also consider non-personalized aspects, particularly video popularity (Ricci et al. 2022). After recommendations, users decide whether to join the video discussion by leaving comments. We regard all users who participate in the discussion of the same video as forming an algorithmic video community (Fig. 1c, video community for short). It is worth mentioning the prerequisite of joining an algorithmic video community for a user is that he/she is recommended the corresponding video by the algorithm. This suggests that discussions within the algorithmic community, including its users, comments, and opinions, are fundamentally influenced by the algorithm. In this paper, our analyses focus on this newly emerging form of communities, as they serve as the basic ground for users’ interactions and communications on AI-dependent platforms. We consider the most prevailing opinion within a video’s comments to be the dominant opinion of the corresponding video community.

a Traditional social media, where users are socially organized into communities, and their social networks serve as the channel for information dissemination. b Short video platforms, where the AI-driven recommendation algorithm propagates information and virtually links users together. c Interactions on the short video platform. The recommendation algorithm first recommends appropriate videos to users, and then users decide whether to join the discussion of the recommended videos by leaving comments. Users in the discussion of the same video are regarded in the same algorithmic video community, where they can interact and communicate with each other.

Full size image

Spectrum of opinions on feminism

A key step of this study is to build a spectrum to evaluate users’ opinions on feminism in China. Chinese opinions on feminism are deeply intertwined with domestic politics (Pan, 2001; Zheng, 2005), and are also shaped by Chinese culture and traditional values (Croll, 2011). Consequently, two main forces are confronting each other: the liberal camp versus the conservative camp (Zheng, 2005).

The liberal camp is deeply inspired by the modernity and globalization of feminism. Emerging during the Mao era and flourishing after the Opening and Reform in 1978, this camp had increasingly incorporated ideas from Western feminism and shifted from Marxist and socialist feminism to liberal feminism (Li, 1995; Mohajan, 2022; Vogel, 2013; Zheng and Zhang, 2010). The 1995 UN Fourth World Conference on Women in Beijing marked a turning point (Federation, 2011; Zheng and Zhang, 2010). Diverse Western feminist theories, including radical feminism, began to influence Chinese feminism (Zheng and Zhang, 2010). Therefore, within the liberal camp, opinions toward gender inequality are divided into moderate and radical groups. The moderate group, absorbing ideas from Marxist, socialist, and liberalism feminism, focuses on equal rights and opportunities for men and women (Mill, 1869; Wollstonecraft, 2014; Zheng, 2005). By contrast, the radical group, influenced by radical and cultural feminism, argues that women should be granted more rights than men, rather than merely seeking equality (Atkinson et al. 2000).

The conservative camp is deeply rooted in traditional Chinese culture, e.g., Confucianism and Taoism, and values like collectivism (Croll, 2011). The government also promotes the concept of “Smiling Chinese feminism” and emphasizes “gender harmony” to maintain stable social development (Spakowski, 2011). Within the conservative camp, there are both moderate and radical groups. The moderate group upholds traditional views on women’s roles, blending these with the modern concept of “gender harmony” in Chinese society (Bacchetta and Power, 2013). This group reaffirms and respects traditional roles for women, such as mothers and caregivers, while also advocating for equality and autonomy within these roles (Bacchetta and Power, 2013). The radical group in the conservative camp, referred to as patriarchy, originated from the historical and cultural traditions of feudal China. The group promotes patriarchal values that establish male superiority, promote a gendered division of labor confine women to domestic roles, and emphasize traditional virtues for women, thereby restricting their social and political participation (Legge et al. 2022).

These two opposing camps represent two polarized forces in the online discussion on feminism. We assign the liberal camp as the representatives of both radical and liberal opinions, recognizing women’s status as equal or even higher than males. This camp holds radical and liberal as its subgroups, including those labeled as radical feminism, cultural feminism, liberal feminism, and Marxist and socialist feminism. Conservative camp advocates men’s dominance and traditional gender roles, containing conservative and patriarchal subgroups. However, when analyzing users’ comments, we observe that many discussions on feminism neither align with liberal or conservative values, nor consider equal rights, responsibilities, and opportunities for men and women. Instead, they adopt neutral narratives. For example, “Men and women are the same, and there are both good people and bad people”. In this example, the comment remains descriptive but does not engage with values related to feminism. Therefore, we categorize them into the neutral camp, characterized by the neutral position on feminism. We supplement a detailed review of studies related to the development of this spectrum in SI Section S1.

Codebook and labeling

As discussed above, we develop a spectrum of opinions on feminism based on theories (Li, 1995; Mohajan, 2022; Vogel, 2013; Zheng and Zhang, 2010) and China’s background (Croll, 2011; Pan, 2001; Zheng, 2005). Based on this spectrum, we compile our primary labeling manual. It consists of 5 labels, which represent the five subgroups, i.e., radical feminism (R), liberal feminism (L), neutral (N), conservatism (C), and patriarchy (P). After pre-labeling and discussion (See SI Section S2 for details of the whole process), the team of four researchers has reached agreements and finalized the codebook as shown in Table 2.

Based on the codebook, we employ a deductive coding method to systematically label a dataset of 11,856 comments. This process involves five rounds of labeling. In each round, experts independently label an equal number of comments. Following this, a discussion is held to resolve any discrepancies and align the labeled results, ensuring consistency and reliability in the coding process. With a team of four researchers specializing in feminism, gender studies, and public policy, all comments were cross-labeled, and agreements were reached on all labels. These procedures ensure the consistency and quality of the labeled dataset.

Opinion classification

Although we form a team for labeling, evaluating the opinions of all the user comments remains challenging. Therefore, we employ natural language processing (NLP) methods to automate the labeling process. However, unlike classical NLP tasks, e.g., sentimental analysis (Wang et al. 2022), there are no available models or dictionaries for this task (Pennebaker et al. 2001). Therefore, we apply BERT techniques (Cui et al. 2021; Kenton and Toutanova, 2019), which can provide high-dimensional pre-trained representations of textual data. Moreover, considering user comments are in simplified Chinese, we adopt the Chinese BERT model that is equipped with the whole word masking technique and pre-trained with large-scale Chinese texts (Cui et al. 2021). Based on the pre-trained model, we further fine-tuned it using the manually labeled 11,856 comments (Cui et al. 2021; Wolf et al. 2020). Specifically, we split the labeled comments into the training and validation sets in an 8:2 ratio. We fine-tune the model using the training set and evaluate the model using the validation set. The model finally achieves the accuracy score of 0.72 and F1 score of 0.70 in the validation set (see detailed evaluation in SI Section S2.3).

Given online user comments are typical of their informality and noise, we first clean the comment data following several steps: (i) We remove all the special characters, which are difficult to be processed by language models. (ii) We remove duplicate punctuation and spaces. We set the minimum length for a valid comment to 10 characters to ensure semantic integrity. It is worth noting that we remove all identifiers about individuals and groups to prevent privacy breaches. (iii) We train a model to distinguish between comments related to feminism and those irrelevant ones, with an accuracy score of 0.93 and F1 score of 0.92.

Based on the fine-tuned model and the cleaned comment dataset, we obtain the predicted label and prediction probabilities for each comment. We then measure the prediction confidence by calculating the difference in prediction probabilities between the predicted and the secondary labels. For example, if a comment is predicted to be liberal feminism with a probability of 0.99, and the secondary label is radical feminism with 0.009, the confidence score is 0.981. To enhance the reliability, we filter out the comments with a confidence score lower than 0.9. After extracting opinions expressed in user comments, the next step is to infer users’ opinions. However, users’ ground-truth opinions (e.g., obtained from survey data) are mostly unavailable in studies on social media (Falkenberg et al. 2022). Therefore, following prior practices (Falkenberg et al. 2022; Flamino et al. 2023), we infer users’ opinions from their online behaviors and the content they produce. In particular, we aggregate all comments related to feminism for each user, identify their camps based on the camp they most frequently align with and then assign them to sub-groups according to their most commonly expressed opinions.

Growing attention to feminism-related issues

We begin by assessing public attention to feminism-related issues on the platform. Based on the curated keywords, we have identified 15.7 million videos related to feminism-related issues from the overall 29.6 billion videos. By sorting these videos according to their upload time, we observe a seven-fold increase in the number of new feminism-related videos from 2018 to 2023 (blue bars in Fig. 2a). Meanwhile, the percentage of such videos to the overall new videos also demonstrates a super-linear growth rate (orange dashed lines in Fig. 2a). These observations suggest that feminism-related issues are becoming increasingly prominent in the platform’s information landscape.

a Trends in feminism-related videos from 2018 to 2023, where blue bars represent the number of feminism-related videos and orange dashed lines represent the number as a percentage of the overall videos. Note that the dashed bar represents the estimated number based on time duration. b Recommendations and user engagement of feminism-related videos, where blue and orange lines represent the average numbers of recommendations and comments respectively, and shades represent the corresponding 95% confidence intervals (CIs). Feminism-related videos attract increasing recommendations and user comments.

Full size image

Along with the increase in these videos, user engagement with feminism-related issues has also risen markedly. Among the 62K users who are active in these videos, we find that their watching time and comment frequency increased by 6 times and 10 times, respectively over 2 years (orange line in Fig. 2b and SI Fig. S4a). Corresponding to the rise of user engagement, the recommendation algorithm improves the exposure of these videos (SI Fig. S4b,c), eventually leading to a five-fold increase (blue line in Fig. 2b). Overall, these results consistently suggest that feminism-related issues have drawn growing attention and sparked heated discussions on the platform.

Polarization on feminism

To explore public opinions on feminism, we extract opinions from users’ comments by leveraging the power of language models (see the “Methods” section for details). Figure 3a presents the distribution of comments across different opinions, revealing an apparent polarization pattern. In this distribution, two opposing camps, i.e., liberal and conservative camps, take up 34.8% and 51.1%, respectively. In the distribution of user opinions (Fig. 3b), the polarization pattern becomes more apparent: over 88.7% of the users hold opinions in either liberal or conservative camps. Moreover, the asymmetry between the two camps is further amplified, where the conservative camp takes up 54.3% of the users while the liberal camp only takes up 34.4%. This suggests that the conservative camp, in terms of both the number of comments and users, constitutes the majority on the platform. By dividing each camp into two sub-groups, we observe that although moderate users currently take a central role, a substantial number of extreme comments (40.3%) and users (39.0%) exist in polarization on feminism (Fig. 3). These observations lead to a natural question of how polarization on feminism becomes so severe.

a Distribution of comments across opinions on feminism. b Distribution of users across opinions on feminism. Here error bars represent 95% CIs.

Full size image

To depict the dynamics of polarization, we further measure polarization levels for each month by quantifying how opinions deviate from the center of the spectrum. Specifically, assuming an opinion distribution to be [f₋₂, f₋₁, f₀, f₁, f₂] (\(\mathop{\sum }\nolimits_{i = -2}^{2}{f}_{i}=1\)), its polarization score is computed as \({s}_{{\rm {pol}}}=\frac{\mathop{\sum }\nolimits_{i = -2}^{2}{(i)}^{2}\cdot {f}_{i}}{4}\). As shown in Fig. 4a, the level of polarization is significantly intensified in terms of both comments and users over two years (Student’s t-test t = −12.25, p ≪ 0.001, t = −7.24, p ≪ 0.001). Notably, the largest increase occurred between June 2021 and June 2022, from 0.436 to 0.513, which is equivalent to 30.8% of users becoming more polarized.

a Trend of polarization over two2 years, where the gray box illustrates how the metric is calculated. b Trend of two extreme sub-groups, i.e., radical feminism and patriarchy. c Trend of neutral group. d Trend of the liberal camp. e Trend of the conservative camp. Coupled with the intensification of polarization on feminism, extreme opinions increase while opinions regarding neutral and liberal feminism decline.

Full size image

As polarization on feminism intensifies, we find a striking increase in the two extreme sub-groups, i.e., radical feminism and patriarchy (Fig. 4b), which eventually encompass almost half of the comments (41.4%) and users (38.8%). Meanwhile, neutral opinions are rapidly eroded by the intensifying polarization between liberal and conservative camps (Fig. 4c). Over 2 years, neutral comments and users have decreased by 2.6% and 4.3%, respectively. Along with these decreases, we have also witnessed a gradual decrease in the liberal camp (Fig. 4d) but an increase in the opposing conservative camp (Fig. 4e) in recent months. Especially, users in the liberal camp have dropped from a high of 40.8% to 36.7%, while those in the conservative camp have increased to 50.0%. These observations reflect the recent resurgence of traditional values and growing challenges faced by feminist movements in today’s China society (Fincher, 2016; Ji, 2015; Wang and Chang, 2023).

Echo chambers created by the recommendation algorithm

To encounter information overload, the recommendation algorithm is essentially designed to filter similar items to what users favored in the past (Piao et al. 2023; Ricci et al. 2022). Despite its effectiveness in personalization, this could potentially gather like-minded users in similar videos and create “echo chambers”. To validate it, we examine the extent to which users are recommended to video communities dominated by similar opinions (i.e., echo chambers). As shown in Fig. 5a, we find that the algorithm recommends significantly more video communities that align with users’ pre-existing opinions than those with opposing opinions (two-sided Student’s t-test, t = −36.39, p < 0.001). This suggests that the recommendation algorithm indeed guides users into echo chambers. Facing these echo chambers delivered by the algorithm, users tend to participate in discussions within echo chambers rather than those dominated by opposing opinions (Fig. 5b): users write 70% more comments on the most similar videos compared to the most dissimilar ones on average. These observations emphasize the role of the recommendation algorithm in forming echo chambers, even in the absence of social networks.

a Relationship between the video similarity and the number of recommendations, where bars represent the average values in terms of users and error bars represent the corresponding 95% CIs. b Relationship between the video similarity and the number of comments. Users are more likely to (a) receive recommendations and (b) engage in discussions of videos dominated by similar opinions. c Relationship between fractions of videos dominated by radical feminism and users' changes in opinions. d Relationship between fractions of videos dominated by patriarchy and users' changes in opinions. Users who are excessively engaged in extreme videos are likely to adopt more extreme opinions.

Full size image

We further explore whether these algorithmic echo chambers also contribute to the increased polarization. Specifically, we take users with unchanged opinions as the control group (L → L and C → C bars, where → denotes the direction of change in opinion) and compare them with those who have shifted toward more extreme opinions after 2 years (L → R and C → P bars). As shown in Fig. 5c and d, we observe that users adopting more extreme opinions have engaged in a significantly higher number of videos dominated by extreme opinions within their respective camp (two-sided Student’s t-tests t = 16.41, p ≪ 0.001, t = 27.92, p ≪ 0.001). Even initially neutral users are not irrelevant to these echo chambers of extreme opinions (N → R and N → P bars). After two years of excessive interactions with extreme opinions, they eventually transform into one of two extreme sub-groups, i.e., radical feminism or patriarchy (N → N versus N → R and N → N versus N → P bars, t = 21.58, p ≪ 0.001 and t = 14.11, p ≪ 0.001). Overall, we find that the increased polarization of users’ opinions is positively correlated with extensive interactions with echo chambers dominated by extreme opinions.

Given the potential consequences of algorithmic echo chambers, we examine the severity of echo chambers on the platform. We first assess the degree to which users are recommended opinion-aligning information. In particular, we compute the fraction of recommendations from the same camp and depict its distribution among users in Fig. 6a, b. We find that there are very few users who are trapped in severe echo chambers, for example, only 0.75% of users receive over 75% of the recommendations from their respective camp. This finding suggests that, despite the existence of echo chambers, users still have access to opposing opinions. Furthermore, the distribution displays a unique bi-modal pattern, featuring one peak around 0.3 and the other around 0.5. By dividing the overall population into two opposing camps, we discover that these two peaks correspond to the liberal and conservative camps, respectively. This implies that the recommendation algorithm creates asymmetric echo chambers, where users in the conservative camp receive a higher proportion of recommendations in line with their pre-existing opinions, but those within the liberal camp can still access a wide range of communities, users, and opinions (two-sided Student’s t-test, t = −186.76, p ≪ 0.001). To further understand how users interact within their camp, we assume two users commenting in the same video community represent a single interaction and depict the distribution of same-camp interactions in Fig. 6c, d. The bi-modal distribution suggests that users in the conservative camp are more inclined to interact with like-minded peers than those in the liberal camp (two-sided Student’s t-test, t = −500.79, p ≪ 0.001). This observation further highlights the asymmetry of echo chambers between the two camps, with the conservative camp more likely to engage in echo chambers than the liberal camp.

a, b Distribution of recommendations from the same camp among users, where a shows the overall distribution, and b shows the decomposed distributions based on users' camps. c, d Distribution of interactions from the same camp among users, where c shows the overall distribution, and d shows the decomposed distributions based on users' camps. Users in the conservative camp are more likely to have in-camp recommendations and interactions than those in the liberal camp.

Full size image

Cross-camp interactions and algorithmic battlefields

As discussed above, there are a substantial number of cross-camp interactions, raising a question of their role in polarization on feminism. We begin by measuring the opinion heterogeneity of video communities using the variance of opinion distribution in each community, where a lower heterogeneity value indicates more in-camp interactions in the discussion while a higher value indicates more cross-camp interactions. We select the top 25% most opinion-heterogeneous video communities and separate them according to their dominant opinions (Fig. 7a–d). Specifically, in the communities dominated by extreme opinions (Fig. 7a, d), all subgroups participate in the discussion, expressing either support for or opposition against the prevailing opinions. On the other hand, for communities where moderate opinions take up the majority (Fig. 7b, c), two opposing groups emerge: liberal feminism versus patriarchy, and radical feminism versus conservatism. This reflects the inherent value conflicts in these two groups. For example, in the realm of women’s status, liberal feminism advocates for women’s equality, while patriarchy advocates for male supremacy. In terms of gender roles, conservatism upholds traditional gender roles, while radical feminism challenges the very concept of gender roles.

a–d Opinion distributions in battlefields, where solid bars represent average fractions of the dominant opinions, hatched bars represent those of the other opinions, and error bars represent the corresponding 95% CIs. Note that these video communities are the top 25% most opinion-heterogeneous ones. e, f Relationship between the opinion heterogeneity of commented video communities and users' changes in opinions. Users who are excessively engaged in opinion-heterogeneous video communities are likely to adopt more extreme opinions. g Relationship between the opinion heterogeneity of video communities and the number of comments. h Relationship between the opinion heterogeneity of video communities and the number of recommendations.

Full size image

However, cross-camp interactions do not correlate with the anticipated reduction in polarization on feminism; rather, they are correlated with the exacerbated polarization, serving as “battlefields” (Fig. 7e, f). We find that users who have previously engaged in opinion-heterogeneous communities, characterized by more frequent cross-camp interactions, tend to adopt more extreme opinions compared to those who avoid such interactions (L → R versus L → L, two-sided Student’s t-test t = 5.76, p ≪ 0.001; C → P versus C → C, t = 21.17, p ≪ 0.001). Even initially neutral users are not immune to this effect, as they exhibit a tendency to shift towards extreme positions after extensive exposure to cross-camp interactions (N → R versus N → N, two-sided Student’s t-test t = 14.74, p ≪ 0.001; N → P versus N → N, t = 15.19, p ≪ 0.001). To sum up, we find that the increased polarization of feminism is also positively correlated with users’ excessive interactions with battlefields.

Remarkably, these opinion-heterogeneous video communities tend to gain great popularity among the population (Fig. 7g). In particular, comparing the top 25% most heterogeneous videos with the bottom 25%, we observe that the top videos attract 3.2 times more comments than the bottom counterparts (two-sided Student’s t-test, t = 33.99, p ≪ 0.001). Given their great popularity, the recommendation algorithm naturally directs more exposure to these videos (Fig. 7h). It is worth noting that while the recommendation algorithm is designed to filter similar items for personalization, it also takes non-personalized factors into account, particularly video popularity in its recommendations (Ricci et al. 2022). Moreover, we find that like echo chambers, these battlefields are also asymmetrically applied by the algorithm, where the liberal camp is more likely to engage in the battlefields than the conservative camp (SI Fig. S6).

Polarization on feminism stands as a prominent and intricate phenomenon in contemporary society (Barroso, 2020), characterized by extreme divisions of opinions within discussions regarding women equality (Faludi, 2009; Fudge and Cossman, 2002; Mansbridge and Shames, 2008). In recent years, a rise in conservative political forces and their civil society allies have posed challenges to women’s rights and domestic protections, resulting in the emergence of conservatism and a rapid growth in anti-feminist movements or activities globally (Sanders and Jenkins, 2022). In today’s China, this polarization has become a particularly complex issue: it is shaped not only by traditional Chinese culture and values (Croll, 2011), but also by domestic political dynamics (Barlow, 2004; Zheng, 2005). We offer, to the best of our knowledge, the first large-scale study of polarization on feminism in China, allowing us to illustrate how the general public perceives and engages in discussions on feminism-related issues in today’s China. In this study, we first build a spectrum of opinions on feminism based on feminist theories and Chinese cultural background. Then we accurately measure public opinions by leveraging the developed spectrum and the fine-tuned large language model. We observe the increased polarization of public opinions on feminism in China (as suggested by Figs. 3 and 4) and imply the need for policymakers to develop regulations aimed at reducing polarization and promoting women’s equality.

Polarization is a widely recognized and serious societal issue. It permeates aspects of our society (Bail et al. 2018; Falkenberg et al. 2022; Jost et al. 2022; Lorenz-Spreen et al. 2023; Santos et al. 2021), amplifies ideological extremes (Hetherington, 2009; Jost et al. 2022), exacerbates hatred (Dimant, 2024; Jost et al. 2022), and threatens democracy (Arbatli and Rosenberg, 2021; Lorenz-Spreen et al. 2023). In the specific context of women’s equality, this polarization has fostered misogyny, for example, remarks like “Bearing children is a sacred mission and divinely ordained duty for women”, which undermines long-term efforts of feminist movements in China and obstructs progress toward achieving the Sustainable Development Goals. Therefore, given the consequence of polarization, we need to reduce polarization, which is also the common goal of prior studies on polarization (Bail et al. 2018; Balietti et al. 2021; Barberá, 2020; Chen et al. 2024; Interian et al. 2023; Jost et al. 2022; Lorenz-Spreen et al. 2023; Santos et al. 2021).

Moreover, the role of social media in polarization has been a central focus in recent decades, drawing substantial attention from various fields, including sociology, political science, social psychology, and information science (Bail et al. 2018; Bakshy et al. 2015; Cinelli et al. 2021; Falkenberg et al. 2022; Flamino et al. 2023; Lorenz-Spreen et al. 2023, ?; Murphy et al. 2021; Nyhan et al. 2023). However, it is crucial to note that that social media, serving as the ground for polarization, are rapidly transforming: short video platforms emerge and traditional social media giants incorporate short videos to maintain their influences. This transformation highlights the role of recommendation algorithms in managing the information accessible to users and substantially affects the dynamics of their opinions. Our pioneering study, based on over 62,000 users and their 67 million records over the entire two years, uncovers how the short video platform, a typical new form of social media in the era of AI, offers grounds for polarization. We find the increased polarization of users’ opinions is positively correlated with extensive interactions with the recommended videos dominated by extreme opinions from the same camp (Fig. 5). Additionally, it is positively correlated with users’ excessive interactions with recommended videos that frequently involve cross-camp interactions (Fig. 7). These findings highlight that on short video platforms, the AI-driven recommendation algorithm, without requiring real social networks, can contribute to the increased polarization.

On social media, the dynamics of opinions are inherently complex. However, for simplification, we often assume homogeneity and monotonicity in opinion dynamics (Bakshy et al. 2015; Cinelli et al. 2021; Flamino et al. 2023): for example, users with similar opinions tend to group together, while those with opposing opinions should avoid gathering. In-camp interactions typically reinforce existing opinions, while cross-camp interactions are expected to challenge and weaken pre-existing opinions. Our study offers vital evidence that two simultaneous algorithmic polarization drivers, i.e., echo chambers and battlefields, co-exist, representing an initial step in unraveling the complexity of opinion dynamics on social media. Moreover, these two polarization drivers manifest asymmetric effects between the two opposing camps. Specifically, the liberal camp participates more in battlefields while the conservative camp engages more in echo chambers (Fig. 6 and Fig. S6 in SI). These findings have implications on the design of intervention strategies to reduce polarization. Traditional strategies such as assembling users with opposing opinions or recommending opposing content, may not yield the intended outcomes (Fig. 7) and could even trigger a “backfire” effect (Bail et al. 2018). Therefore, the design of recommendation algorithms should recognize the intricate nature of opinion dynamics and adopt more nuanced strategies.

Our study has several limitations. First, we conduct the study on one of the largest short video platforms and provide a detailed examination of polarization on feminism. Future work should consider extending the study to multiple platforms in order to mitigate potential data biases. Second, this study leverages a large-scale real-world observational dataset, offering valuable insights into polarization on feminism in the real-world context of China. However, the observational nature of the dataset limits our ability to establish rigorous causal relationships. Future work should consider using experimental approaches to validate these causalities. Third, this study mainly focuses on polarization on feminism. To empirically measure this, we develop a spectrum rooted in feminist theories and the Chinese cultural background. While the developed spectrum encompasses a broad range of radical feminism, liberal feminism, neutrality, conservatism, and patriarchy, it cannot fully capture the complexity of opinions on feminism. Future work should consider measuring them in a more fine-grained manner. Moreover, we focus this study on the salient factor of the AI-driven recommendation algorithm. Other factors, e.g., changes in the social environment, changes in the dissemination of information through new media, etc., could also affect the polarization, which should be further explored in future work. Finally, while this study aims to reveal the concerning phenomenon of polarization on feminism and its underlying drivers, further efforts in algorithm design and policy development are imperative in the pursuit of advancing women’s equality.