Polarization of public opinions on feminism in China

Jinghua Piao, Fang Zhang, Tian‐Ling Ren, Fengli Xu, Jinglei Zhou, Jun Su, Peng Ru, Yong LiHumanities 2025 · DOI

Abstract

Global efforts on feminism and women’s equality are undergoing intense challenges under the transformative impact of social media and artificial intelligence (AI). This study explores the polarization of public opinions on feminism—a prevalent yet underexplored phenomenon—by analyzing large-scale empirical data from one of China’s most prominent short video platforms. Facilitated by a fine-tuned large language model, we assess 62K users’ opinions on feminism based on their 67M fine-grained behavioral records over two years. We observe a severe polarization of opinions on feminism, where users are asymmetrically polarized into two opposing camps—54.3% in the conservative camp and 34.4% in the liberal camp. We further reveal that the AI-driven recommendation algorithm contributes to two primary drivers of this polarization: (i) echo chambers, where the algorithmic gathering of like-minded users reinforces their shared similar opinions and (ii) battlefields, where opposing users conflict, strengthening pre-existing opinions. We point out that these two drivers are asymmetrically applied to the polarized camps, with conservatives more prone to echo chambers while liberals are more inclined to battlefields.

Introduction

Feminism and gender equality have become flashpoints for intense social conflict in China, shaped by a unique intersection of traditional Confucian values, state-led "socialist feminism," and modern liberal influences. As social discourse shifts from traditional social networks, where users follow friends, to AI-driven recommendation platforms, where algorithms curate content, the nature of public opinion changes. This research investigates how "black-box" recommendation algorithms contribute to the polarization of sensitive social issues like feminism, asking whether these platforms merely reflect existing social divides or actively exacerbate them through algorithmic curation.

Methodology and Data

The study utilizes a massive, fine-grained dataset from a prominent Chinese short video platform, comprising 62,224 active users, 71,405 feminism-related videos, 2.16 million comments, and 65 million recommendation records over a two-year longitudinal window (April 2021 to April 2023).

The researchers employed a multi-stage computational approach:

  1. Opinion Spectrum Design: A five-point spectrum was developed to categorize views: Radical Feminism (R), Liberal Feminism (L), Neutral (N), Conservatism (C), and Patriarchy (P).
  2. NLP Classification: A Chinese BERT model (using whole-word masking) was fine-tuned on a manually labeled dataset of 11,856 comments to classify user stances.
  3. User Inference: Individual stances were inferred by aggregating fine-grained behavioral records (comments and recommendations) over time.
  4. Polarization Modeling: A polarization score ($s_{pol}$) was used to quantify how far the community's distribution of opinions deviated from a neutral center.

The Role of Algorithms in Shaping Opinion

The transition from traditional social media to algorithmic platforms fundamentally alters how users interact. While traditional media relies on social ties, short video platforms use bipartite graphs to match users with content based on predicted interest, creating a different kind of community dynamic.

Conceptual framework comparing traditional social media (a) with algorithmic short video platforms (b) and the resulting algorithmic video community (c).

The paper identifies two distinct drivers of polarization within these algorithmic communities:

  • Echo Chambers: These reinforce views by gathering like-minded users. The algorithmic gathering of similar users reinforces their shared opinions.
  • Battlefields: These are high-heterogeneity spaces where opposing sides clash. Crucially, the authors find that these clashes do not moderate views; they actually strengthen pre-existing biases.

A striking finding is Algorithmic Asymmetry: the algorithm treats the two camps differently. Conservatives are more prone to being placed in echo chambers (receiving content that aligns with their views), while liberals are more frequently pushed into "battlefields" (confrontational spaces). Furthermore, because controversial "battlefield" videos attract 3.2x more comments than average, the algorithm's prioritization of popularity creates a feedback loop that promotes polarizing content.

Key Results

  • Severe Polarization: 88.7% of users hold polarized views. The population is asymmetrically split: 54.3% in the conservative camp and 34.4% in the liberal camp.
  • Intensification and Erosion: Polarization increased by 30.8% over the study period. Neutral opinions are being rapidly "eroded" as users migrate toward the extreme ends of the spectrum (Radical Feminism and Patriarchy), which now encompass nearly 40% of users.
  • Classifier Performance: The fine-tuned BERT model achieved an accuracy of 0.72 and an F1 score of 0.70 for the complex 5-way classification task.

Conclusion and Future Directions

This study uncovers the severe polarization of public opinions on feminism and reveals how AI-driven recommendation algorithms act as primary drivers. While the findings are observational and based on a single platform, they offer critical implications for reducing polarization on increasingly AI-dependent social media. The authors suggest that future research should focus on designing "de-polarizing" recommendation strategies that can bridge divides without triggering a "backfire effect," where exposure to opposing views makes a person more extreme.

Summary updated

Abstract

Global efforts on feminism and women’s equality are undergoing intense challenges under the transformative impact of social media and artificial intelligence (AI). This study explores the polarization of public opinions on feminism—a prevalent yet underexplored phenomenon—by analyzing large-scale empirical data from one of China’s most prominent short video platforms. Facilitated by a fine-tuned large language model, we assess 62K users’ opinions on feminism based on their 67M fine-grained behavioral records over two years. We observe a severe polarization of opinions on feminism, where users are asymmetrically polarized into two opposing camps—54.3% in the conservative camp and 34.4% in the liberal camp. We further reveal that the AI-driven recommendation algorithm contributes to two primary drivers of this polarization: (i) echo chambers, where the algorithmic gathering of like-minded users reinforces their shared similar opinions and (ii) battlefields, where opposing users conflict, strengthening pre-existing opinions. We point out that these two drivers are asymmetrically applied to the polarized camps, with conservatives more prone to echo chambers while liberals are more inclined to battlefields. Our study not only uncovers the polarization of public opinions on feminism and its drivers but also offers implications for reducing polarization on increasingly AI-dependent social media.
BibTeX
@inproceedings{piao2025polarization,
  author = {Jinghua Piao and Fang Zhang and Tian‐Ling Ren and Fengli Xu and Jinglei Zhou and Jun Su and Peng Ru and Yong Li},
  title = {{Polarization of public opinions on feminism in China}},
  booktitle = {Humanities},
  year = {2025},
  doi = {10.1057/s41599-025-04635-z},
}
Figure