Gender Bias in YouTube Exposure: Allocative and Structural Inequalities in Political Information Environments

Abstract: Recommendation algorithms have become the dominant mechanism for information distribution on digital platforms, profoundly shaping personalized information consumption environments. However, gender bias, as a significant form of algorithmic discrimination, may cause users to experience unequal exposure within different political information environments. Taking YouTube as a case, we conduct a controlled social-bot field experiment, where male-coded and female-coded profiles are constructed. We track the exposure and click patterns of these bots to analyze their recommendation trajectories. We analyze the distribution of recommended content from two dimensions: allocative bias and structural bias. First, we find statistically significant differences in allocative bias across male-coded and female-coded profiles, particularly in terms of issue distribution, ideological orientation, and political entities. Secondly, we observe structural bias in the political information environments, characterized by distinct clustering patterns. Additionally, time-series analysis shows that exposure pathways continue to be shaped over time by both communities detected in the co-occurrence network and individual profile-level dynamics. Finally, we construct a simple collaborative-filtering model that reproduces the observed gender bias. We argue that gender bias in recommendation systems is reflected not only in the allocation of political content, but also in how community structures shape these environments, reinforcing societal inequalities and highlighting the need for algorithmic fairness.

• The paper demonstrates that controlled social-bot experiments reveal significant gender bias in YouTube's political content allocation, with female-coded profiles receiving higher exposure.
• It uses a robust methodology with 160 bots and network analysis to show that male-coded profiles exhibit higher content diversity while experiencing lower structural diversity.
• The paper validates a collaborative-filtering model, illustrating that even minimal gender-homophily in algorithms can lead to enduring community segregation and bias in content recommendations.

Gender Bias in YouTube Exposure: Allocative and Structural Inequalities in Political Information Environments

Overview and Objectives

This study systematically examines gender bias in YouTube's recommendation system within political information environments, focusing on both allocative and structural inequalities. Through a controlled social-bot field experiment, the research isolates the effects of algorithmic design from user-side confounds by programming bots with male-coded and female-coded behavioral signals. By analyzing the resulting recommendation streams, the paper establishes not only statistically significant allocative differences in political content exposure, but also persistent structural segregation in the political information ecologies encountered by different gender-coded profiles. Furthermore, the authors propose and validate a collaborative-filtering-based model to explore the minimal algorithmic conditions under which such group differentiation can emerge.

Experimental Design

The methodological foundation rests on creating 160 controlled social bot accounts (80 male-coded, 80 female-coded), initialized with platform-representative gendered interests but no prior political preference differentials. The bot interaction protocol comprises an initial training phase (to establish behavioral profile), followed by 150 consecutive recommendation steps emulating typical user engagement dynamics including viewing patterns and click distributions. Political content in the recommendation streams is annotated using an LLM-assisted hybrid approach, with additional manual expert validation for classification accuracy in issues, ideology, and political entities.

Allocative bias is measured across share and diversity of political exposures, issue and ideological distributions, and prominent entities. Structural bias is quantified via account co-exposure networks, analyzing density, clustering, modularity, and community structure, and employing Social Network Analysis (SNA) to track the evolution of community boundaries and their feedback effect on subsequent exposure trajectories.

Key Findings

Allocative Bias

• Statistically significant differences in the proportion of political content: Female-coded profiles are exposed to more political content than male-coded (final 50 steps: 17.2% ± 10.1% for females vs 13.4% ± 7.8% for males, $p < 0.01$), despite no a priori difference in political engagement.
• Male-coded profiles exhibit higher content category entropy (fine-grained diversity), but lower structural diversity, suggesting a narrower recommendation focus on initial behavioral interests.
• Issue assignments differ systematically: Male-coded profiles are concentrated on hard/domestic-order issues (law, crime, defense), while female-coded profiles receive a broader and more multidimensional spectrum, including international, macroeconomic, and cultural policy topics ($p < 0.01$ for multiple issue classes).
• In ideological exposure, female-coded profiles receive higher proportions of neutral content and lower left-leaning exposure compared to males, challenging the prevailing assumption that partisan content allocation tracks gender preference exclusively. Right-leaning content exposure is not statistically different between groups.
• Salient political entities in recommendations differ by profile group, with male-coded streams featuring state-power and security organizations, while female-coded streams include more international and public-affairs institutions.

Structural Bias

• Distinct network topologies: Co-exposure networks for male-coded profiles have higher density and clustering coefficients, indicating high within-group homogeneity and repeated exposure to overlapping content. Female-coded networks show increased modularity and community segmentation, leading to a more diffuse and internally heterogeneous information environment.
• Community detection and feedback: Once established, these community structures persist and become predictive of subsequent exposure and engagement patterns. Lagged regression and similarity analyses reveal that recommendations propagate primarily along community boundaries, reinforcing structural segregation.
• Structural bias is not static but dynamically reinforced via exposure–click–re-exposure loops, compounding group differentiation over time.

Modelled Mechanism

A minimal agent-based collaborative-filtering simulation demonstrates that even weak gender-homophily in the similarity computation, coupled with low-randomness recommendation sampling, suffices for the emergence and reinforcement of both allocative and structural bias. This effect occurs absent any pre-specified group differences in political preferences, indicating the critical role of algorithmic design rather than user-side inputs.

Implications

Practical Implications

The findings implicate YouTube's recommendation system as an active organizing force in structuring asymmetric political information environments according to gender-coded behavioral cues. Allocative bias manifests not as crude exclusion or visibility suppression, but rather as differentiated allocation of topics, ideological valence, and political actors. More consequentially, structural bias entrenches group segregation over time, potentially restricting cross-group political learning and reducing pluralism in digital public spheres.

For platform operators and regulatory stakeholders, this demonstrates that algorithmic fairness cannot be ensured solely by balancing surface-level exposure metrics. Fairness interventions must address both content allocation and the emergent network structure—especially persistent community boundaries produced by similarity-based models.

Theoretical Implications

The research substantiates that allocative and structural biases are inherently intertwined phenomena in platform information environments. Identity-driven information ecologies (IDIE) are thus not merely aggregates of individualized exposure histories, but collective structures shaped by feedback-amplified group differentiation processes. This highlights the need for dynamic, ecology-aware models of algorithmic bias that move beyond static fairness or representation frameworks.

The agent-based model elucidates a potential minimal mechanism—a slight, often implicit homophily in collaborative filtering—sufficient for complex bias dynamics absent any in-built ideological segmentation. This raises new challenges for AI ethics and transparency, specifically around the opacity of subtle behavioral-cue amplification in recommender systems.

Future Directions

This paper suggests multiple directions for further inquiry:

• Expanding beyond binary gender coding to other identity dimensions or more granular behavioral typologies.
• Longitudinal studies aligning observed information ecologies with downstream attitudinal or behavioral changes in real-world users.
• Development and evaluation of de-biasing algorithms or structural interventions targeting both allocative and network-level biases in large-scale recommendation systems.
• Analysis of platform algorithm updates and their short/long-run effects on ecological bias dynamics.

Conclusion

This study rigorously demonstrates that YouTube's recommendation system generates both allocative and structural gender bias in the political information it distributes, even after tightly controlling for user-side variables. The results emphasize the necessity for fairness auditing protocols that consider not only exposure shares, but also content allocation, community structure, and dynamic feedback loops. Algorithmic platforms thus play a constitutive role in the reproduction and amplification of societal inequalities within digital political communication environments, warranting enhanced scrutiny and informed intervention.