Value Alignment of Social Media Ranking Algorithms (2509.14434v1)

Abstract: While social media feed rankings are primarily driven by engagement signals rather than any explicit value system, the resulting algorithmic feeds are not value-neutral: engagement may prioritize specific individualistic values. This paper presents an approach for social media feed value alignment. We adopt Schwartz's theory of Basic Human Values -- a broad set of human values that articulates complementary and opposing values forming the building blocks of many cultures -- and we implement an algorithmic approach that models and then ranks feeds by expressions of Schwartz's values in social media posts. Our approach enables controls where users can express weights on their desired values, combining these weights and post value expressions into a ranking that respects users' articulated trade-offs. Through controlled experiments (N=141 and N=250), we demonstrate that users can use these controls to architect feeds reflecting their desired values. Across users, value-ranked feeds align with personal values, diverging substantially from existing engagement-driven feeds.

• The paper presents a framework that aligns feed ranking with human values via LLM-driven annotation using Schwartz's 19-value model, achieving a low MAE of 0.95.
• It validates the approach through experiments where users recognized and configured feeds reflecting their prioritized values over standard engagement metrics.
• The findings imply that incorporating human values into ranking algorithms can enhance user agency and mitigate filter bubbles on social media.

Value Alignment of Social Media Ranking Algorithms

Introduction

This paper presents a comprehensive framework for aligning social media feed ranking algorithms with human values, operationalized via Schwartz's theory of Basic Human Values. The authors argue that engagement-driven ranking is not value-neutral and tends to amplify individualistic, short-term values. They propose a method for classifying and integrating value expressions in social media posts, enabling user-driven value prioritization in feed curation. The approach is validated through controlled experiments, demonstrating that users can recognize and configure feeds that reflect their articulated values, with value-ranked feeds diverging substantially from engagement-based rankings.

Operationalizing Human Values in Feed Ranking

The core of the framework is the adoption of Schwartz's circumplex model of 19 basic human values, grouped into four clusters: self-transcendence, conservation, self-enhancement, and openness to change. This model provides structural coverage of the value design space, capturing both complementary and conflicting motivations.

Figure 1: Visualization of Schwartz's 19 Basic Human Values, organized into four broader groups and mapped onto a circumplex to illustrate motivational proximities and tensions.

To annotate social media content, the authors employ LLMs, specifically GPT-4o, using few-shot prompting to rate the presence and magnitude of each value in tweets on a 0–6 scale. The LLM classifier achieves a mean absolute error (MAE) of $0.95 \pm 1.10$ against consensus human labels, outperforming individual human annotators ($1.07 \pm 1.05$ MAE) for most values. This demonstrates that LLMs can reliably scale value annotation for large content inventories.

Value-Based Ranking Algorithm

The ranking algorithm integrates user-specified value weights with LLM-derived value labels for each post. Weights are assigned in $[-1, 1]$ for each value, and the ranking score for a post is computed as the dot product of the value label vector and the weight vector. This linear aggregation allows for additive trade-offs and supports multi-value prioritization.

Figure 2: Example feeds ranked by different value weight configurations, illustrating the impact of prioritizing "Achievement" and "Dominance" versus "Humility" and downranking "Face".

The system supports both implicit learning of user preferences and explicit user control via interfaces. The latter is implemented through slider-based controls, allowing users to adjust the prominence of each value in their feed.

Experimental Validation

Study 1: Single-Value Ranking

Participants ($N=141$) were asked to identify which of two feeds—one engagement-ranked, one value-ranked by their top personal value—reflected the specified value. The recognizability rate was $76.1\%$, significantly above random chance ($p < 0.001$). Most values produced recognizable feeds, with the exception of "Interpersonal Conformity".

Figure 3: Experimental platform rendering side-by-side feeds for recognizability assessment.

Study 2: Multi-Value, User-Controlled Ranking

A larger experiment ($N=250$) evaluated recognizability when users could adjust 1–19 value sliders. Recognizability dropped as the number of sliders increased, but remained above chance ($63.4\%$ for multi-value feeds). Complexity ratings did not significantly increase with more sliders, though qualitative feedback indicated some choice overload in the full-slider condition.

Figure 4: User interface for adjusting value sliders to re-rank the feed.

Figure 5: Recognizability as a function of the number of sliders changed, showing a decline but remaining above random chance even for high-dimensional control.

Empirical Insights into Value Selection and Feed Composition

Analysis of user selections revealed that configured value weights were positively correlated with personal values ($r = 0.39 \pm 0.25$), and that users often mapped topical interests to underlying values. Engagement feeds were found to prioritize individualistic values such as "Hedonism" and "Stimulation", with value-ranked feeds amplifying self-transcendence and openness to change values.

Figure 6: Engagement feeds prioritize individual/personal values over societal ones, as measured by value strength across participants.

Figure 7: Value-ranked feeds amplify self-transcendence and openness to change values relative to engagement feeds.

The distribution of value expressions across tweets was sparse, with most tweets not containing any given value, but 94.4% of tweets were value-laden in at least one dimension.

Figure 8: Distribution of labeled values across 212,663 tweets, showing sparsity and prevalence of "Stimulation" and "Hedonism".

Implications and Future Directions

Practical Implications

• User Agency: The framework enables end-users to exert granular control over the values amplified in their feeds, addressing limitations of centralized, engagement-driven curation.
• Algorithmic Diversity: Value-based ranking can counteract algorithmic monoculture and potentially mitigate filter bubbles by supporting pluralistic value configurations.
• Scalability: LLM-based annotation is robust and scalable, supporting real-time value-based feed curation for large inventories.

Theoretical Implications

• Value Trade-offs: The circumplex model supports nuanced trade-off navigation, but interface design must balance expressivity and cognitive load.
• Generalizability: While Schwartz's values are comprehensive and cross-cultural, the method generalizes to other value systems (e.g., Moral Foundations Theory) with similar structural properties.
• Subjectivity in Value Perception: Future work should address personalization of value classifiers to account for individual and cultural differences in value interpretation.

Limitations

• Bias in LLMs: Potential overrepresentation of Western-centric values in LLM outputs remains a concern; pluralistic alignment methods are needed.
• Content Inventory Constraints: The ability to satisfy user value preferences is limited by the available content; platform-level integration could alleviate this.
• Longitudinal Effects: The studies were single-session; long-term impacts on user behavior and engagement require further investigation.
• Cultural Generalizability: Experiments were US-centric; cross-cultural validation is necessary.

Conclusion

The paper demonstrates that social media feed ranking algorithms can be systematically aligned with human values using a scalable, LLM-driven annotation and ranking framework grounded in cultural psychology. Users can recognize and configure feeds that reflect their articulated values, with value-ranked feeds diverging substantially from engagement-based rankings. The approach offers a pathway for democratizing feed curation, supporting pluralistic value expression, and informing future research on value alignment in AI systems.