💬 NLP

Polarization by Default: Auditing Recommendation Bias in LLM-Based Content Curation

This study reveals that LLM-based content curation systematically amplifies polarization and exhibits distinct, provider-specific biases in toxicity handling and sentiment, while consistently over-representing left-leaning authors on Twitter/X regardless of prompting strategies.

Original authors: Nicolò Pagan, Christopher Barrie, Chris Andrew Bail, Petter Törnberg

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicolò Pagan, Christopher Barrie, Chris Andrew Bail, Petter Törnberg

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a massive, chaotic library where millions of people are shouting stories, jokes, news, and rants all at once. In the past, a human librarian would stand on a ladder, listen to a few shouts, and decide which stories to put on the "Best of the Week" shelf for everyone to read.

Now, imagine that human librarian has been replaced by three different super-smart robots (OpenAI's GPT, Anthropic's Claude, and Google's Gemini). These robots are tasked with reading 100 shouting voices and picking the top 10 to put on the shelf.

This paper is a massive experiment where the researchers asked: "If we give these robots different instructions, do they pick different stories? And are they fair?"

Here is the breakdown of what they found, using some everyday analogies:

1. The "Default" Robot is Already Biased

Even when the researchers told the robots to be "neutral" (just "Rank these posts"), the robots didn't act like blank slates. They had their own hidden preferences baked into their brains from all the data they were trained on.

The Analogy: It's like asking a chef to "just make a sandwich" without specifying ingredients. The chef doesn't make a plain sandwich; they make a sandwich with their favorite toppings because that's what they learned to love.
The Finding: Even with no specific instructions, the robots consistently picked more polarized (extreme) content. They loved drama and conflict.

2. The "Mood Ring" Effect (Prompting)

The researchers tested six different "moods" or instructions for the robots, like asking them to find "the most popular post" or "the most educational post."

The Analogy: Think of the robots as actors. If you tell an actor to play a "villain," they act scary. If you tell them to play a "hero," they act brave. The script (the prompt) changes how they behave, but the actor (the specific robot model) still has their own personality.
The Twist:
- Toxicity: When asked to find "engaging" content (likes/shares), the robots actually liked toxic, mean-spirited posts. But when asked for "informative" content, they immediately rejected the mean posts.
- Sentiment: When asked to be "engaging," the robots preferred negative, angry, or sad content. When asked to be "informative," they preferred positive content.

3. The "Political Lens" on Twitter

On Twitter (X), the researchers could guess the political views of the people posting based on their bios.

The Setup: In their test pool, there were actually more right-leaning people shouting than left-leaning people (a 43% vs. 17% split).
The Result: Despite the right-leaning crowd being larger, all three robots consistently picked more left-leaning voices.
The Analogy: Imagine a town where 70% of the people are wearing Red hats and 30% are wearing Blue hats. But every time the town mayor (the robot) picks a speaker for the stage, they pick someone wearing a Blue hat. It's not because the mayor is looking at the hats directly; it's because the mayor has learned that "Blue hat" is associated with "interesting speech" in their training data.
Crucial Point: The robots weren't told who was left or right. They just read the text. They inferred the bias from the style of the writing.

4. The Three Robots Have Different Personalities

The study compared the three big AI providers, and they acted like three different people:

OpenAI (GPT-4o Mini): The Steady Eddie. No matter what instruction you gave, this robot stayed the most consistent and balanced. It didn't swing wildly between being mean or nice.
Anthropic (Claude): The Chameleon. This robot was very sensitive to the "safety" instructions. It was great at avoiding toxic content when asked to be informative, but it was the most likely to let toxic content slide when asked to be "engaging."
Google (Gemini): The Grump. This robot had the strongest preference for negative, angry, and toxic content, especially when asked to be engaging.

5. The "Hidden Hand" of Bias

The most scary part of the study is how the bias happens.

The Finding: The robots didn't explicitly say, "I am picking this person because they are a woman" or "because they are left-leaning."
The Analogy: It's like a hiring manager who says, "I didn't look at the resume's name, but I rejected the candidate because their writing style was 'too casual'." The manager didn't look at the name, but the "casual style" was actually a proxy for the candidate's background.
The Reality: The robots picked left-leaning or specific demographic groups because those groups tended to write in a certain style (polarized, specific topics) that the robots had learned to love. Even if you hid the author's name, the robots would still pick the same people because of how they wrote.

The Big Takeaway

We are handing over the keys to what billions of people see on social media to these robots.

The Problem: These robots naturally love drama, anger, and polarization.
The Illusion: We think we can fix this just by changing the instruction (the prompt). But the study shows that while prompts change things a little, the robots' "default setting" is still biased.
The Warning: If we let these robots curate our news feeds without strict oversight, they will likely show us a world that is more angry, more divided, and more skewed toward certain political views than reality actually is.

In short: The robots aren't neutral librarians; they are biased curators with their own favorite genres, and they are currently amplifying the loudest, angriest voices in the room.

1. Problem Statement

Large Language Models (LLMs) are increasingly being deployed to curate and rank human-created content on social media platforms (e.g., X/Twitter, Bluesky, Reddit). While prior research has documented biases in traditional recommender systems (e.g., popularity bias) and in LLMs used for generation or Q&A, there is a critical gap in understanding LLM-based content curation. Specifically, it is unclear:

Which biases are structural (inherent to the model training/alignment) versus incidental (prompt-dependent).
How biases vary across different LLM providers (OpenAI, Anthropic, Google).
Whether prompt engineering can effectively mitigate these biases.
How these systems handle sensitive demographic attributes (political leaning, gender, minority status) in a social media context.

2. Methodology

The authors conducted a controlled simulation study involving 540,000 simulated top-10 selections across 54 experimental conditions.

Datasets: Real social media posts from three platforms:
- Twitter/X: 5,000 posts (allows for demographic inference via bios).
- Bluesky: 5,000 posts.
- Reddit: 5,000 posts.
LLM Providers: Three major models were tested:
- OpenAI: GPT-4o Mini
- Anthropic: Claude Sonnet 4.5
- Google: Gemini 2.0 Flash
Prompt Strategies: Six distinct prompting styles were used to vary the recommendation objective:
1. Neutral: "Rank these posts."
2. General: "Most interesting to a general audience."
3. Popular: "Most popular/viral."
4. Engaging: "Generate the most engagement (likes, shares, comments)."
5. Informative: "Most informative and educational."
6. Controversial: "Thought-provoking or generate debate."
Experimental Design:
- For each condition, 100 independent trials were run.
- In each trial, 100 posts were randomly sampled from the pool, and the LLM was asked to rank the top 10.
- Crucial Constraint: Models received only raw post text. No metadata (author names, engagement counts, profile bios) was provided. This isolates model-level biases learned during pre-training from explicit signal usage.
Feature Engineering: Posts were characterized by 13 features:
- Content: Polarization score, toxicity, sentiment, subjectivity, word length, topic.
- Demographics (Twitter/X only): Gender, political leaning, minority status (inferred via a two-LLM agreement protocol using profile bios).
Bias Metrics: Bias was quantified by comparing feature distributions in recommended posts vs. the candidate pool using Cramér's V (categorical) and Cohen's d (numerical), converted to $R^2$ for comparability.

3. Key Contributions

Systematic Multi-Provider Audit: The first study to simultaneously compare bias across three major LLM providers, three platforms, and six prompting strategies in a social media curation context.
Differentiation of Bias Types: Distinguishes between structural biases (robust across prompts) and prompt-sensitive biases (modifiable via instruction).
Demographic Inference in Curation: Demonstrates how demographic biases manifest even when models are not explicitly given demographic data, revealing indirect discrimination via content correlations.
Prompt Engineering Limits: Provides empirical evidence that while prompt engineering can alter content selection (e.g., toxicity), it is insufficient for correcting deep-seated demographic biases.

4. Key Results

A. The Landscape of Bias (RQ1)

Polarization is Universal: Polarization is the strongest predictor of selection across all models and conditions. All models consistently prefer more polarized content, even under "neutral" or "informative" prompts.
Prompt Sensitivity:
- Content Biases: Highly sensitive to prompts. For example, "Informative" prompts reduce toxicity, while "Engaging" prompts increase it.
- Demographic Biases: Highly resistant to prompts. Political leaning bias persists across all prompt styles.
Mechanism: SHAP analysis reveals that demographic bias is indirect. Models do not explicitly use author attributes (which were hidden); instead, they select content based on features (like polarization or topic) that correlate with specific demographics in the training data.

B. Provider-Specific Behaviors (RQ2)

OpenAI (GPT-4o Mini): Exhibits the most stable and consistent behavior across all prompts. It shows moderate positive bias for polarization but maintains a balanced profile regarding toxicity and sentiment.
Anthropic (Claude): Shows high adaptivity in toxicity handling. It actively avoids toxic content under "informative" prompts but amplifies it under "engaging" prompts.
Google (Gemini): Displays the strongest negative sentiment preference and the highest overall amplification of toxic content, particularly under engagement-focused prompts.

C. Demographic Biases on Twitter/X (RQ3)

Political Leaning: A robust left-leaning bias was observed. Despite the candidate pool being predominantly right-leaning (43.4% right vs. 17.9% left), left-leaning authors were systematically over-represented in recommendations (+9.0 percentage points). This pattern held across all providers and prompts, though "Controversial" prompts amplified it, and "Popular" prompts nearly eliminated it.
Gender & Minority Status: Results were weaker and less consistent.
- Gender: Bias direction varied by model (e.g., Claude under-represented females; Gemini/GPT-4o-mini slightly over-represented them).
- Minority Status: High "unknown" rates (48.4%) limited conclusions, but non-minority authors were generally under-represented.

D. Content vs. Demographics

Content Features: Models prefer News & Social Concern topics and longer, more formal vocabulary.
Toxicity Inversion: A striking dichotomy exists: "Engaging" prompts lead models to tolerate/prefer toxic content, while "Informative" prompts lead to strong toxicity aversion.

5. Significance and Implications

"Polarization by Default": The study concludes that polarization amplification is a default behavior of current LLMs, surfacing even without explicit instructions to be controversial. This suggests that simply removing explicit engagement goals does not create a neutral feed.
Prompt Engineering is Insufficient: While prompts can shift content preferences (e.g., reducing toxicity), they cannot easily fix demographic biases (e.g., political leaning) because these biases are structural to the model's learned associations.
Real-World Impact: As platforms like X (Grok) and Bluesky (Attie) integrate LLMs for feed curation, these biases risk compounding over time, potentially distorting public discourse and reinforcing political divides.
Recommendations: The authors argue that mitigation requires more than prompt tuning. It necessitates adversarial debiasing, fairness constraints during ranking, training data curation, and hybrid human-AI oversight. Furthermore, the field must prioritize standards for demographic data collection and inference validation.

Conclusion

The paper establishes that LLM-based content curation introduces systematic, often structural, biases that favor polarized content and exhibit a default left-leaning political skew, regardless of the provider or prompt. While prompt engineering offers some control over content safety (toxicity), it fails to address fundamental fairness issues regarding demographic representation, highlighting the urgent need for deeper architectural and regulatory interventions before widespread deployment.