TL;DR
Large Language Models (LLMs) exhibit political biases influenced by the names of media outlets they process, impacting tasks like news bias prediction and summarization. This paper introduces a novel metric (SIPS) to quantify this bias and proposes an automated prompt optimization framework to effectively mitigate it, making LLMs fairer in news-related applications.
What Problem Does This Paper Address?
While extensive research has explored political biases in LLMs' text generation and perception, a critical and underexplored dimension is the bias induced by the names of media outlets themselves. Humans are known to develop biases towards media outlets based on their perceived political leanings, and this tendency is rooted in our inclination to favor information aligning with pre-existing beliefs. This paper investigates whether LLMs, similar to humans, internalize these biases and how they manifest as "anchoring effects" when processing news content attributed to specific media outlets. This is particularly crucial given the increasing deployment of LLMs in news generation and analysis, where impartiality and factual integrity are paramount. The study aims to quantify this media outlet name bias and develop methods to mitigate it.
Methods
The paper employs a two-pronged approach:
- Measuring Media Outlet Name Bias:
- Political Bias Prediction: LLMs are prompted with news articles, and their predicted political bias is analyzed while varying the attributed media outlet name. A novel metric, the Source-Induced Prediction Shift (SIPS), is introduced. SIPS quantifies bias by considering two components: Absolute Sensitivity (AS), measuring the magnitude of prediction shifts regardless of direction, and Agreement Coherence (AC), measuring how well the direction of the shift aligns with the known political orientation of the outlet.
- Article Summarization: LLMs are prompted to summarize news articles with varying attributed media outlet names. The sentiment of named entities within the generated summaries is analyzed to detect systematic changes influenced by the outlet name.
- Mitigating Media Outlet Name Bias:
- Automated Prompt Optimization: An LLM (GPT-4.1) is used as an optimizer within a framework inspired by Yang et al. (2024b). This framework iteratively refines prompts to reduce the SIPS score, aiming to make the LLM less sensitive to media outlet names. The SIPS, AS, and AC scores serve as the objective signal for the optimization process.
Additionally, saliency analysis is used to understand the mechanistic basis of the bias, quantifying the attention LLMs pay to media outlet name tokens versus content tokens.
Datasets
The study utilizes the following datasets:
- AllSides Dataset: Used for political bias prediction and summarization experiments. A sample of 1,500 articles per bias class (left, center, right) was used for prediction, and a subset of 450 articles for summarization.
- Hyperpartisan News Detection dataset: Used for political bias prediction, including all 1,273 articles.
- Fictitious Media Outlet Names: Created using two methods: a "formulated approach" combining parts of real outlet names (e.g., "Millennial Times") and a "generated approach" where GPT-4.1 created politically biased fictional names. These were used to test LLM sensitivity to political connotations of names.
Key Insights and Ideas
- LLMs are susceptible to media outlet name bias: The study confirms that LLMs consistently exhibit political bias influenced by the names of media outlets, even when the article content is the same.
- Bias is driven by linguistic cues, not just factual knowledge: Saliency analysis reveals that LLMs pay disproportionately more attention to media outlet name tokens than to article content tokens, suggesting that the bias is mechanistically rooted in this attention.
- Alignment and model size amplify bias: Larger models and those that have undergone alignment tuning (like instruction tuning or RLHF) tend to show stronger biases.
- Fictitious names also trigger bias: LLMs demonstrate sensitivity to the implied ideological cues in fabricated outlet names, indicating they can infer political connotations.
- Near-neutral articles are more vulnerable: Articles with less inherent bias are more susceptible to shifts in predicted political stance due to media outlet attribution.
- Summarization sentiment is affected: Media outlet names can subtly alter the sentiment of generated news summaries.
- Prompt optimization is an effective mitigation strategy: The proposed automated prompt optimization framework successfully reduces media outlet name bias.
Key Results
- Quantifiable Bias: The SIPS metric effectively quantifies media outlet name bias, showing significant scores across various LLMs. For instance, Qwen-2.572B-Instruct achieved a high SIPS score of 0.529 on the AllSides dataset.
- Model Variations: LLMs exhibit varying degrees of bias. Qwen-2.572B-Instruct and Gemma-227B-IT showed strong sensitivity to right-leaning sources, while GPT-4.1 exhibited modest bias magnitude but clear direction.
- Alignment Amplification: Alignment tuning consistently led to higher SIPS scores compared to base versions of models.
- Summarization Sentiment Shifts: For Mistral-Small24B-Instruct, the proportion of positive entities in summaries changed by an average of 8.45% when outlet names were included, with negative and neutral entities showing even larger fluctuations.
- Prompt Optimization Success: The automated prompt optimization framework reduced SIPS for Qwen-2.572B-Instruct from 0.529 to 0.292 after seven iterations, falling below the 0.3 threshold. This mitigation was also effective for Gemma-227B-IT.
- Human Evaluation: Four out of five annotators detected bias perception shifts in LLM-generated summaries when the attributed media outlet name was varied, validating the LLM findings.
Limitations and Future Work
- Geographic and Linguistic Scope: The study focused exclusively on English-language U.S. news media, limiting generalizability to other languages or global contexts.
- Task Specificity: Bias was analyzed only in prediction and summarization tasks. Architectural mitigation strategies were not explored.
- Sentiment Analysis Granularity: The sentiment-based analysis provided limited directional assessment.
- Dataset Sensitivity: The metrics showed sensitivity to article selection.
- Ground Truth Subjectivity: The study relied on datasets like AllSides and Hyperpartisan, whose annotations are subjective and may encode annotator biases.
- LLM as Political Actors: The work highlights LLMs' potential to shape public perception, emphasizing that while they lack intent, their outputs reflect statistical patterns that can have sociopolitical consequences.
Future work could explore bias in other tasks, investigate architectural mitigation, expand to multilingual settings, and conduct more in-depth analyses of the interaction between article content and outlet bias. The authors also call for community-driven evaluations to verify and extend their findings.
Summary updated