Unequal Verdicts: Investigating Gender Bias in LLM-Based Fake News Detection

Technical Summary: Unequal Verdicts: Investigating Gender Bias in LLM-Based Fake News Detection

Problem Statement
While Large Language Models (LLMs) are increasingly deployed for automated fact-checking and fake news detection, their susceptibility to gender bias in this specific context remains underexplored. Existing literature acknowledges that LLMs inherit societal stereotypes from training data, yet no study has systematically quantified how gender cues in speaker attributes affect the veracity judgments of identical statements. This gap poses a risk to the reliability and fairness of automated fact-checking systems, as biased labeling can distort public discourse by amplifying or suppressing specific voices based solely on perceived gender.

Methodology
To address this gap, the authors conducted the first systematic investigation of gender bias in LLM-based fake news detection using real-world data.

• Dataset Augmentation: The study utilized the LIAR benchmark, a collection of manually labeled short statements from PolitiFact. The authors augmented this dataset by modifying the speaker_job column. For each statement, they generated three gendered variants of the speaker's job title: Neutral (e.g., "Attorney"), Male (e.g., "Councilman"), and Female (e.g., "Chairwoman"). The statements themselves remained identical, ensuring that only the gender presentation varied.
• Data Curation: The process involved filtering out non-person entries and cases where the speaker's gender was indirectly revealed (e.g., by name or specific context). The final curated dataset consisted of 8,243 samples (4,805 True, 3,438 False), with each instance paired with three gender variants.
• Experimental Setup: Six state-of-the-art LLMs were evaluated: five open-source models (Phi-4 14B, Llama-3.1 8B, Llama-3.2 3B, Gemma-3 12B, Qwen-3 14B) and one proprietary model (GPT-4.1 Mini). The models were prompted to classify statements as "TRUE" or "FALSE" based on the statement and the specific gender variant of the speaker.
• Evaluation Metrics: The authors designed a comprehensive framework to measure bias across three dimensions:
  1. Instability (Flip Rates): Measuring how often predictions change when the gender variant changes (Pairwise Flip Rate, Directional Flip Rate, Conditional Flip Rate).
  2. Aggregate Disagreement: Measuring overall sensitivity to gender cues (Gender-Cue Sensitivity Index - GSI, Unique Disagreement Rate - UDR).
  3. Fairness Disparities: Evaluating systematic differences between Male and Female variants using Demographic Parity (DP), Equalized Odds (EO), and Equal Opportunity (EOpp).

Key Contributions

1. Dataset Creation: Construction of an augmented LIAR dataset with gender-variant speaker job titles, enabling controlled investigation of gender bias in fake news detection.
2. Evaluation Framework: Design of a multi-metric framework to systematically examine gender bias, distinguishing between prediction instability and systematic directional favoritism.
3. Empirical Evidence: Demonstration that all evaluated models exhibit gender sensitivity, revealing two distinct bias manifestations: instability (inconsistent judgments) and directionality (systematic favoritism).
4. Resource Release: Public release of the augmented dataset to support future research and mitigation strategies.

Experimental Results
The study found that all six models exhibited gender sensitivity, though the magnitude and nature of the bias varied significantly:

• Instability: Total flip rates (where a statement received different labels across gender variants) ranged from 9.79% to 35.13%. Specifically, Male–Female comparisons showed flip rates between 6.5% and 23.6%. This indicates that up to nearly one in four statements received different verdicts based solely on gender cues.
• Directional Bias: Five of the six models showed statistically significant directional effects.
  • Male-Skeptic Bias: Gemma-3 12B and Llama-3.1 8B displayed the strongest patterns, systematically assigning more "False" labels to male-coded speakers.
  • Female-Skeptic Bias: Phi-4 14B and Qwen-3 14B showed minor tendencies to assign more "False" labels to female-coded speakers, though with negligible practical magnitude.
  • GPT-4.1 Mini: This model demonstrated superior fairness with no significant directional effects and the lowest overall flip rates (5.8%–7.1%).
• Sensitivity to Truth: Five of the six models showed higher gender-cue sensitivity for statements with a "True" ground truth, suggesting gender cues modulate skepticism more strongly when claims lack obvious falsifying signals.
• Safety Alignment: GPT-4.1 Mini exhibited a unique pattern where the Neutral variant had a higher disagreement rate than the Male or Female variants, potentially indicating that explicit gender terms trigger safety alignment mechanisms that enforce consistency.

Significance and Claims
The paper claims that gender bias fundamentally undermines both the reliability and fairness of LLM-based fake news detection. The authors assert that current LLMs cannot serve as unbiased fact-checking systems without targeted mitigation. However, they also highlight that the model-specific nature of these biases suggests actionable opportunities for intervention in architecture, training data, and alignment strategies. The finding that GPT-4.1 Mini achieved highly equitable performance serves as evidence that fair automated fact-checking is achievable. The authors conclude that bias-aware evaluation and mitigation strategies are essential to harness the potential of LLMs in information verification responsibly.