LLM Bias Evaluation: Gender, Racial, and Age Disparities in Occupational and Crime Scenarios

Technical Summary: Evaluating Gender, Racial, and Age Biases in Large Language Models

Problem Statement
Despite significant advancements in Large Language Models (LLMs), widespread enterprise adoption is hindered by persistent bias issues that compromise usability, reliability, and fairness. Bias in LLMs is not merely a technical artifact but a societal concern capable of reinforcing stereotypes and propagating misinformation. While researchers have developed mitigation strategies—such as debiasing layers, specialized reference datasets (e.g., Winogender, Winobias), and Reinforcement Learning with Human Feedback (RLHF)—residual bias remains. This paper posits that current leading LLMs, despite employing these debiasing techniques, still exhibit significant deviations from real-world statistical distributions in occupational and crime scenarios. Furthermore, the authors hypothesize that mitigation efforts may inadvertently lead to "over-indexing" on specific sub-classes, potentially exacerbating bias rather than resolving it.

Methodology
The study conducts a comparative analysis of four state-of-the-art LLMs released in 2024: Gemini 1.5 Pro, Llama 3 70B, Claude 3 Opus, and GPT-4o. The research focuses on two primary bias categories:

1. Occupational Gender Bias: Assessing the gender association of 20 distinct professions.
2. Demographic Bias in Crime Scenarios: Assessing gender, race, and age associations across six crime types (robbery, fraud, murder, drug abuse, DUI, vandalism).

Data Generation and Classification:
The authors utilized APIs to generate a robust dataset comprising 10,400 stories (8,000 occupational and 2,400 crime scenarios).

• Occupational: Prompts requested 500-word inspiring stories for specific professions.
• Crime: Prompts requested 500-word stories about individuals committing specific crimes, explicitly instructing the models to include demographic details.
• Classification: The same four LLMs were employed to classify the generated stories. Occupational stories were classified by gender (binary: male/female). Crime stories were classified by gender (male/female/non-binary), age groups (<18, 19-30, 30-50, >50), and race (Black, White, Asian).

Benchmarking:
The generated distributions were benchmarked against authoritative real-world data:

• Occupational Data: U.S. Bureau of Labor Statistics (BLS).
• Crime Data: U.S. Federal Bureau of Investigation (FBI) reports.
• Validation: A manual human quality assurance (QA) review was conducted to verify automated classification results.

Key Results
The study quantified bias by calculating the standard deviation of model outputs from real-world statistics.

• Occupational Gender Bias:

  • All four models showed a consistent pattern: they largely adhered to traditional gender stereotypes for female-dominated (e.g., nurse, receptionist) and male-dominated (e.g., firefighter, construction worker) roles.
  • However, significant deviations occurred in high-status or technical roles. For professions like Software Engineers, Lawyers, and CEOs, models frequently generated female characters at rates significantly higher than BLS data, suggesting an attempt at correction that resulted in overcompensation.
  • The average deviation from BLS data across models was approximately 37%.

• Crime Scenario Biases:

  • Gender: Models exhibited divergent behaviors. GPT-4o closely mirrored FBI male statistics in some areas but overrepresented males in others. Gemini 1.5 Pro and Llama 3 70B significantly underrepresented males in crime scenarios (e.g., predicting 96% female involvement in DUI cases where FBI data shows 74% male). Claude 3 Opus presented a more balanced distribution. The average gender deviation was 54%.
  • Race: Most models (Claude, GPT-4o, Llama) overrepresented White individuals in crime scenarios compared to FBI data, while underrepresenting Black individuals. Gemini 1.5 Pro notably showed lower representation of White individuals. The average racial deviation was 28%.
  • Age: Deviations were generally lower than gender and race biases (average 17%), though Llama 3 70B tended to overrepresent the 31-50 age group, and Gemini 1.5 Pro overrepresented the 18-30 group.

• Model Performance Summary (Standard Deviation from Real-World Data):

  • Llama 3 70B showed the highest deviation in Crime Gender bias (71.87%).
  • Gemini 1.5 Pro showed the highest deviation in Crime Racial bias (45.31%).
  • Claude 3 Opus showed the highest deviation in Crime Age bias (30.59%).
  • All models showed similar levels of deviation for Occupational Gender bias (ranging from 33% to 39%).

Significance and Claims
The paper claims that despite the integration of advanced debiasing techniques like RLHF and red teaming, leading LLMs continue to exhibit substantial bias. A critical finding is that mitigation strategies often result in over-indexing on specific sub-classes (e.g., over-representing women in male-dominated professions or White individuals in crime scenarios) rather than achieving statistical parity.

The authors argue that these findings highlight the limitations of current bias mitigation approaches. The study suggests that simply applying debiasing layers or fine-tuning on specific datasets is insufficient and may lead to poor generalization or new forms of bias. The paper concludes that there is an urgent need for more effective, nuanced approaches to bias mitigation to ensure responsible AI adoption, particularly as LLMs are increasingly integrated into enterprise applications, hiring processes, and judicial support systems. The proposed framework serves as a baseline for evaluating future models and other bias types (political, socio-economic) through open-ended prompt engineering.