Abstract
Large Language Models (LLMs) are increasingly used as information retrieval tools, yet their internal biases can skew the visibility of different demographics. This study investigates gender bias in LLMs by leveraging the "parallel" structure of Olympic Games data. Findings indicate that while models demonstrate nearly equal factual knowledge of both men’s and women’s sports when explicitly prompted, they exhibit a pervasive "default to male" bias in ambiguous contexts. This bias is only reversed in disciplines stereotypically associated with women, such as artistic gymnastics.
Introduction
As LLMs become primary interfaces for information, the way they handle underspecified queries becomes critical. If a user asks, "Who won the gold in rowing in 2012?", the model's choice to surface men's results, women's results, or both reflects underlying factual favoritism. This "default to male" heuristic reinforces societal biases, effectively rendering the achievements of women invisible in general queries. This paper moves beyond toxic language or simple stereotypes to address bias in factual information retrieval.
Methodology and Data
The research utilizes a novel dataset of Olympic Games results from 1988 to 2021, specifically focusing on 338 team events with parallel competitions for men and women. Six instruction-tuned models, including GPT-4o, Llama 3.1, and Mistral-Large, were evaluated using two prompt types:
- Specified Prompts: Explicitly requesting a gender (e.g., "Women's Rowing").
- Underspecified Prompts: Using neutral phrasing (e.g., "Rowing").
Bias was quantified through three metrics:
- Knowledge-based Bias: Accuracy difference (F1 score) when gender is specified.
- Explicit-Ambiguous Bias: Frequency of providing only one gender's results while labeling them as such in response to neutral prompts.
- Implicit-Ambiguous Bias: Frequency of providing one gender's results without labeling them in response to neutral prompts.
Results and Discussion
The study found that factual knowledge is balanced across genders; accuracy for women's events was nearly identical to men's when explicitly requested. However, a significant "retrieval preference" for male data emerged in ambiguous scenarios.
As shown in the table above, models like GPT-4o and Mistral-Large achieve high accuracy but still exhibit notable implicit and explicit bias. GPT-4o-mini, for instance, showed the highest explicit bias (0.22), while Llama 3.1-70b demonstrated a strong implicit bias (0.29), often presenting men's medalists as the "default" winners without qualification.
The Role of Stereotypes
The bias is not uniform across all sports. The "default man" heuristic appears to be influenced by societal stereotypes regarding masculinity and femininity in athletics.
Artistic Gymnastics stood out as the only discipline where the bias flipped (-0.32), favoring women's results. Conversely, disciplines like Fencing (0.43) and Swimming (0.36) showed the heaviest tilt toward male results. This suggests that LLMs mirror specific cultural associations rather than applying a blanket male default.
Conclusion
The findings highlight that even highly aligned models like GPT-4o struggle with neutral information retrieval. While the models "know" the facts about women's sports, their architecture or training data leads them to prioritize male achievements in the absence of specific instructions. Future work should explore whether this "defaulting" behavior extends to other domains with parallel structures, such as corporate leadership or academic awards, to further understand the scope of factual favoritism in AI.
Summary updated