Gender Bias in Decision-Making with Large Language Models: A Study of Relationship Conflicts

TL;DR

Gender equity within LLMs is studied through a decision-making lens with a new dataset, DeMET Prompts, containing scenarios related to intimate, romantic relationships, and while all models exhibit the same biases, safety guardrails reduce bias.

Abstract

Large language models (LLMs) acquire beliefs about gender from training data and can therefore generate text with stereotypical gender attitudes. Prior studies have demonstrated model generations favor one gender or exhibit stereotypes about gender, but have not investigated the complex dynamics that can influence model reasoning and decision-making involving gender. We study gender equity within LLMs through a decision-making lens with a new dataset, DeMET Prompts, containing scenarios related to intimate, romantic relationships. We explore nine relationship configurations through name pairs across three name lists (men, women, neutral). We investigate equity in the context of gender roles through numerous lenses: typical and gender-neutral names, with and without model safety enhancements, same and mixed-gender relationships, and egalitarian versus traditional scenarios across various topics. While all models exhibit the same biases (women favored, then those with gender-neutral names, and lastly men), safety guardrails reduce bias. In addition, models tend to circumvent traditional male dominance stereotypes and side with 'traditionally female' individuals more often, suggesting relationships are viewed as a female domain by the models.

Figures (6)

• Figure 1: Average scores across all models and scenarios for each relationship type (M=man, N=gender neutral, and W=woman). Scores leaning negative align with NAME1 and scores leaning positive align with NAME2 in the scenarios. Within a pairing (e.g. W-M), NAME1 refers to the first label (e.g. W) and NAME2 refers to the second (e.g. M). Section A shows the baseline model decisions in same-gender relationships. Sections B, C, and D show differences in paired mixed-gender relationships, where biases are shown in how each bar deviates from the same-gender baseline (towards NAME1 or NAME2). Differences between each pair of bars indicate how amplified the bias is within each paired mixed-gender relationship.
• Figure 2: Differences in score distributions between egalitarian and traditional scenarios.
• Figure 3: Average scores for five models across each relationship type (M=man, N=gender neutral, and W=woman) on DeMET Prompts. Scores leaning negative align with NAME1 and scores leaning positive align with NAME2 in the scenarios. Within a pairing (e.g. W-M), NAME1 refers to the first label (e.g. W) and NAME2 refers to the second (e.g. M).
• Figure 4: Average scores for eight models across each relationship type (M=man, N=gender neutral, and W=woman) on DeMET Prompts. Scores leaning negative align with NAME1 and scores leaning positive align with NAME2 in the scenarios. Within a pairing (e.g. W-M), NAME1 refers to the first label (e.g. W) and NAME2 refers to the second (e.g. M).
• Figure 5: Average scores for each model across each relationship type (M=man, N=gender neutral, and W=woman) on GPT-4 generated scenarios. Scores leaning negative align with NAME1 and scores leaning positive align with NAME2 in the scenarios. Within a pairing (e.g. W-M), NAME1 refers to the first label (e.g. W) and NAME2 refers to the second (e.g. M).