Examining Gender and Racial Bias in Large Vision-Language Models Using a Novel Dataset of Parallel Images

TL;DR

This work introduces PAIRS, a novel parallel-image dataset designed to isolate gender and racial cues in LVLMs by keeping scenarios visually identical while varying demographic attributes. Using four state-of-the-art LVLMs, the authors probe bias through binary occupation/status/crime prompts and open-ended prompts analyzed with PMI-based lexical methods. Key findings include significant gender bias in occupation associations and racial bias in status judgments, with open-ended text revealing intersectional stereotypes; crime-related prompts showed no consistent racial bias. The study highlights the need for robust bias-mitigation strategies and multi-faceted evaluation to ensure safe, fair deployment of multimodal models. It also discusses limitations and ethical considerations related to dataset scope, prompts, and generation costs.

Abstract

Following on recent advances in large language models (LLMs) and subsequent chat models, a new wave of large vision-language models (LVLMs) has emerged. Such models can incorporate images as input in addition to text, and perform tasks such as visual question answering, image captioning, story generation, etc. Here, we examine potential gender and racial biases in such systems, based on the perceived characteristics of the people in the input images. To accomplish this, we present a new dataset PAIRS (PArallel Images for eveRyday Scenarios). The PAIRS dataset contains sets of AI-generated images of people, such that the images are highly similar in terms of background and visual content, but differ along the dimensions of gender (man, woman) and race (Black, white). By querying the LVLMs with such images, we observe significant differences in the responses according to the perceived gender or race of the person depicted.

Figures (5)

• Figure 1: Sample images from the Occupations subset. In the first row, we ask whether the person is a doctor or a nurse; in the second row, we ask whether the person is a pilot or a flight attendant; and in the third row we ask whether the person is an architect or an event planner.
• Figure 2: LVLMs tend to label images of men as the male-dominated occupation (positive association score), and images of women as the female-dominated occupation (negative score). The differences are statistically significant for all four models ($p < 0.05$).
• Figure 3: In three out of four cases, LVLMs are more likely to label images of white people as higher-status (positive score) and Black people as lower-status.
• Figure 4: There are no differences in the association scores for criminality (positive values indicate the neutral or positive interpretation; negative values indicate the criminal interpretation).
• Figure 5: Sample images of a person in an orange prison jumpsuit from the open-ended questions data.