Auditing Multimodal Large Language Models for Contextualized Hate Speech Detection Using Conjoint Experiments

• In Event: 20268 - Culture and Computational Social Science (Co-sponsored by Section on Mathematical Sociology)

Abstract

Multimodal large language models (MLLMs) that process text, images, and other data open new possibilities for automated content moderation by integrating richer contextual information. In this study, we audit five state-of-the-art MLLMs of varying sizes and architectures using conjoint experiments to assess how these models incorporate context when detecting hate speech. We systematically vary the features of synthetic social media posts---including the presence of identity-based slurs and the demographic identity of the author---and compare MLLM evaluations of these posts with human judgments (N=1854). We find that larger, more advanced models generally make evaluations more closely aligned with human evaluations, particularly in discerning how the meaning of a slur depends on the author's identity. However, we identify persistent racial and lexical biases, and explicit prompts to incorporate or disregard context have little impact on performance. Additionally, our analyses also reveal that MLLMs are especially sensitive to visual identity cues when evaluating hate speech. These findings underscore both the promise and limitations of MLLMs for contextualized hate speech detection and illustrate the potential of conjoint analysis as a scalable tool for auditing AI applications in complex, context-sensitive settings.

Author

• Thomas Davidson, Rutgers University-New Brunswick