Abstract
This study investigates gender bias in multimodal AI training data by analyzing occupational representations in large-scale image–text datasets. We reveal strong bimodality in bias scores across occupations, with roles such as nurse skewing feminine and soldier skewing masculine. To address this, we propose a novel sentence-level bias quantification method and an embedding-space mitigation strategy that preserves semantic meaning. Using Iterative Nullspace Projection (INLP), we reduce occupational disparities, and experiments with DistilGPT-2 show how caption-level biases propagate into generated outputs and how targeted mitigation can alleviate them. The findings provide methodological tools and practical insights for researchers building fairer NLP systems, industry practitioners designing inclusive platforms, and hiring professionals seeking to avoid gender-coded job postings.
Recommended Citation
Kwan, Kin and Yang, Xingwei, "AI Training Sets Fair? Evidence of Occupational Gender Bias in the Multimodal Dataset" (2025). Proceedings of the 2025 Pre-ICIS SIGDSA Symposium. 36.
https://aisel.aisnet.org/sigdsa2025/36