Bias in Multimodal AI:
Representation, Risk, and Repair

Multimodal foundation models that integrate vision, language, and time-series data are redefining AI capabilities but also exposing new forms of bias and representation inequity. From gendered stereotypes in image generation to cultural bias in language-vision alignment, these systems can unintentionally reinforce social inequalities embedded in their training data.

The Bias in Multimodal AI: Representation, Risk, and Repair workshop aims to advance understanding of how bias emerges, propagates, and can be mitigated in multimodal AI systems. We invite researchers, practitioners, and policy experts to submit work that explores bias from technical, social, and ethical perspectives, fostering dialogue toward fairer and more inclusive AI.

Submissions may include, but are not limited to:

• Bias propagation across modalities (text, image, audio, time-series)

• Case studies of stereotypical outputs and misrepresentation in MLLMs

• Auditing frameworks and benchmark datasets for multimodal bias evaluation

• Inclusive and representative dataset construction for foundation models

• Prompt engineering and parameter tuning for bias reduction

• Community-driven and participatory design for responsible multimodal AI

• Explainability and transparency methods for bias detection

• Ethical and policy considerations for fair deployment of multimodal AI

• Multidisciplinary and cross-cultural perspectives on representation and fairness

If you have any questions, please contact Dr. Caren Han (caren.han@unimelb.edu.au)