This work, developed by the Universität der Bundeswehr München, investigates the fairness and bias implications of multimodal fusion techniques in AI-driven recruitment systems using the FairCVdb dataset. While fairness-aware learning has been extensively studied for individual modalities like tabular data, images, and text, multimodal AI remains underexplored. The integration of multiple modalities introduces unique challenges, such as compounding biases, alignment complexities, and imbalanced representations. We focus on two widely used fusion techniques: early fusion, where modalities are combined before processing, and late fusion, where each modality is processed separately before merging outputs.

Our results show that early fusion closely matches the ground truth, achieving lower Mean Absolute Errors (MAEs) by effectively integrating each modality’s unique characteristics. In contrast, late fusion produces more generalized predictions and higher MAEs, potentially leading to fairness concerns due to oversimplified decision boundaries. These findings highlight the advantages of early fusion for more accurate and fair AI applications, even in the presence of demographic biases. Future research could explore alternative fusion strategies and modality-specific fairness constraints to enhance fairness in multimodal AI systems.