Background: The advent of Multimodal Artificial Intelligence has been accelerated by contrastive approaches to self-supervised learning, enabling systems to learn rich, robust feature representations without the need for expensive manual labeling. However, these "black box" models often produce high-dimensional latent spaces that are opaque to human interpretation, posing significant risks in high-stakes environments such as healthcare and criminal justice.

Methods: This study proposes a theoretical framework that bridges the gap between unsupervised contrastive learning and Explainable AI (XAI). We integrate principles of Granular Computing and Fuzzy Set Theory to impose interpretable structures upon the latent feature spaces generated by contrastive losses. Furthermore, we apply the National Institute of Standards and Technology (NIST) principles of explainability to evaluate the ethical standing of these systems.

Results: Our analysis demonstrates that while contrastive methods maximize feature richness, they often sacrifice semantic clarity. By applying granular modeling, we show that continuous feature vectors can be discretized into interpretable "information granules," thereby allowing for post-hoc explainability without retraining the foundational model. We further analyze the impact of confidence calibration on user trust.

Conclusions: We conclude that learning rich features without labels is viable for critical systems only when paired with robust XAI mechanisms. The integration of granular computing provides a mathematical foundation for extracting meaning from unlabeled data. We advocate for a "human-in-the-loop" governance model to ensure that contrastive AI systems remain ethical, transparent, and socially responsible.