AuditMM: A Framework for Auditing Bias and Interpretability in Multimodal AI Systems

Multimodal AI systems (combining text, image, audio, etc.) are increasingly used in high-stakes domains, but are under-audited for fairness, transparency, and bias amplification across modalities. We propose AuditMM, a high-level framework for systematically auditing multimodal systems: covering data, preprocessing, feature extraction & encoding, fusion, decision output, and post-deployment monitoring. AuditMM includes synthetic and counterfactual benchmarks, modality-specific and fused-model metrics, and diagnostic tools to locate sources of bias. We present a small experiment applying AuditMM on a benchmark text+image classification task, showing how fused models can amplify disparity relative to single-modality baselines, and demonstrate probing and attention-weight diagnostics. Our results indicate actionable insights for mitigation of bias in the multimodal training and auditing pipeline.