A Review of Multimodal Explainable Artificial Intelligence: Past, Present and Future

AI systems are becoming harder to understand as they grow more powerful: The paper identifies a critical challenge where advances in computational power and massive datasets have made modern AI models increasingly difficult to interpret, creating a "black-box" problem that undermines human trust in AI decision-making—a concern that has only intensified with the rise of Large Language Models.

Explainable AI is essential for responsible deployment: The review emphasizes that eXplainable AI (XAI) methods focusing on transparency and interpretability are necessary to enhance human understanding and trust, suggesting that explainability should be a core requirement rather than an optional feature when deploying AI systems in real-world applications.

Multimodal AI adds complexity to the explainability challenge: When AI systems process multiple types of data simultaneously (text, images, audio, etc.), explaining their decisions becomes significantly more complex, requiring specialized Multimodal eXplainable AI (MXAI) approaches that can account for how different data types interact in the reasoning process.

Different AI eras require different explanation approaches: The paper categorizes MXAI development across four distinct periods—traditional machine learning, deep learning, discriminative foundation models, and generative LLMs—indicating that explanation methods must evolve as AI technology fundamentally changes, and no single approach works across all AI systems.

Building trustworthy AI requires ongoing research and standardization: The review's discussion of evaluation metrics and datasets suggests that creating truly transparent, fair, and trustworthy AI systems requires continued development of standardized methods to measure and validate explainability across different AI applications.