Explainable Generative AI: Concepts, Challenges and Future Directions

Large language models (LLMs), diffusion-based image generators, speech synthesizers, code assistants, and multimodal foundation models are examples of generative AI systems that are now deeply integrated in scientific research, creative work, and decision-making. Understanding why these models produce specific outputs has become crucial for safety, accountability, trust, debugging, and regulatory compliance as they continue to acquire autonomy and agency. However, explainability for generative AI is significantly more challenging than explainability for traditional discriminative models, because the generation unfolds over time, the output space is open-ended, and the model internals operate in high-dimensional spaces. Explainable Generative AI (XGAI) is the new interdisciplinary field examined in this paper. The paper defines the scope and fundamental concepts of XGAI, presents a taxonomy of interpretability techniques for generative models, examines technical, social, and regulatory challenges, and outlines future research directions toward transparent and controllable generative systems. This work argues that explainability for generative AI requires coverage of intent, process, uncertainty, and provenance and cannot be reduced to local feature attribution alone.