Exploring XAI for the Arts: Explaining Latent Space in Generative Music

Making AI systems transparent helps creative collaboration: The paper demonstrates that when AI music generation models are made explainable—allowing users to see which parts of the system control which musical qualities—people can better understand and work alongside the AI as creative partners rather than treating it as a "black box."

Meaningful controls require intentional design: Rather than leaving the AI's internal "latent space" as abstract numbers, the researchers deliberately trained the system so that specific dimensions correspond to recognizable musical attributes (like tempo or instrumentation). This targeted approach makes the system debuggable and allows creators to predict what will happen when they make adjustments.

Real-time feedback loops empower users: By providing a user interface where people can adjust the AI's internal parameters and immediately hear the musical results, the system becomes more interactive and understandable. This direct feedback helps users learn how the AI works through experimentation rather than relying on technical documentation.

Visualization bridges the gap between code and creativity: The paper emphasizes that showing visual representations of how musical attributes map to the AI's internal structure helps non-technical users understand and predict the system's behavior—addressing a key barrier to responsible AI adoption in creative fields.

XAI for the arts remains underdeveloped: The authors note there is "very little XAI for the arts" currently, suggesting that explainability standards common in other AI applications have not yet been widely applied to creative AI systems, representing an important gap in responsible AI development.