Abstract
As generative AI models for music move toward mainstream adoption, the industry faces a critical challenge: the lack of transparency and fair attribution for original creators. This paper, presented at the NeurIPS 2025 Workshop on AI for Music, proposes a shift from "black-box" generation to a transparent, agentic ecosystem. By integrating Retrieval-Augmented Generation (RAG) with a granular database of musical components, the authors outline a framework where AI-driven creativity and automated artist compensation can coexist.
Introduction: The Attribution Gap in Generative Music
Current generative AI models for music often operate as "single-shot" systems, where a user prompt results in a finished audio file. This paradigm presents two major issues: creative limitation for the user and a significant attribution gap for the artists whose work trained the models. As we enter a "post-streaming" era, existing royalty models are ill-equipped to handle the scale and complexity of AI-driven production.
The authors argue for a structural shift where AI acts as a mediator that tracks provenance from the moment of creation. This requires moving beyond traditional copyright and royalty frameworks toward automated, real-time attribution tracking.
Proposed Architecture: The Music AI Agent
The core of the proposal is a multi-agent architecture grounded in Content-based Retrieval-Augmented Generation (RAG). This system moves away from monolithic generation toward a modular approach involving three main pillars:
- BlockDB: A database where music is stored as "Blocks"—granular components like drum stems or basslines. These are indexed using a two-dimensional ontology of timbral (instrumental role) and temporal (song structure) axes.
- Multi-Agent Orchestrator: A suite of specialized agents, including an Intent Agent to build task plans and a Query Agent to retrieve attributable assets from BlockDB.
- Attribution Layer: A "settlement engine" that logs a "Block usage event" every time a component is retrieved to serve as a condition for new content, enabling transparent royalty distribution.
Building a Collaborative Ecosystem
Instead of a static catalog, the authors reframe music as a collaborative and adaptive ecosystem of reusable parts. The system utilizes a Creator Contribution Workflow where artists submit tracks that are decomposed into indexed Blocks enriched with multi-modal embeddings. This architecture supports a "superfan economy" where fans can remix or co-create using an artist’s official assets, directly supporting them through usage-based micropayments.
Discussion and Future Work
While the framework provides a technical path toward fair AI, several challenges remain. The success of such an ecosystem depends on overcoming the "cold start" problem of incentivizing artist and label adoption. Furthermore, future research must address how to handle royalties when retrieved blocks are heavily transformed by AI, making them unrecognizable from their source. Aligning this Attribution Layer with global copyright registries (ISRC, ISWC) remains a primary goal for ensuring these AI-native royalties are recognized by traditional rights management organizations.
Summary updated