Every Song Leaves a Trail

Preserving the Connection to Original Music

‍

Our 2025 project with NYU centered on a broader attribution system that included new generative music models and an evaluation framework. The work examined whether a generated song retains a recognizable relationship to known music, a necessary part of understanding this relationship between training data and outputs.

At UC Riverside, Computer Science and Design teams studied generative music models alongside attribution frameworks, examining the connection between source music, the model, and what the model produces. Cal Poly explored new generative-AI music applications designed to carry provenance from the start, so the relationship to creative sources does not have to be reconstructed only after a song is generated.

‍

Testing What the Evidence Can Show

‍

With UC Berkely joining in 2026, benchmarking more approaches side by side to test what each method can reliably show about the relationship between an output and original works.

Sound Ethics AI-detection research plays an important part in the process as well as in better understanding the relationship between training data (human songs) and generated music.

Together, the projects examine different parts of the same chain. Historical research identifies music used in the development of AI. Provenance research asks how links to creative sources can be carried forward. Attribution compares generated songs with original works, the benchmarking tests the methods, while detection remains an important link in this chain revealing more about AI's footprint and how people are using AI in new music production.

Each project helps clarify a different part of the relationship between specific songs, recordings, and the systems built from and around them.

‍