Large-Scale Training Data Attribution for Music Generative Models via Unlearning

AI music generators can now identify which artists' work influenced specific outputs: The paper demonstrates that "unlearning-based approaches can be effectively adapted to music generative models," meaning researchers can trace which training data contributed to generated music—a capability previously unavailable at scale for this technology.

This addresses a major copyright and credit gap in AI music: The authors highlight that "proper recognition and credit for original artists are generally overlooked" in AI-generated music. By identifying source training data, this work creates a pathway for fairly attributing and potentially compensating artists whose work trained these systems.

The method works by strategically "unlearning" data to measure its importance: The approach uses unlearning techniques to determine which training examples matter most to a model's outputs—essentially removing data and observing what changes, which reveals each artist's contribution to the final generated music.

Validation shows the approach is reliable and consistent: The researchers performed "grid search over different hyperparameter configurations and quantitatively evaluate[d] the consistency of the unlearning approach," indicating the method produces dependable results that could be trusted for real-world attribution decisions.

This is a step toward more accountable AI systems: The paper frames this work as "paving the way for more ethical and accountable AI systems for music creation," suggesting it could support future policies requiring transparency about which artists' work influenced AI-generated content.