Sony Builds Tech to Trace Your Songs Inside AI Music

Sony Cracks the AI Music Attribution Problem

Sony Group has unveiled a groundbreaking technology that could reshape the relationship between human musicians and artificial intelligence. Developed by Sony AI researchers, the system uses a technique called machine unlearning to identify which existing songs most influenced a piece of AI-generated music, effectively creating a traceable lineage from machine composition back to human creativity.

How It Works

The approach, dubbed training data attribution, works by selectively making an AI model forget specific songs from its training data and then measuring how the model's output changes. If removing a particular track significantly alters what the AI produces, that track likely had a strong influence on the generated music. The researchers tested their method on a latent diffusion model trained on one hundred and fifteen thousand high-quality music tracks spanning diverse genres.

Why It Matters

The research, accepted at the NeurIPS twenty twenty-five Creative AI Track, represents the first work exploring attribution on a text-to-music model trained on a large and diverse dataset. It arrives at a critical moment for the music industry, where AI-generated content is proliferating rapidly and artists have had little recourse when their work is used without permission or compensation.

Industry-Wide Push for Protection

Sony Music Group has previously warned more than seven hundred technology companies against unauthorised use of its catalogue to train AI systems. The company has also partnered with SoundPatrol alongside Universal Music Group to deploy neural fingerprinting technology for detecting copyright infringement in AI-generated works. Meanwhile, major labels have struck licensing deals with AI music startups Suno and Udio, with both companies agreeing to retrain their models exclusively on licensed content in twenty twenty-six.

Looking Ahead

The attribution technology is part of Sony AI's broader Creative and Protective AI for Music and Entertainment research programme, which also includes methods for recognising musical relationships between works and benchmarking audio watermarking resilience. As licensed AI music creation services prepare to launch, tools like these could form the backbone of a fairer, more transparent AI music ecosystem.