AI music attribution aims to identify the recordings and compositions that influenced a generated song. Its findings could inform licensing and payment, but identifying an influence does not establish permission or determine what a creator is owed.
What attribution can show
Several related tools answer different questions. Audio fingerprinting seeks an exact or recognizable match, AI detection assesses whether music was generated by AI, and provenance records a file’s origin and changes. Dataset disclosure identifies training material. Attribution attempts to estimate which examples influenced an output and by how much.
Attribution can examine either the training process or a particular generation. A system may be able to document a reference recording, uploaded track or licensed artist model supplied for a specific output. Tracing the influence of training material is harder because a model may have learned from hundreds of thousands or millions of examples, and methods can yield different results.
Researchers including Choi and colleagues and Morreale and colleagues have proposed methods for tracing training influence. No universally accepted industry method produces consistent results across different models.
Warner Music Group agreed to acquire Sureel AI in 2026. Warner said Sureel’s registry contains millions of music assets and can track the use of components of a work in AI systems. Such tracking could provide evidence of use, but permission requires a license, while payment depends on a contract or law.
Scale and accounting
Meta disclosed that its MusicGen research model was trained on roughly 400,000 recordings, representing about 20,000 hours of music owned by Meta or licensed for the project. Many commercial AI services do not publish a complete training manifest, the number of model versions they have created or their users’ generation totals.
In April 2026, Deezer reported receiving almost 75,000 fully AI-generated tracks daily, more than two million a month and approximately 44 percent of its daily uploads. It said those tracks accounted for only 1 to 3 percent of streams and that as much as 85 percent of those streams were fraudulent. The figures cover one streaming service and exclude generations kept within AI platforms, discarded, revised or used privately.
Any payment proposal must account for both the number of works used in training and the number of outputs a model produces. A proposed percentage alone does not show what an individual creator would receive.
Separate rights in a song
A recorded song normally contains at least two copyrighted works: the musical composition, which covers music and lyrics, and the sound recording, which covers the recorded performance and production. The U.S. Copyright Office treats them separately, so permission for one does not automatically cover the other.
Recordings and compositions can each have multiple rights holders, with ownership shares that vary by territory or remain disputed or missing. A model may also learn from a singer’s phrasing, an instrument’s sound, production choices or a recording’s timbre. Voice, name and likeness can involve rights outside the composition and sound recording licenses.
