AI music generators produce audio recordings from text prompts, in the manner of Suno and Udio. The policy questions they raise differ from those in text generation in two respects: the training corpus consists of sound recordings whose copyright ownership is concentrated in a small number of rightsholders, and the output competes in the same market as the inputs.
Proving training-data composition from outputs
Because developers have declined to disclose training data, the litigation has had to establish copying by inference from outputs. The UMG v. Suno complaint sets out three lines of evidence, and the third is the most distinctive.
Targeted prompting. Prompts specifying "subject matter, genre, artist, instruments, vocal style" produced outputs "that closely matched the targeted copyrighted sound recording," on the argument that "this degree of similarity in output would be impossible if Suno were not training on the Copyrighted Recordings."
Transcribed comparison. The complaint attaches transposed side-by-side musical transcriptions, colour-coding notes matching the original in both pitch and rhythm against those matching one but not the other — importing melodic-similarity analysis from conventional infringement practice.
Producer tags. The strongest line concerns artefacts with no expressive function. An output titled "Rains of Castamere" opened with the "CashMoneyAP" producer tag though no prompt referenced the producer, and another reproduced Jason Derulo's spoken-name tag. A tag is not a stylistic feature that could be independently arrived at; its presence is evidence that the recordings carrying it were in the training set.
The market-substitution argument
The complaint's harm theory is that synthetic outputs "could saturate the market with machine-generated content that will directly compete with, cheapen, and ultimately drown out the genuine sound recordings on which the service is built." That framing — the output competing with its own inputs — recurs in the text and publishing cases (Hachette et al. v. Google (Gemini training data)), but is sharper here because a generated track substitutes for a specific recording in a way a generated paragraph rarely does for a specific book.
The fair-use posture
The labels' pre-emptive fair-use argument is framed as a category distinction rather than a factor-weighing exercise: fair use "promotes human expression by permitting the unlicensed use of copyrighted works in certain, limited circumstances, but Suno offers imitative machine-generated music—not human creativity or expression." The plaintiffs also read the defendant's own assertion of fair use as conceding the structure of the claim, since the defence "only arises as a defense to an otherwise unauthorized use of a copyrighted work."
The labels position themselves as licensors rather than opponents, citing existing catalogue licences to streaming, user-generated content, social, fitness, gaming, and metaverse platforms.
Relationships
- instance-of: AI Copyright
- related: UMG Recordings v. Suno (AI music training data) — the principal test case
- related: Suno, Recording Industry Association of America (RIAA), Media, Journalism & Entertainment — AI Deployment, Complaint, UMG Recordings et al. v. Suno, Inc. (D. Mass., June 24, 2024), Web Scraping for AI Training