Two complaints landed in the US District Court for the Northern District of California on November 6, both against AI music generator Suno, and together they mark a change in how rights holders are fighting generative audio.

The first, filed by Elvis Presley Enterprises, ABG EPE IP, Raleigh Music Publishing and other rightsholders, covers 859 musical works. Applying the statutory maximum of $150,000 per work for willful infringement, the plaintiffs plead $128.85 million, plus $2.15 million for alleged circumvention of access controls and $21.48 million for alleged removal of copyright management information. Separately, under Tennessee's Ensuring Likeness Voice and Image Security Act, they claim damages "believed to exceed $50 million" for use of Presley's voice and likeness, pleaded twice under two state claims. Counting it once, the total exceeds $202 million.

The second, from Eight Mile Style and Martin Affiliated, covers 269 compositions associated with Eminem and pleads just over $47.7 million. Eminem is not a plaintiff.

Why the legal theory matters

For two years, the core question in AI training cases has been fair use: is ingesting copyrighted works to train a model transformative, and does it substitute for the original market? That question is unresolved in the US, and defendants have built their defenses around it.

The Presley complaint routes around it. A performer's recognizable voice is not a federal copyright interest, so copyright preemption does not swallow the claim. Tennessee's ELVIS Act was written for exactly this: unauthorized AI reproduction of a singer's voice becomes independently actionable under state law.

The complaint also attacks Suno's safeguards with specifics. It alleges that even without naming the artist, using style descriptions or spelling the name with spaces between letters, users could generate performances recognizably in Presley's voice. If a filter is that easy to step around, calling it a safeguard is a hard argument to sustain.

Suno's position

The company says its purpose is to help people create original music rather than trade on anyone's name or likeness, that it prohibits prompts naming specific artists or songs, and that it works with third-party providers to screen uploaded audio and lyrics.

The structural weakness is that these controls live at the input. The disputed output comes from weights that already encode a timbre. Keyword blocklists do not fix that; training-data licensing, output similarity detection and traceable usage logs are the layers that do.

A compliance checklist for AI content products

License at design time, not after launch. Data sources, scope of rights, commercial use: assemble a file a lawyer can read.

Separate the three rights. Composition copyright, sound recording rights and a performer's right of publicity belong to different parties. One license does not cover all three.

Screen the output. Similarity comparison and voice-print detection get closer to the problem than prompt filtering.

Keep records. Generation logs, prompts, terms of service, takedown response times: these become mitigation evidence.

Nothing here is adjudicated. But the direction is clear. Risk in generative media is moving from "whose data trained the model" to "does the output resemble a specific person."