Performer rights when the performer is a model

Performer rights when the performer is a model

By Ethan Martinez, Head of Sound Design, RadioPro LLC. Legal review by Priya Gupta, Head of Legal Affairs.

When a session singer records a song, three sets of rights come into existence at once. The songwriter owns the composition. The singer owns a performer's right in the performance. Whoever paid for the session owns the producer's right in the recording. The music industry's royalty plumbing is built on that triangle. When the singer is a model, one corner of the triangle goes missing, and the money and the obligations that used to flow through it have to go somewhere else.

This article explains what happens to each corner when the performance is machine-made, and ends with the questions we think a composer or publisher should settle in writing before any of their work is performed this way. It draws on research our legal team carried out across the European Union during the past year. It is a guide, not legal advice, and the position varies by country.

The composition: still human, still protected

Nothing about an AI performance changes who wrote the song. Under the Court of Justice of the European Union's settled test, a work is protected where it is "its author's own intellectual creation," which is the case where the author "was able to express his creative abilities in the production of the work by making free and creative choices." (Infopaq, C-5/08; Painer, C-145/10) Lyrics and a melody written by a person pass that test regardless of what later performs them.

The reverse is also settled. In the United States, the Copyright Office concluded in January 2025 that "if content is entirely generated by AI, it cannot be protected by copyright," and that "prompts alone do not provide sufficient human control to make users of an AI system the authors of the output." (US Copyright Office, Part 2) The D.C. Circuit held in March 2025 that "the Copyright Act requires all work to be authored in the first instance by a human being," and the Supreme Court declined to review that decision on 2 March 2026. (Thaler v. Perlmutter; SCOTUSblog)

The practical consequence is the first thing a rightsholder should understand. A song written by a person and performed by a model is a protected composition in a recording. A song written by a model is, on both sides of the Atlantic, nobody's property. The value of an AI performance rests on the human authorship underneath it. That is why AuralVerse is built to perform lyrics and compositions that people bring to it, and not to write them.

The performer: a right that needs a person

The Rome Convention of 1961 defines performers as "actors, singers, musicians, dancers, and other persons who act, sing, deliver, declaim, play in, or otherwise perform literary or artistic works." (Rome Convention, Article 3) The 1996 WIPO Performances and Phonograms Treaty repeats the list and adds "interpret." (WPPT, Article 2) Neither treaty uses the phrase "natural persons," but every item on the list is a human role, and the prevailing reading is that a performance requires a person performing. The scholars behind the European Union's reCreating Europe study put it this way: "A mere act of activating AI-enabled speech or sound generation (e.g., by pushing the button 'generate') without any further action is not a performance in a sense intended by the international treaties." (Bulayenko, Quintais, Poort and Gervais, Kluwer Copyright Blog, 2022)

This matters because performer rights in Europe carry something no contract can remove. Under the Rental and Lending Directive, Member States must ensure "that a single equitable remuneration is paid by the user" whenever a commercially published phonogram is broadcast or communicated to the public, "and to ensure that this remuneration is shared between the relevant performers and phonogram producers." (Directive 2006/115/EC, Article 8(2)) In several countries that remuneration is collected only by the national collecting society and cannot be waived by the performer. The French Cour de cassation confirmed in December 2019, in a case about background music supplied under Creative Commons licences, that a performer's statutory remuneration is a matter of public order that the performer cannot sign away, and that a catalogue offered to businesses for a fee is a phonogram "published for commercial purposes" whatever the licence says.

Where there is no human performer, there is no performer to remunerate. That is the single most consequential legal fact about machine performance, and it cuts both ways. A business playing a recording performed entirely by a model has no performer's share to account for. A singer whose work is replaced by a model loses a share of income that the law had deliberately made impossible to bargain away.

The producer: a right that survives

The third corner does not disappear. The WPPT defines the producer of a phonogram as "the person, or the legal entity, who or which takes the initiative and has the responsibility for the first fixation of the sounds of a performance or other sounds, or the representations of sounds." (WPPT, Article 2(d)) Nothing in that definition requires a performer. The reCreating Europe authors are direct: "phonogram producers benefit from legal protection regardless of whether the underlying sound was created by a human or AI-generated. There is no threshold for protection other than that for the subject matter to qualify as a phonogram, i.e., it must be a fixation of sound." The same authors note a limit worth remembering: "AI-enabled continuous unfixed music generation and streaming projects are not phonograms," so a live, unfixed generation has no producer right at all.

So a recording performed by a model has an owner. Whoever took the initiative and responsibility for fixing it, which in a publisher's or production company's arrangement with us is normally that company, holds the producer's right, including the reproduction right under the Information Society Directive. (Directive 2001/29/EC, Article 2) Producer rights, unlike performer rights, are commercial rights held by companies and can be licensed, assigned or waived in the ordinary way. A contract can therefore say exactly who owns the master of an AI rendition, and it should.

The voice: the line nobody may cross

The absence of a performer's right in a machine performance does not mean a machine may perform in anyone's voice. That question is governed by a different and rapidly hardening body of law.

  • United States. Tennessee's ELVIS Act added voice to the right of publicity from 1 July 2024, defining voice as a sound "readily identifiable and attributable to a particular individual, regardless of whether the sound contains the actual voice or a simulation of the voice." (Tennessee Public Chapter 588) The federal NO FAKES Act, reintroduced in May, would create a national right against unauthorised "digital replicas" of a person's voice or likeness. (S.1367 text)
  • Denmark. The government has notified the European Commission of amendments to its copyright act that would prohibit realistic digitally generated imitations of a person's characteristics, and separately of a performer's rendition of a work, without consent. The European Parliament's research service notes these are not copyright in the strict sense, because "voice and image considered in isolation" do not meet the originality test. (EPRS, January 2026)
  • European Parliament. On 10 March 2026 Parliament called on the Commission to investigate measures against AI-generated content "imitating their personal characteristics (deep fakes) without consent," expressly including "artists' works and performances," and stressed "the need to clearly label purely AI-generated content." (OEIL summary, 2025/2058(INI))
  • Platforms and societies. Spotify's policy since September 2025 is that "Vocal impersonation is only allowed in music on Spotify when the impersonated artist has authorized the usage." GEMA's AI charter names "the right to one's own voice, one's own name and one's own image" among the moral rights that generative AI must respect. (GEMA AI charter)

Our own rule is simpler than any of them. AuralVerse does not render in the voice of an identifiable living performer, and we decline prompts that ask it to. A performance model that imitates a person is not a performance model. It is a replica, and the law is converging on treating it as one.

Disclosure: what the recording must say about itself

From 2 August 2026 the European Union's AI Act requires providers of systems "generating synthetic audio" to ensure that outputs "are marked in a machine-readable format and detectable as artificially generated or manipulated." (Regulation (EU) 2024/1689, Article 50(2)) Where such content "forms part of an evidently artistic, creative, satirical, fictional or analogous work," the disclosure duty is limited to one "that does not hamper the display or enjoyment of the work." That carve-out was written for exactly the case of a song.

The industry's metadata has caught up. The DDEX release standard that labels and distributors use to deliver recordings now carries a flag for whether a recording, and each contribution to it, was made fully or in part by a generative AI tool. YouTube asks music partners to use it, with three values: fully, partly or no generative AI. (YouTube Help) Spotify displays AI credits by role: lyrics, vocals, instrumental performance, production. An AuralVerse rendition of a human-written song is, in that vocabulary, a recording with human lyrics and composition and AI vocal and instrumental performance. We supply that description and expect our partners to pass it on.

Where the money goes

Collecting societies have begun to answer the question the law leaves open. GEMA's licensing model for generative AI, introduced in September 2024, "provides a 30% share of all net income generated by the generative AI model or system of the provider," plus a share of the income from AI-produced songs that "must be at least equivalent to what would have been provided for purely human-generated works." The licence "is not dependent on where the training takes place; instead, it is linked to the generation of output." (GEMA, generative AI licensing model) SACEM withdrew its repertoire from text and data mining in October 2023, so that use of its works to build AI tools requires its prior authorisation. GEMA's registration practice, as described by its general counsel in March 2026, is that a work can be registered only to human authors, with no deduction for any AI-generated portion.

For a human-written song performed by a model, the author's share is therefore intact and flows as it always did. The performer's share has no beneficiary. The producer's share belongs to whoever owns the master. The one new question is whether the arrangement with the model's owner takes a share of its own, and that is a matter for the contract.

The contract checklist

Before any composition is performed by AuralVerse or any comparable system, we would expect a rightsholder's agreement to answer eight questions. Bloomberg Law's advice to artists in January 2026 was to review existing agreements "for clauses on new technologies, derivatives, text and data mining, machine learning training, voice likeness/right of publicity, and any clauses that let labels or publishers license on their behalf." (Bloomberg Law, 13 January 2026) Our list is the same advice from the other side of the table.

  1. Chain of title in the composition. Who holds the publishing, in which territories, and does that holder have the right to authorise a performance of this kind. Existing publishing and recording contracts may already assign or reserve it.
  2. Ownership of the master. Who takes the initiative and responsibility for the fixation, and therefore holds the producer's right. Name the party.
  3. Scope of the performance licence. Which compositions, which arrangements, which territories, which media, and for how long.
  4. No training grant unless intended. Several European recording and distribution contracts now contain express grants of the right to use recordings to train AI models. A performance licence is not a training licence, and the contract should say so.
  5. Voice. A warranty from the model's operator that no identifiable living performer's voice is imitated, and a prohibition on the rightsholder requesting one.
  6. Moral rights and credit. The author's right to be identified and to object to prejudicial distortion survives in most of Europe and cannot be assigned. State how the author is credited and who approves arrangements.
  7. Disclosure. Which party supplies the AI-use metadata to distributors and platforms, in which standard, and with what values.
  8. Remuneration. How the author's share, the producer's share and any society licence are accounted for, and how the parties handle the performer's share where national law allocates it to a fund rather than to no one.

None of these questions requires the rightsholder to know how the model works, and none of our agreements disclose it. They require the rightsholder to know what they own, what they are licensing, and what the recording will say about itself. Those are the questions the law now asks, and they are answerable in a page.