Can Collective Management Presume Copyright in AI-Generated Music? Lessons from Brazil

On 31 July 2025, in Spitz Park v ECAD (Agravo de Instrumento No. 5032376-37.2025.8.24.0000), the Court of Justice of Santa Catarina, a state in southern Brazil, considered whether a collective management organisation may charge for the public performance of music said to have been generated entirely by AI. ECAD (the Central Collection and Distribution Office) is Brazil’s private, non-profit body responsible for licensing public performances of music and collecting and distributing royalties. Spitz Park Aventuras argued that it used only tracks made with Suno, an AI music generator, that those tracks were not in ECAD’s database, and that no royalties were due. ECAD replied that Suno may have been trained on protected recordings and submitted a technical report finding significant similarity between one generated track and a pre-existing work. The court upheld the refusal to suspend the charges.

The order did not find that AI-generated music is protected by copyright. This issue arose at the interim stage and the court stressed the need for evidence and full adversarial proceedings on this matter. Yet the court’s reasoning raises a broader concern, as it shows that rules developed to facilitate collective management may displace the logically prior question of whether the material being performed is protected at all. An evidentiary presumption in favour of ECAD’s authority risks becoming, indirectly, a substantive presumption of copyright.

What the Court Decided (and What It Did Not)

The court relied on two principal considerations. First, it noted ECAD’s statutory role in collecting and distributing royalties for the public performance of music. Brazilian case law allows ECAD to collect without identifying each work performed or proving that every relevant right holder is affiliated with a member association. The court treated that case law as supporting continuing collection while the dispute is investigated.

Second, the court considered the uncertainty surrounding Suno’s technology. The decision referred to litigation brought by major record companies against Suno and Udio in the United States (UMG Recordings v Suno, D. Mass.; UMG Recordings v Uncharted Labs, S.D.N.Y.), the possible unauthorised use of protected recordings in training, and ECAD’s similarity report. It then stated that the absence of an identifiable human author does not, by itself, imply the absence of copyright or of obligations arising from public use.

The court’s caution at the interim stage was understandable. The concern lies in the movement between distinct legal questions. The rule that ECAD need not identify every work is an enforcement mechanism designed for situations in which protected repertoire is publicly performed. It does not establish that a particular AI output is a protected musical work or phonogram.

Allegations about training do not answer that question either. A model may have been trained unlawfully even when a particular output is non-infringing. Conversely, a particular output may reproduce protected expression regardless of whether the training process was lawful. Similarity involving one track may justify an infringement inquiry, but it does not show that all Suno outputs are protected, belong to ECAD’s repertoire, or generate royalties that ECAD can meaningfully distribute.

The result is a category error with institutional consequences: copyrightability, infringement and collective-management entitlement are treated as though uncertainty about any one of them were enough to preserve all three.

A Threshold Other Systems Address First

Recent approaches elsewhere differ, but they generally identify a doctrinal bridge between an AI output and copyright before moving to economic rights. The U.S. Copyright Office’s 2025 report protects only human-authored expressive elements; Thaler v Perlmutter confirmed that an autonomously generated work cannot have an AI author; and Zarya of the Dawn protected the human text and arrangement of a comic book while excluding its generated images.

The UK takes an expressly statutory route by attributing authorship of a computer-generated work to the person who made the arrangements necessary for its creation, although the UK Government has acknowledged the tension between that provision and an originality standard associated with human qualities. In China, Li v Liu emphasised the user’s prompts, choices and intellectual investment. The Munich Local Court, in its judgment of 13 February 2026 (142 C 9786/25), rejected protection where the user’s instructions did not sufficiently determine the final expression.

These approaches produce no universal answer. Some focus on expressive control, some on personal intellectual choices, and the UK relies on statutory attribution. Each nevertheless attempts to explain why an output is protected and who may claim rights in it. The Brazilian decision moved towards collection without identifying an equivalent bridge under Brazilian law. This is despite the fact that Article 7 of the Brazilian Copyright Act describes protected works as “creations of the spirit”, making the underlying conception of human creation unusually explicit.

Could Neuroscience Supply the Missing Bridge?

Could science tell us whether a “creation of the spirit” extends to AI-generated output? As I have explored elsewhere in Portuguese, Antonio Damásio’s critique of Cartesian dualism offers a reason to resist treating human creation as abstract information processing. For Damásio, decision-making is not purely rational: it depends on bodily signals, emotion, memory and experience. Applied to authorship, this supports caution before equating a plausible machine output with human expression merely because the two look or sound alike. A Suno track that sounds convincing is not, for that reason alone, a “creation of the spirit”.

Lisa Feldman Barrett complicates the idea of emotion as a stable inner cause of action. In her theory of constructed emotion, emotional experience emerges through prediction, bodily regulation, learned concepts, language and context. If emotions are interpreted and categorised through those processes, it becomes difficult to claim that a determinate emotion transparently caused a particular artistic choice.

Daniel Kahneman’s work on bounded rationality, heuristics and bias creates a different tension. Human choices may reflect anchors, framing and influences that decision-makers do not experience as reasons. Authorship therefore cannot be romanticised as the sovereign expression of an autonomous inner self. This does not make human and machine processes equivalent. Rather, it shows that “human contribution” is normatively important but scientifically less self-explanatory than copyright doctrine sometimes assumes. What copyright can do is make explicit, and defend, the assumptions it adopts when deciding which forms of human involvement count as authorship, precisely the step the Santa Catarina decision skipped.

The Economic Question the Decision Leaves Open

If ECAD may collect whenever AI-generated music is publicly performed, who should ultimately receive the money? The user who wrote the prompts, Suno, the authors and producers whose recordings may have been used in training, or the owners of specific works reproduced in a particular output? If no protected work and no entitled right holder can be identified, the payment begins to resemble a levy on AI-generated music rather than a royalty attached to a recognised copyright.

The opposite outcome also has consequences. If commercial venues can avoid collective-management charges by replacing licensed repertoire with low-cost AI tracks, the rule may encourage substitution away from human music and weaken existing remuneration systems. That is a serious economic concern. Presuming that every generated track is protected or derived from ECAD-managed repertoire does not resolve it.

A sound legal analysis should proceed sequentially. It should identify the relevant subject matter (musical composition, phonogram, performance, or some combination), ask whether it is protected and who holds the relevant rights, determine whether a specific output infringes an identified work, and only then address ECAD’s authority to collect and distribute. If policymakers instead want a compensatory levy for AI-generated music because of its market effects or dependence on human culture, that proposal should be defended openly and designed with rules on beneficiaries and distribution.

Generative AI forces copyright to confront the fact that authorship is more than a doctrinal label: it organises cultural production by deciding whose choices count, whose labour is rewarded, whose control is enforceable and what remains available for others to use. A sound framework keeps four inquiries separate: the meaning of creation; the attribution of expressive features to users, automated processes, training materials or later editing; the appropriate legal regime; and the distribution of gains and losses under each possible rule.

The provisional posture of Spitz Park v ECAD justified caution. It did not require the court to collapse authorship, training, derivation and public-performance liability into a single presumption. Collective-management presumptions are tools for enforcing existing rights; they should not create those rights by default.

The case does not establish that AI-generated music must be royalty-free or that ECAD must always be paid. Its central lesson concerns the passage from technological uncertainty to copyright liability, which requires both a legal bridge and an account of economic consequences. Granting or denying protection shapes markets, incentives, bargaining power, cultural access and the public domain. Without that analysis, familiar institutions may silently decide questions that copyright doctrine, neuroscience and public policy have not yet answered.