On May 5, 2026, the Brazilian Senate's Committee on Communications and Digital Law (CCDD) advanced amendments to PL 2338/2023 — the Marco Legal da Inteligência Artificial — that would require generative AI developers to disclose copyrighted training data and pay remuneration to rights holders. The Escritório Central de Arrecadação e Distribuição (ECAD), Brazil's powerful music collection society, is pressing for mandatory licensing of Brazilian musical works used by OpenAI, Google, and Meta. The political momentum is unmistakable: Brazilian artists, having watched generative tools proliferate, want a piece of the AI economy.
That instinct is understandable. The execution is dangerous.
What PL 2338 Already Got Right
The Senate plenary approved an earlier version of PL 2338/2023 in December 2024, sending it to the Chamber of Deputies. That text — building on Senator Rodrigo Pacheco's original draft — wisely organized AI governance around risk classification, fundamental rights protections, and transparency duties. Article 42 of the December 2024 version already required developers to provide a summary of copyrighted works used in training and to respect opt-outs by rights holders. That approach mirrors the European Union's AI Act, which under Article 53(1)(c)–(d) obliges providers of general-purpose AI models to publish a sufficiently detailed summary of training content and to comply with reservations under Article 4(3) of the 2019 Copyright in the Digital Single Market Directive.
Transparency-plus-opt-out is a defensible equilibrium. It gives rights holders agency without freezing model development behind a wall of pre-clearance. The May 2026 CCDD amendments now risk abandoning that equilibrium for something far more rigid: a compulsory collective licensing regime modeled on ECAD's monopoly over public performance royalties.
The ECAD Model Doesn't Translate to AI Training
ECAD operates under Lei 9.610/98 (Brazil's Copyright Law) and Lei 12.853/2013, which centralized public-performance collection. It works — imperfectly — for a narrow use case: identifying when a song is played in a bar or on the radio and distributing royalties accordingly. Generative AI training is structurally different.
- Training is not performance. A model ingesting a corpus to learn statistical patterns does not reproduce or communicate the work to the public in the sense Article 29 of Lei 9.610/98 contemplates. Treating it as such requires a legal fiction that even European courts have hesitated to adopt.
- Per-work licensing at training scale is mathematically infeasible. Frontier models are trained on trillions of tokens drawn from billions of documents. Bilateral negotiation with every rights holder — or even every collecting society — is not a regulatory inconvenience; it is a categorical bar to entry.
- Compulsory tariffs distort incentives. ECAD has faced repeated antitrust scrutiny from CADE, Brazil's competition authority, including the landmark 2013 ruling that found anticompetitive conduct in tariff-setting. Extending that institution's reach to AI training would compound those concerns at a moment when Brazil is trying to build a domestic AI sector.
What the Comparative Evidence Shows
Jurisdictions that have engaged seriously with AI and copyright are converging on a more proportionate model. Japan's Article 30-4 of the Copyright Act expressly permits text-and-data mining for non-enjoyment purposes, a position the Agency for Cultural Affairs reaffirmed in its March 2024 "General Understanding on AI and Copyright." The United States Copyright Office's 2025 report on Copyright and Artificial Intelligence, Part 3: Generative AI Training declined to recommend a compulsory license, instead emphasizing the role of fair use analysis and emerging voluntary licensing markets. The first substantive U.S. ruling on AI training, Thomson Reuters v. Ross Intelligence (D. Del. 2025), turned on the specific competitive substitution facts of that case — not a categorical rule against training.
Even the EU, often caricatured as maximalist, did not impose mandatory licensing in the AI Act. It chose disclosure plus a robust opt-out, leaving the market to price the rest.
A Proportionate Path for Brazil
Brazil's policymakers should retain PL 2338's strengths and resist the CCDD's licensing overreach. A workable framework would include:
- Granular but bounded transparency. Require developers to publish training-data summaries at the level of dataset categories, languages, and major sources — sufficient for rights holders to assert claims, insufficient to expose proprietary curation methods.
- A statutory opt-out registry. Brazil could create a national machine-readable registry, administered by the Instituto Nacional da Propriedade Industrial (INPI) or the Biblioteca Nacional, where rights holders flag works excluded from training. Compliance would be a safe harbor.
- Voluntary collective licensing, not compulsory. ECAD and similar bodies should be free to negotiate blanket licenses with willing developers — and many will. But mandatory tariffs on training, set by an entity with a documented competition history, invite both constitutional challenge under Article 170 of the Federal Constitution and economic harm.
- A limited TDM exception for research and SMEs. Following Japan's lead, Brazil should ensure that academic researchers and small Brazilian AI firms — exactly the actors least able to absorb licensing costs — can train on lawfully accessed data without prior authorization.
The Stakes for Brazilian Innovation
Brazil has a genuine opportunity in AI. Portuguese-language models, Latin American context, and a young technical workforce are real comparative advantages. The country's Plano Brasileiro de Inteligência Artificial (2024–2028) committed R$23 billion to building that capacity. A compulsory licensing regime that none of Brazil's major trading partners impose would be a self-inflicted wound — handing the Portuguese-language AI market to firms training in jurisdictions with clearer rules.
Creators deserve compensation when AI products substitute for their work in the marketplace. That is properly the domain of output-side liability, competition law, and voluntary licensing. Loading those concerns onto the training stage, through a compulsory regime modeled on a 20th-century performance-rights body, will not protect Brazilian artists. It will simply ensure that the models shaping their cultural future are built somewhere else.