AI Regulation Tracker / Legislation pending
Brazil's AI bill would make firms pay authors to train AI on their work
The Senate-approved text of PL 2338/2023 would require prior authorization and payment before protected works train an AI. A proposed exception that removes that duty is now the central copyright battle in the Chamber of Deputies.
A coalition of Brazil's creative industries has moved to defend a single principle inside the country's pending artificial-intelligence law: that companies should get permission and pay before they train AI on someone's protected work. In May 2026, according to gov.br/cultura and press reporting, more than 50 industry entities plus ECAD, the country's central music-royalty collection office, delivered a letter to the Chamber of Deputies special commission handling PL 2338/2023. Their demand was narrow and firm. Keep author protection in the bill, and add no exception for AI training.
The fight is over text that already exists. The Senate approved its version of PL 2338/2023, the proposed "Marco Legal da IA," with language that would require prior authorization and compensation to the rightsholder before protected works are used to train an AI system. What alarmed the creative sector is a proposed carve-out, a text-and-data-mining style exception, that would permit training on protected content without consent or payment. The coalition wants that exception rejected.
What the Senate text does
As passed by the Senate, the bill treats training data as something a developer must clear in advance. Using protected works to train AI would require prior authorization from, and compensation to, the rights holder. That is a consent-and-pay rule, closer to licensing than to a broad research or fair-use freedom. The creative federations behind the May letter want to keep exactly this, and they are asking the Chamber not to weaken it with a training exception.
This piece does not assert the specific article number that carries the copyright provision in the current text. That statutory detail is not confirmed here, and reporting on a bill still being amended in committee should not attach a section number it cannot verify. What is verified is the substance: prior authorization and payment in the Senate text, and an active push to keep it intact.
Why the disagreement runs deep
Author remuneration was not an isolated flashpoint. During the special commission's public hearings, three questions drew the most expert disagreement: whether and how authors of training-data works should be paid, whether the national data-protection authority (ANPD) should lead the AI governance body, and how AI risk assessment should work. Copyright sits at the center of that unresolved list, which is part of why the vote has been slow to arrive and why the creative sector treated a single committee letter as worth 50 signatures.
What it does not do yet
Nothing here is law. PL 2338/2023 is still moving through the Chamber of Deputies, and the copyright provision could be preserved, softened, or replaced before any final vote. There is no effective date, no enforcement mechanism, and no penalty schedule to point to. A professional advising a client on Brazilian training data cannot yet say a licensing duty exists. What they can say is that the default the Senate chose is paid authorization, and that removing it now requires an affirmative legislative act that the creative sector is actively contesting.
The cross-border read
For a US reader, the contrast is the story. The same clash, whether developers may train on protected works without permission, is being fought in US courtrooms, most visibly in The New York Times' litigation against OpenAI, where the answer will emerge case by case through fair-use doctrine. Brazil may resolve the identical question the opposite way, by statute, with a mandatory licensing rule written into a national AI law. A firm operating in both markets could face judge-made fair-use analysis in the United States and a codified pay-to-train regime in Brazil. US rightsholders who want a legislative template rather than a litigation gamble will watch which way the Chamber votes.
Frequently Asked Questions
What changed in Brazil's AI bill regarding AI training on copyrighted work?
Nothing is enacted yet. The Senate-approved text of PL 2338/2023 requires prior authorization and payment to rightsholders before protected works are used to train AI. In May 2026, more than 50 creative-industry entities plus ECAD asked the Chamber of Deputies to keep that rule and reject a proposed exception that would allow training without consent or compensation.
Who is affected by this provision?
AI developers that train or fine-tune models on Brazilian protected works, and the rightsholders who own those works, including authors, performers, publishers, and collecting societies. If the paid-authorization model becomes law, sourcing training data in Brazil becomes a licensing question rather than an open one.
Is this a binding law right now?
No. PL 2338/2023 is a bill still under debate in the Chamber of Deputies. There is no effective date and no enforcement mechanism. The Senate has approved its version, but the Chamber can still amend the copyright language before any final vote.
How does this compare to what is happening in the United States?
The United States is resolving the training-data question through litigation, such as The New York Times v. OpenAI, under fair-use doctrine. Brazil may resolve the same question by statute, through a mandatory authorization-and-payment rule. That would make Brazil a licensing jurisdiction rather than a case-law one.
Sponsored Training
Browse the full AI Regulation News tracker
Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.