Brazil's AI Law (PL 2338/2023): What 199 Amendments Reveal About the Real Regulatory Battle

TL;DR: Brazil's Projeto de Lei 2338/2023 is a risk-based AI regulatory framework modeled partly on the EU AI Act. It prohibits AI systems that violate fundamental rights, mandates transparency and human oversight for high-risk applications, and creates a liability framework for operators. The bill attracted 199 proposed amendments in the Senate, reflecting fierce disagreement over definitions, oversight authority, and the scope of prohibitions. Organizations with Brazil exposure need to map their AI deployments to the bill's risk tiers before enforcement tightens.

What PL 2338/2023 Actually Is

Brazil's Projeto de Lei 2338/2023 is a proposed federal law establishing a comprehensive framework for the development, deployment, and oversight of artificial intelligence systems. Senator Rodrigo Pacheco introduced it in May 2023, and it has since become the primary vehicle for AI governance in a country of 215 million people, the largest economy in Latin America.

The law takes a risk-based approach. It classifies AI systems into four tiers: excessive risk (prohibited), high risk (strict requirements), limited risk (transparency obligations), and minimal risk (largely unrestricted). This architecture resembles the EU AI Act, though the details diverge in ways that matter for compliance teams.

At its core, the bill is about fundamental rights. Brazil's constitutional tradition puts individual rights at the center of regulatory design, and PL 2338/2023 reflects that directly. Systems that manipulate human behavior, exploit vulnerable populations, or enable government social scoring fall in the prohibited category without exception.

The Four Risk Tiers: What They Actually Require

Excessive risk (prohibited systems): These are flatly banned. The category includes AI that performs subliminal manipulation causing harm to individuals, AI used by public entities for social scoring based on personal behavior, and systems that exploit the vulnerabilities of children, elderly persons, or people with disabilities. Real-time biometric identification in public spaces also falls here, though the exact scope of that provision was among the most contested in the amendment process.

High-risk systems: This is where most of the compliance burden lives. HR and employment decisions, credit and insurance scoring, critical infrastructure management, educational assessment, administration of justice, border control, and law enforcement applications all qualify. Operators must conduct algorithmic impact assessments, maintain technical documentation, implement human oversight mechanisms, and allow individuals to contest automated decisions that affect them.

Limited risk: Mainly chatbots and synthetic media. The core obligation is transparency: users must be told they are interacting with an AI system. Deepfake-generated content requires clear labeling.

Minimal risk: Everything else. No specific AI-law obligations apply beyond general legal compliance.

The tiered structure sounds clean on paper. In practice, classifying a specific system requires judgment calls that the bill's definitions do not always resolve. That ambiguity is exactly why 199 amendments landed.

199 Amendments: What the Debate Reveals About the Hard Problems

The 199 proposed amendments submitted during Senate review tell you more about the real difficulties of AI regulation than the bill text itself does. Reading them as a body, five fault lines stand out.

Definitional battles. A large bloc of amendments targeted the definitions of "artificial intelligence system," "operator," and "high-risk." Getting those definitions wrong either creates regulatory gaps or sweeps in systems no one intended to regulate. The EU AI Act went through years of similar definitional fights before its final text settled.

The oversight authority question. Significant disagreement arose over whether ANPD (Brazil's national data protection authority, which enforces the LGPD) should serve as the primary AI regulator, or whether a new dedicated authority should be created. ANPD has existing enforcement infrastructure. Critics argued it lacks the technical capacity for AI-specific oversight. Some amendments pushed for a multi-body structure; others wanted a single new agency. The resolution of this question determines who issues guidance, who conducts audits, and who imposes fines.

Biometric surveillance. Real-time facial recognition in public spaces generated the sharpest conflict. Industry groups and law enforcement advocates pushed for broad carve-outs. Civil society organizations pushed for near-total prohibition. The final configuration of this provision carries real operational consequences for smart-city vendors, security integrators, and retailers using in-store analytics.

Generative AI and foundation models. PL 2338/2023 was drafted before generative AI dominated the public conversation. Several amendments tried to patch in explicit rules for large language models, synthetic media, and AI-generated content. Debate over whether foundation model developers bear liability for harms caused by downstream applications did not produce consensus. This gap mirrors arguments still happening in Brussels and Washington.

Innovation sandboxes. Brazilian technology industry groups lobbied for regulatory sandbox provisions allowing startups to test AI systems under relaxed conditions. Some amendments incorporated sandbox language; others argued it created unacceptable consumer protection gaps. The sandbox debate is not unique to Brazil, but the stakes are higher in a market where most AI startups lack the compliance infrastructure of a large European firm.

199 amendments is a high count by Brazilian legislative standards. It signals that no major stakeholder group found the original text fully workable. That is not unusual for first-generation AI legislation anywhere. What it means practically is that practitioners cannot treat any early summary of this bill as stable compliance guidance.

How Brazil Compares to Other Major AI Regulatory Frameworks

No two significant AI frameworks are identical. That creates real compliance complexity for any organization operating across jurisdictions.

EU AI Act: The most detailed framework currently in force. It requires conformity assessments for many high-risk systems, involves accredited notified bodies, and includes a General-Purpose AI tier for foundation models above a compute threshold. Brazil's bill is structurally similar but lighter on conformity assessment mechanics. An AI system fully compliant with the EU AI Act will satisfy most but not all of Brazil's high-risk obligations. Impact assessment formats, documentation standards, and transparency disclosure language will need Brazil-specific versions.

United States: Still no federal comprehensive AI law. The U.S. approach remains sector-specific: FDA rules for AI in medical devices, CFPB guidance on algorithmic credit decisions, EEOC guidance on automated hiring tools. This means U.S.-designed compliance programs are not structured to map onto Brazil's horizontal framework. Organizations that built their AI governance around U.S. sector regulators will need to add a new layer for Brazil.

United Kingdom: Principles-based and deliberately light-touch. The UK chose not to pass a standalone AI law, instead distributing regulatory responsibility to existing sector regulators under shared principles. Lower compliance burden than Brazil, but also lower legal certainty for operators.

China: Enacted specific regulations for generative AI (2023), algorithmic recommendations (2022), and deep synthesis (2022). China's rules are operator-facing and require security assessments and government filings for services offered to Chinese users. Brazil's territorial reach is narrower, but the substantive rights obligations are broader.

OECD AI Principles: Voluntary and widely referenced as a drafting source. Brazil's bill draws on them, as does the EU AI Act. They are useful for benchmarking internal governance frameworks but carry no enforcement weight.

How to Keep Pace With AI Regulation as It Evolves

The uncomfortable reality is that AI regulation is not going to stabilize. Brazil's law will generate secondary implementing regulations after enactment. The EU AI Act is already producing delegated acts. The U.S. regulatory landscape shifts with each administration. Any compliance program built around a single snapshot of any framework will be outdated within months.

Three actions that actually hold up over time:

1. Build a living AI inventory. Know what AI systems you operate, in which jurisdictions, and what decisions they affect. Without this, you cannot map to risk tiers under any framework, and you will miss new regulatory triggers as they emerge.

2. Write modular compliance documentation. Impact assessments, transparency notices, and human oversight procedures written in a modular format can be adapted to different jurisdictional requirements faster than rebuilding from scratch. A single monolithic compliance document is a liability.

3. Track amendment outcomes and secondary regulations, not just bill passage. The operative details of PL 2338/2023 live in the 199 amendments and in the implementing rules that follow. Set up a dedicated monitoring process for Brazilian regulatory publications, not just news coverage of the headline law.

Brazil's AI law applies to any organization offering AI-enabled products or services to people in Brazil. That is a large, fast-growing market. Starting the classification and documentation work now costs a fraction of what a retrofit costs once enforcement begins.

FAQ

What is PL 2338/2023?

PL 2338/2023 (Projeto de Lei 2338 de 2023) is a proposed Brazilian federal law establishing a comprehensive regulatory framework for artificial intelligence. It was introduced in the Senate by Senator Rodrigo Pacheco in May 2023. The law classifies AI systems by risk level, bans those that violate fundamental rights, and creates transparency and accountability requirements for high-risk applications.

How does Brazil's AI law compare to the EU AI Act?

Both laws use a risk-based classification with a prohibited tier and a high-risk tier requiring impact assessments and human oversight. The EU AI Act is more prescriptive, requiring conformity assessments and notified bodies for many high-risk systems. Brazil's framework places greater emphasis on fundamental rights protection and individual contestation of automated decisions. An organization compliant with the EU AI Act will cover most but not all of Brazil's requirements.

What AI systems does PL 2338/2023 prohibit outright?

The bill prohibits AI systems that perform subliminal manipulation causing harm, AI used by public entities for behavioral social scoring, and systems that exploit vulnerabilities of children, elderly persons, or disabled individuals. Real-time biometric identification in public spaces is also targeted, though the exact scope of that prohibition was one of the most contested points across the 199 amendments.

Why were 199 amendments proposed to a single AI bill?

The high amendment count reflects unresolved contested issues: the definition of 'artificial intelligence system,' which government body should serve as the primary regulator, how to treat generative AI and foundation model liability, biometric surveillance carve-outs for law enforcement, and whether to include regulatory sandboxes for startups. No major stakeholder group found the original text acceptable, which is common for first-generation AI legislation.

What should organizations operating in Brazil do right now?

Build an inventory of all AI systems deployed or offered in Brazil and classify each one against the bill's risk tiers. Prepare modular compliance documentation (impact assessments, transparency notices, human oversight procedures) that can be updated as final law and secondary regulations are published. Track amendment outcomes, not just headline news about the bill, because the operative compliance obligations live in the details.