Brazil - National AI Regulation (2.338/2023)

Overview

Projeto de Lei nº 2.338/2023 proposes national rules on the development, promotion and ethical and responsible use of artificial intelligence, centred on the human person. Article 1 combines fundamental-rights protection with responsible innovation, competitiveness and safe, trustworthy systems benefiting people, democracy and social, scientific, technological and economic development. The Senate text is dated 31 January 2025 and was presented to the Chamber of Deputies on 17 March 2025. It remains a bill under review, rather than an enacted national AI law. The framework combines individual and collective rights, risk categories, value-chain obligations, coordinated supervision, sanctions and sustainable innovation. Its proposed provisions appear in the official legislative text.

Article 1 excludes exclusively private, non-economic use by natural persons, AI developed and used solely for national defence, and pre-market investigation, research, testing and development. Consumer, data-protection, environmental and copyright laws remain applicable, and testing in real conditions must observe the bill. Services limited to infrastructure for storing and transporting data used in AI are also excluded. These exclusions are bounded by their stated purposes. Articles 2–3 establish human centrality, equality, privacy, accessibility, information integrity, worker protection and environmental sustainability as foundations and principles.

Definitions

Article 4 defines an AI system as a machine-based system that, with varying autonomy and explicit or implicit objectives, infers from received data or information how to generate outputs, particularly predictions, content, recommendations or decisions influencing virtual, physical or real environments. Its lifecycle covers conception, planning, development, training, retraining, testing, validation, deployment and monitoring; discontinuation may occur at any stage. General-purpose AI is based on a model trained on large-scale datasets and can perform varied tasks, including tasks for which it was not specifically developed or trained. Generative AI is specifically intended to generate or significantly modify text, images, audio, video or software code.

AI agents comprise developers, distributors and applicators. Developers create systems directly or by commission for market placement or service application under their own name or brand. Distributors make systems available for third-party application. Applicators employ systems in their own name or benefit, including configuration, maintenance and data support for operation or monitoring. These roles cover public or private natural and legal persons, including paid or free provision where specified. Article 18 assigns developer responsibilities to applicators or distributors making substantial modifications or changing a system's purpose. Article 4 also defines affected persons and groups, vulnerability, preliminary assessment, algorithmic impact assessment, synthetic content, systemic risk and experimental regulatory environments.

Governance and Institutional Framework

Article 45 names the Autoridade Nacional de Proteção de Dados (ANPD) as the competent authority coordinating the Sistema Nacional de Regulação e Governança de Inteligência Artificial (SIA). The SIA also comprises sectoral authorities, the Conselho Permanente de Cooperação Regulatória de Inteligência Artificial (Cria) and the Comitê de Especialistas e Cientistas de Inteligência Artificial (Cecia). The Executive is authorised to establish the system and determine participating bodies and institutional arrangements. Coordination preserves sectoral regulatory, supervisory and sanctioning powers, without hierarchical subordination between coordinating and sectoral authorities.

Articles 46–49 give ANPD general rulemaking, coordination, audit and supervisory functions. It would issue binding general rules collaboratively, support sector-specific regulation, receive anonymous complaints with identity protections and publish annual activity reports. Article 47 gives ANPD residual regulatory and enforcement jurisdiction over economic activities without a specific sectoral regulator. Sectoral authorities retain statutory competence and issue market-specific technical rules. Cria develops guidelines and supports cooperation, studies and public debate; Cecia provides technical and scientific orientation and supervision under regulation. Article 74 requires necessary ANPD resources, including administrative restructuring, within two years and assigns SIA membership and institutional composition decisions to the Executive.

Rights and Risk Classification

Articles 5–11 distinguish general rights from additional high-risk protections. General rights include accessible, free and understandable information about AI interactions, privacy and personal-data protection, and protection against unlawful or abusive discrimination. Interaction information has a regulated exception for systems exclusively dedicated to cybersecurity and cyberdefence. Persons affected by high-risk AI additionally receive explanation, contestation, requested review and human-review rights, considering context, risk and technological development. Under Article 8, demonstrably impossible or disproportionately burdensome human oversight must be replaced with effective alternative measures. Systems intended for vulnerable groups require transparency and language appropriate to age and cognitive capacity. Article 11 provides administrative and individual or collective judicial protection.

Article 13 prohibits specified harmful manipulation, exploitation of vulnerabilities, personality-based assessments of criminal offending risk, facilitation of child sexual abuse material, illegitimate or disproportionate public universal social scoring and autonomous weapons systems. Real-time remote biometric identification in publicly accessible spaces is prohibited subject to defined exceptions and safeguards. Article 14 identifies high-risk contexts: critical infrastructure, education, employment, essential services, justice, autonomous vehicles, health, criminal analysis, administrative investigations, specified biometric uses and migration control. Narrow procedural tasks and intermediate technologies that do not influence or determine outcomes are excluded from high-risk classification on that basis. Article 19 requires synthetic-content identifiers according to technological development, use context and regulation.

Implementation and Assessment Requirements

Article 12 generally permits preliminary risk assessment before market introduction or use. This good practice may support favourable sanctioning consideration or priority in conformity procedures. Sectoral authorities may request an assessment or access to it, and authorities may reclassify systems with procedural safeguards. Article 29 separately requires developers of general-purpose and generative AI to conduct preliminary assessment, including identification of potential systemic risk. Articles 25–28 require algorithmic impact assessment for high-risk systems or uses according to the developer's or applicator's role. Assessment precedes market introduction, continues throughout the lifecycle and is periodically updated, including when significant changes occur.

Articles 17–18 establish role-specific safeguards, including documentation, testing, operational records, interpretability support, bias mitigation and value-chain cooperation. Distributors must support and verify compliance before circulation. These mandatory duties are separate from optional internal governance programmes and voluntary good-practice codes under Article 40. Article 34 permits accreditation of national or international conformity-assessment bodies. Sectoral regulation may flexibilise specified obligations according to context and role. Public-sector high-risk systems require additional access protocols, interoperability and rights safeguards. Sandboxes allow time-limited development and testing under authorised conditions; participants remain responsible for third-party damage under applicable law.

For systemic-risk general-purpose and generative AI, Article 30 sets pre-commercial-availability requirements covering documentation, testing, lawful data, training summaries, efficiency and downstream information. Synthetic-content obligations include regulated identifiers and qualified artistic signalling options. Public-sector applications also require publication and citizen-review safeguards, and discontinuation where relevant risks cannot be eliminated or substantively mitigated.

Monitoring and Evaluation

Article 42 requires AI agents to report serious security incidents to the sectoral authority within its prescribed period and severity criteria. Reportable events include risks to life or physical integrity, critical-infrastructure disruption, serious property or environmental damage, and serious violations of fundamental rights, information integrity, freedom of expression or democratic processes. The authority may order measures to reverse or mitigate effects. Article 25 separately requires immediate communication to the sectoral authority and other value-chain agents when unexpected, relevant risks or impacts on natural persons' rights become known after market introduction or service use.

Article 44 establishes a public high-risk AI database containing public impact-assessment documents, subject to commercial-confidentiality, data-protection and access-to-information safeguards. Sectoral databases may operate if interoperable and structured. ANPD may conduct or order necessary audits of high-risk systems or systems producing relevant legal effects, request information from public bodies and accredit institutions for audit or research access under security and confidentiality requirements. Joint investigations with sectoral authorities are contemplated. Accredited conformity bodies receive continuous monitoring and periodic reassessment. Annual activity reports support accountability; Article 79 requires SIA studies and an advisory opinion to Congress every four years on framework improvements.

Penalties, Liability and Redress

Article 50 provides warnings, fines, publication of confirmed infringements, sandbox restrictions for up to five years, partial or total suspension of development, supply or operation, and prohibitions on processing particular datasets. Fines are capped at R$50 million per infringement. For private legal persons, fines may reach 2% of the preceding financial year's Brazilian gross revenue of the entity, group or conglomerate, excluding taxes, subject to that cap. Administrative proceedings must allow full defence and consider severity, good faith, cooperation, remediation, governance and proportionality. Preventive measures are available under statutory conditions. Public bodies may receive the listed non-fine sanctions. Detailed sanction procedures require regulation, public consultation and regulatory impact analysis. Article 50 follows general rather than immediate commencement.

Articles 35–39 preserve existing liability regimes. Consumer relationships remain subject to the Consumer Protection Code and relevant legislation; other damage caused by AI agents remains subject to the Civil Code and special laws. Applicable autonomy, risk and agent-specific regimes inform concrete liability analysis unless law provides otherwise. Article 37 directs judges to reverse the burden of proof where victims are disadvantaged or system operation makes proving liability requirements excessively burdensome. Reversal depends on these statutory conditions rather than AI involvement or high-risk classification alone. Administrative and judicial rights channels remain available.

Relationship to Other Instruments

The bill complements existing Brazilian legislation, preserving the Lei Geral de Proteção de Dados Pessoais, consumer law, the Civil Code, copyright law, environmental legislation and sectoral rules. Article 27 allows algorithmic impact assessment to be prepared jointly with a required personal-data impact report. Article 43 preserves specific cybersecurity, critical-infrastructure and rights-protection laws. Competition safeguards prohibit competitively sensitive information sharing through self-regulation, require SIA authorities to communicate relevant findings to Cade, and permit necessary access to high-risk training, validation and testing datasets when investigating concrete indications of competition infringements. Article 72 preserves other domestic rights and treaty-based rights.

Articles 62–66 address copyright disclosure, narrowly conditioned non-commercial research text-and-data mining, rights-holder prohibitions outside that exception, remuneration subject to domicile and reciprocity conditions, and personality rights in identifiable image or voice content. The research exception requires lawful access, necessary use and protection of economic interests; storage and dissemination are restricted, and specified commercial affiliations are excluded. Article 75 adds critical algorithmic literacy to the National Digital Education Policy. Article 76 prioritises AI research and projects in specified funding provisions. Article 77 reserves online content-circulation regulation affecting freedom of expression, including AI moderation and recommendation, to specific legislation. Article 78 preserves perfected legal acts, acquired rights and res judicata.

International Alignment

Articles 2 and 68 promote international cooperation, knowledge sharing and regulatory and technological interoperability. ANPD would represent Brazil before international AI bodies under Executive coordination and cooperate with foreign, international or transnational authorities. Articles 46 and 48 encourage internationally recognised standards, practices, certification and accreditation. Article 34 permits cooperation and mutual-recognition arrangements with international accreditation bodies to facilitate recognition of foreign conformity assessments. Recognition operates through these arrangements and applicable Brazilian requirements.

Article 19 envisages a software library supporting synthetic-content signalling, ideally using a widely recognised international standard. Article 41 permits voluntary self-regulatory associations to adopt internationally recognised practices and certification models while observing Brazilian competition law and binding SIA rules. Sectoral governance supports technical and jurisdictional interoperability alongside national legislation. The bill also promotes domestic technological autonomy, Portuguese-language and cultural relevance, Brazilian research and sustainable infrastructure, combining international cooperation with domestic safeguards and industrial development. Treaty-based rights remain protected under Article 72.

Implementation Timeline

Sources and References