Brazil's AI bill, PL 2338/2023, passed the Federal Senate in December 2024 but is not law: as of 27 October 2026 it was still waiting for the rapporteur's report in a special committee of the Chamber of Deputies. Many summaries describe it as if it were in force. It is not, and the Chamber can still change the text. What exists is a Senate-approved framework that sets the direction: a risk-based model close in spirit to the EU AI Act, a national governance system coordinated by the data protection authority (ANPD), fines of up to R$ 50 million or 2% of group revenue per violation, and staggered start dates once enacted. This guide explains what the Senate text covers, its exact legislative status, what it would mean for AI agents and the machine identities they use, and the preparation steps that make sense now. It is for security, privacy, legal and engineering leads at organisations that build, sell or use AI systems in Brazil. Data protection duties that already apply are covered in the Brazil LGPD Guide.
Key takeaways
- Not yet law: the Senate approved the substitute text on 10 December 2024 and sent it to the Chamber on 17 March 2025; the Chamber's Special Committee on Artificial Intelligence had recorded no report or vote by 27 October 2026.
- Risk tiers drive obligations: "excessive risk" uses are banned, high-risk uses carry testing, logging, explainability, human oversight and impact assessment duties, and generative and general-purpose AI need a preliminary risk assessment before release.
- ANPD would coordinate: the Senate text creates the National AI Regulation and Governance System (SIA), coordinated by the ANPD, with sectoral regulators setting rules in their own areas.
- Penalties are material: up to R$ 50 million or 2% of the group's gross revenue per violation, plus suspension of a system and bans on processing certain data.
- Staggered start once enacted: most provisions would apply 730 days after publication; prohibitions, generative and general-purpose AI rules and copyright rules after 180 days.
Scope: who and what is covered
The Senate text regulates the development and use of AI systems in Brazil and places duties on "AI agents", meaning the organisations that develop systems and those that deploy them. Its ementa is short: it "provides for the use of artificial intelligence". The approved substitute was presented by Senator Eduardo Gomes and replaced the original bill authored by Senator Rodrigo Pacheco.
| Situation | Treatment in the Senate text |
|---|---|
| AI system placed on the market or used in Brazil | In scope; obligations depend on risk tier and role (developer or deployer) |
| Personal, non-commercial use by individuals | Excluded |
| National defence | Excluded |
| Research, development and testing not yet released to market | Excluded until placed on the market |
| Systems that only provide data infrastructure for other AI systems | Excluded |
| Generative and general-purpose AI | Specific duties, including a mandatory preliminary risk assessment |
| System used only as an intermediate tool that does not influence relevant decisions | Not high risk; developers can ask for reclassification |
The scope overlaps with the LGPD wherever an AI system processes personal data. LGPD duties (legal basis, data subject rights, security under Article 46 and incident communication to the ANPD) apply today regardless of the bill's fate, so they are the baseline for any AI programme in Brazil.
Key obligations
Risk tiers
| Tier | Examples in the Senate text | Consequence |
|---|---|---|
| Excessive risk (prohibited) | Autonomous weapons; subliminal techniques; exploiting vulnerabilities to cause harm; generating child sexual abuse material; predicting crime from personality traits; social scoring that unfairly restricts access to goods, services or public policy | Development and use banned |
| Real-time remote biometric identification in public spaces | Allowed only in narrow cases: finding missing persons or victims, executing arrest warrants, certain criminal investigations, with judicial authorisation in some situations | Narrow exceptions only |
| High risk | Autonomous vehicles; traffic and utility management where physical safety is at stake; student selection; hiring, firing and worker management decisions; eligibility for public services; certain law enforcement and justice uses; emergency prioritisation; medical diagnosis; border control; biometric emotion recognition; crime-prevention data analysis | Full set of high-risk duties; list is open and can be extended by regulators |
| Generative and general-purpose AI | Foundation models and content generators | Preliminary risk classification before release; risk mitigation; synthetic content identifiers |
| Other systems | Everything else in scope | Governance and transparency basics; preliminary assessment optional but recommended |
Social media content-distribution algorithms were removed from the high-risk list during Senate debate.
Duties by tier
- All AI agents: internal governance structures that ensure system safety and protect the rights of affected people. People must be told when they are interacting with an AI system.
- High-risk systems: documented reliability and safety testing; mitigation of discriminatory bias; automatic logs of system operation; ability to explain results; reporting of serious security incidents to the authority within a timeframe set by regulation; human oversight, with alternatives where oversight is impossible or disproportionately costly.
- Algorithmic impact assessment: required for high-risk systems and carried out by qualified professionals. It covers risks to fundamental rights, benefits, likelihood and severity of harm, mitigation, transparency and system logic. Conclusions are public and held in a database run by the competent authority, and newly discovered risks must be reported to the authority and to affected people.
- Generative and general-purpose AI: preliminary risk classification before release; evidence that risks to fundamental rights, the environment, freedom of expression, information integrity and the democratic process were identified and mitigated; design that reduces energy and resource use; lawful data processing; an identifier on synthetic text, image, video and audio content (metadata is acceptable), with labelling details left to later regulation.
- Affected people's rights: privacy, non-discrimination and clear language for vulnerable groups; for high-risk systems, explanation, the right to contest a decision and human review.
- Copyright: a text and data mining exception for research institutions, journalism, museums, archives, libraries and educational bodies for lawfully obtained, non-commercial use; rights holders can object to other uses; commercially available systems trained on protected content owe remuneration; using a person's image or voice needs prior consent.
- Workers: the government and Ministry of Labour must issue guidelines to reduce risks to workers, such as job loss, and to support retraining.
Codes of conduct and certification are encouraged, and adherence counts as evidence of good faith when sanctions are set. For how these duties compare with Europe, see the EU AI Act High-Risk AI Systems Guide.
Timeline and current status
Status checked on the Senate and Chamber of Deputies tracking pages on 8 November 2026.
| Date | Event | Source |
|---|---|---|
| 10 December 2024 | Federal Senate plenary approves the substitute text (Opinions 207 and 208, supplementary round) | Senate |
| 17 March 2025 | Senate sends the text to the Chamber of Deputies for revision under Article 65 of the Constitution (Ofício SF nº 235) | Senate |
| 4 April 2025 | Chamber board creates a Special Committee because the bill was assigned to more than four committees | Chamber |
| 29 April 2025 | Special Committee constituted and the bill referred to it | Chamber |
| 20 May 2025 | Committee receives the bill; rapporteur Deputy Aguinaldo Ribeiro (PP-PB); public hearing requests begin | Chamber |
| 7 April 2026 | PL 6479/2025 attached to PL 2338/2023 | Chamber |
| 27 October 2026 (latest entry) | Situation: awaiting the rapporteur's opinion in the Special Committee on Artificial Intelligence; no opinions, substitutes or amendments recorded | Chamber |
What happens next: if the Chamber approves the bill with changes, it returns to the Senate for the changes to be considered; the text then goes to presidential sanction. Treat everything on this page as a description of the Senate text, which may change. Check the Chamber page before relying on any detail.
Start dates once published
| Provisions | Applies after publication |
|---|---|
| Organisation and powers of the regulatory bodies (except sanctions); incentives for sustainability and small businesses | Immediately |
| Prohibited (excessive risk) practices; generative and general-purpose AI rules; copyright rules | 180 days |
| Everything else, including high-risk duties | 730 days (two years) |
Enforcement and penalties
Under the Senate text, the ANPD coordinates the SIA and is the competent authority to impose sanctions, issue rules (including on certification and impact assessments), receive complaints and handle serious incident reports. It also regulates AI in economic sectors that have no specific regulator. Sectoral authorities can set AI rules within their areas, receive impact assessments and extend high-risk lists. Two advisory bodies sit inside the SIA: Cria, a regulatory cooperation council, and Cecia, a committee of AI experts and scientists. The SIA must recommend to Congress every four years whether the law needs updating.
| Sanction | Detail |
|---|---|
| Fine | Up to R$ 50 million or 2% of the group's or conglomerate's gross revenue, per violation |
| Warning | Administrative warning |
| Data processing ban | Prohibition on processing certain databases |
| Suspension | Partial or total suspension of the system, temporary or permanent |
| Civil liability | Under the Civil Code or the Consumer Protection Code, depending on the case; burden of proof can shift to the developer where proving causation is too burdensome for the victim |
No penalties can be imposed under the bill today. Penalties for AI processing of personal data can already arise under the LGPD.
What it means for AI agents and non-human identities
The bill does not mention service accounts, API keys or agent credentials. Its duties, however, cannot be met without controlling them. An agent that calls tools, reads data and takes actions does so through non-human identities, and the evidence a regulator would ask for is mostly identity evidence.
| Senate text duty | Identity and NHI control that supports it |
|---|---|
| Automatic logs of system operation (high risk) | A distinct identity per agent and per workload, so every log line attributes an action to one system, not a shared key. See the AI Agent Audit Trails for Regulators Guide. |
| Human oversight and the right to human review | Approval gates on high-impact actions, tied to the agent's authorisation policy. See the AI Agent Action Approvals and Spending Limits Guide. |
| Serious security incident reporting | Detection of misuse of agent credentials and a rehearsed revocation path, so an incident can be scoped and contained quickly. |
| Governance structures for system safety | An owner for every agent and machine credential, recorded in an inventory with purpose and expiry. See the NHI Discovery guide. |
| Impact assessment covering system logic | Documented scope of each agent's permissions and data access, reviewed when tools or models change. |
| Data processing bans as a sanction | Ability to revoke an agent's access to a specific data store without breaking unrelated services, which needs per-agent, least-privilege credentials. |
Agents built on general-purpose models would inherit two layers of duty: the model provider's preliminary risk assessment and, where the agent is used for a high-risk purpose such as hiring decisions, the deployer's high-risk duties. The tier follows the use, not the technology. For identity design patterns, see the Agentic AI Identity Guide.
Compliance steps
Because the text can still change, invest first in controls that pay off under the LGPD today and under any version of the bill.
- Inventory AI systems used in Brazil: list each system, its developer, its deployer role and the decisions it informs. Include agents and the credentials they use.
- Classify against the Senate tiers: flag any use that matches the prohibited list or the high-risk examples (hiring, credit-like eligibility, medical, safety-critical). Record the reasoning, including any intermediate-tool argument.
- Close LGPD gaps first: legal basis, data subject rights handling, Article 46 security measures and incident communication to the ANPD apply now.
- Build logging and attribution: per-agent identities, immutable operation logs and retention aligned with your audit needs. This is the hardest control to add late.
- Draft an algorithmic impact assessment template: cover fundamental-rights risks, likelihood and severity, mitigation, transparency and logic, and pilot it on one high-risk candidate system.
- Prepare transparency notices: tell users when they interact with AI, and design an explanation and contest path for high-risk decisions with a human reviewer.
- Plan synthetic content marking: if you ship generative features, assess metadata-based provenance now and track the later labelling regulation.
- Track the bill: assign an owner to check the Chamber page monthly, and re-run steps 2 and 5 when a committee report or a new text appears.
Small organisations should focus on steps 1 to 3 and on transparency; the Senate text gives immediate effect to incentive measures for small businesses, and detail on any lighter duties will depend on the final text. Large groups should note that the 2% fine is calculated on group or conglomerate revenue.
Common failure patterns
- Treating the bill as in force: building programmes and contracts around obligations that may change in the Chamber, while LGPD duties that already apply are neglected.
- Classifying by technology instead of use: a general chatbot used to screen job applicants is a high-risk use under the Senate list.
- Shared credentials for agents: one API key across several agents makes operation logs useless for attribution and makes a targeted processing ban impossible to honour.
- Impact assessments with no evidence: narrative documents that cannot show tests, bias metrics or the access an agent actually holds.
- Ignoring sectoral regulators: under the Senate text they can extend high-risk lists and receive assessments in their own sectors.
Practitioner checklist
- Confirm the bill's current status on the Chamber of Deputies page and record the date checked.
- Maintain an inventory of AI systems and agents used in Brazil, with owner, role and purpose.
- Classify each use against the prohibited and high-risk examples in the Senate text.
- Verify LGPD legal basis, security measures and incident communication for AI processing.
- Give each agent its own identity and least-privilege credentials, with an expiry and an owner.
- Keep operation logs that attribute every agent action to one identity.
- Pilot an algorithmic impact assessment on one high-risk candidate system.
- Publish AI interaction notices and a human review route for contested decisions.
Standards and references
- Senado Federal: PL 2338/2023 tramitação
- Câmara dos Deputados: PL 2338/2023 ficha de tramitação
- Agência Senado: Senado aprova regulamentação da inteligência artificial; texto vai à Câmara
- ANPD: Agência Nacional de Proteção de Dados
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act)
This guide summarises the Senate-approved text of Brazil's PL 2338/2023 and its status on 8 November 2026, and is not legal advice. Related NHI Mgmt Group resources: Brazil LGPD Guide · Global AI Regulation Map · EU AI Act Compliance Guide · Agentic AI Compliance Guide