Brazil
Tools
Tasks
Users
Scope
Training
Regulation
Cases
Insights
Information uploaded as at March 2026
AT A GLANCE
Brazil is a global leader in AI adoption in criminal justice, with tools spanning law enforcement, prosecutors, courts (for criminal and civil matters), and defence. Police use platforms like Detecta and satellite-based monitoring, while cities expand facial recognition and integrated surveillance (such as Smart Sampa). Prosecutors rely on systems such as LuminarIA, Jarvis, and video-analysis AI to manage cases and evidence efficiently. Courts deploy over 170 AI tools for case management, legal research, and predictive analysis, including APOIA, ASSIS, and VICTOR. Public Defenders are considering piloting AI for drafting and case review. Judges receive mandatory training on AI use, risks, and biases.
Brazil is also emerging as a global leader in AI regulation. It is currently in the process of establishing a comprehensive AI regulatory framework inspired by the EU’s AI Act and is one of a few countries with detailed judicial guidelines on the use of AI in the judiciary. The judicial guidelines (Resolution No. 615/2025) establish a risk matrix for AI systems in use and introduce specific provisions on the use of generative AI. They emphasise the need for human oversight, transparency in the use of AI, explainability (i.e., enabling users to understand how and why a specific outcome was produced) as well as respect for fundamental rights, including the ‘right to a full defence’ and due process. Federal judges interviewed noted that while there is an interest in adopting AI broadly within the judiciary, it is a deliberate choice not to extend its use into the criminal sphere, except in areas peripheral to adjudication and, above all, only where it can be ensured that defendants are in no way placed at a disadvantage.
USE
Since September 2020, Brazil’s National Council of Justice (CNJ) – the administrative body overseeing Brazil’s judiciary – in collaboration with the UN Development Programme, has been led by the ‘Justice 4.0 Program’, a technological modernisation initiative, focusing on driving digital transformation across the Brazilian judiciary. Today, Brazil is a global leader in AI adoption. In courts, AI tools span civil and commercial matters, though since the National Council of Justice’s Resolution No. 332/2020, there has been a clear discouragement of the development and use of AI tools in criminal adjudication.
Law enforcement
In August 2020, the CNJ issued guidelines on the development and use of AI in the judiciary by adopting Resolution No. 332/2020. At that time, only predictive AI systems were in use – though none were applied in criminal matters. Indeed, influenced by the US case of Loomis v Wisconsin [2016] 881 N.W.2d 749 (Wisconsin Supreme Court), the resolution expressly discouraged the use of predictive AI models in criminal matters albeit with some narrow exceptions. It was designed for computational solutions ‘aimed at supporting procedural management and enhancing the effectiveness of judicial services’. In March 2025, the CNJ updated these guidelines in response to the emergence of generative AI and their growing use within the judiciary by adopting Resolution No. 615/2025 (‘Resolution’), revoking the earlier Resolution No. 332/2020. The preamble of the new Resolution acknowledges AI’s potential role in supporting decision-making and states that specific regulations on the use of generative AI are ‘indispensable’. To operationalise implementation and oversight, the CNJ subsequently constituted the Comitê Nacional de Inteligência Artificial do Judiciário (National Judiciary Committee for Artificial Intelligence) through Resolution No. 270/2025, which states that the purpose of the Committee is to assist with CNJ with the “implementation, compliance and supervision” of the Resolution and designates the Committee’s members.
Federal judges interviewed, however, noted that while there is an interest in adopting AI broadly within the judiciary, it is a deliberate choice not to extend its use into the criminal sphere, except in areas peripheral to adjudication and, above all, only where it can be ensured that defendants are in no way placed at a disadvantage.
Operational support
São Paulo has implemented a smart surveillance and policing platform called ‘Detecta’, with the objective to connect police data, automate threat detection, and enable preemptive responses. The platform integrates multiple data sources, such as: civil and military police databases, digital incident reports, criminal photo registries, vehicle and driver data, and real-time CCTV feeds. Detecta was promoted as a multi-platform tool capable of automatically detecting potentially suspicious behaviours, for example, identifying a motorcycle parked in the middle of traffic as potentially suspicious.
The Federal Police of Brazil, the Brazilian Ministry of Justice and Public Security, and Planet Data and SSCON Geospatial have collaborated to leverage satellite data and develop a change detection alert system, making near real-time information regarding illicit activities, such as environmental crimes and illegal mining, accessible to Brazilian government agencies.
Data review and analysis
As at March 2026, 506 AI-driven facial recognition systems have been adopted in Brazil, with a projected reach to around 87 million persons (41% of the Brazilian population). These systems have reportedly been deployed across at least 30 Brazilian cities (as of 2019)for public safety and fraud prevention. Issues of discrimination based on race and gender identity have arisen. In 2019, for example, a black woman was arrested in Rio de Janeiro after her face was mistakenly identified by a smart camera installed in the Copacabana region during a pilot project. She was mistaken for a suspect who had already been serving a sentence since 2019.
Nonetheless, the use of technology for public safety remains in the plans of several public managers, mayors, and governors. For example, ‘Smart Sampa’, a project by the city of São Paulo, provides a single video surveillance platform that integrates and supports the operations of emergency and traffic services, the city’s public transport network, and police forces. As at March 2026, the programme has around 40,000 cameras across the city, combining approximately 20,000 municipal cameras with up to 20,000 third-party and privately owned cameras voluntarily connected through a public integration scheme. The system relies on AI-enabled real-time analytics, including facial recognition, automatic licence-plate recognition, and behavioural detection algorithms (e.g. for theft, vandalism, and intrusions) to generate alerts and assist authorities in locating wanted fugitives, missing persons, stolen vehicles, and lost objects.
Operations of the surveillance programme are coordinated by the Guarda Civil Metropolitana (São Paulo’s municipal law enforcement agency) through the ‘Central de Monitoramento Smart Sampa’, inaugurated in July 2024, which the municipal government describes on its website as “Latin America’s largest integrated monitoring centre”. The centre does not function as a full police dispatch centre, such as the automated 911 centres in the US. The operations team oversees the system and, when necessary, notifies law enforcement field units, which then carry out the enforcement action. The centre offers round-the-clock access for both the public and the press to observe the monitoring environment and understand how the programme operates.
According to the city of São Paulo, Smart Sampa operates under data protection and governance rules aligned with Brazil’s General Data Protection Law. These include limits on data retention and defined access controls, and an explicit commitment that AI supports, rather than replace, human decision-making in public safety operations.
Prosecutors
Brazilian prosecutors have signalled their broad support for the deployment of AI. In September 2025, the heads of Brazil’s prosecutorial office approved a declaration promoting the use of new technologies and AI in law enforcement. The document underscores the role of these tools in bringing speed and efficiency to investigations and judicial proceedings. Brazilian prosecutors were joined by the Attorney Generals of the BRICS Countries, including China, Egypt, India, Iran, Russia, South Africa, the UAE and Indonesia.
Case management
The Public Prosecutor’s Office of the Federal District and Territories uses a tool called ‘LuminarIA’, developed to automate the processing of low-complexity cases. The system analyses processes, verifies requirements, and suggests appropriate measures, optimising prosecutors’ time.
Another notable initiative is ‘Jarvis’, a hearing transcription and analysis system. It allows prosecutors to access structured summaries of testimonies, enabling them to compare versions and identify inconsistencies.
In October 2025, the Public Prosecutor’s Office of Paraíba launched ‘apoIA.MP’, a tool that uses AI to analyse a prosecutor’s case material imported from the case management portal, MPVirtual, and proposes next steps for indictments and texts, which are editable.
Charging support
'Etiquetas Inteligentes' ('Smart Tags') s used by the Ministério Público de São Paulo (Public Ministry of São Paulo, MPSP). It is an AI-driven feature that automatically identifies the procedural stage of case files, such as penalty-calculation adjustments, sentence progression, and remission. It assists by suggesting the proper type of petition to be used, reducing manual triage. Since April 2023, it has been applied in over 3,000 cases across four MPSP units: Bauru, Ribeirão Preto, Presidente Prudente, and São José do Rio Preto.
Evidence review and analysis
The Ministério Público do Estado do Rio Grande do Sul (Public Ministry of the State of Rio Grande do Sul), in partnership with Xertica.ai, has deployed a generative-AI solution that reduced video analysis time by up to 90%. This tool enables automatic transcription with diarisation, generation of summaries, detection of contradictions, and sentiment and bias analysis. The tool has processed over 23,400 videos from November 2024 to May 2025, saving over 11,500 hours of work.
In 2026, the Ministry of Justice and Public Security issued Ministerial Ordinance No. 1122/2026, which regulates the formal identification procedures used in criminal investigations or prosecutions (see below).
Courts
AI has already been adopted by a majority of the e Brazilian courts, including the Brazilian Supreme Federal Court (Brazil’s constitutional court) and the Superior Court of Justice (Brazil’s highest court for non-constitutional matters and matters not reserved for specialised courts). As at March 2026, the most recent research conducted by the CNJ in its Artificial Intelligence Research in the Judiciary 2024 report, Brazilian courts have adopted 178 AI initiatives, the majority of which are predictive or analytical systems, many of which may be used in civil proceedings. These systems have been deployed through both top-down institutional programmes and bottom-up innovation by individual courts. In parallel, generative AI tools have begun to emerge, with nearly half of Brazilian courts reporting some use of generative AI in 2024, primarily to support drafting, document analysis and legal research. These predictive systems are used for, for example: case classification, similarity and clustering, mass-litigation detection, and forecasting workloads.
Case management
As part of the Justice 4.0 Program, ‘Plataforma Codex’ was developed by the Tribunal de Justiça de Rondônia in partnership with the CNJ to serve as a ‘data lake’ for procedural data, consolidating content from judicial case files into a centralised and standardised repository. In March 2022, the CNJ launched ‘Plataforma Codex’ as the official, mandatory tool for extracting judicial data across courts in Brazil through Resolution No. 446/2022. As at March 2026, the database contains more than 386 million lawsuits, and decisions can be uploaded in less than two hours. This tool can be incorporated into AI systems.
Examples of simple automation systems (without an ‘intelligent’ model) used for case management in Brazilian courts are:
|
MANADUS and SCRIBA |
Developed by the Roraima Court of Justice, MANADUS has the goal to assist in case distribution to bailiffs according to zoning and location criteria. The tool seeks to guarantee enforcement of warrants and provide data updates and real-time court summons: the bailiff can immediately register in the court’s system that a party has been officially served with court papers, and the tool can be used as an app on the bailiff’s mobile device. SCRIBA conducts the automatic transcription of hearings and sessions, but still cannot discern from different voices and it is up to a civil servant to manually identify each speech to its corresponding interlocutor. Both projects do not yet use AI in their working structure, but they must incorporate machine learning techniques for risk classification of compliance with the warrant and the allocation of bailiffs according to their ability to comply. |
|
POTI |
A project conducted by the Rio Grande do Norte Court of Justice in partnership with the Federal University of Rio Grande do Norte, delivering products to automate bank account blocking procedures. The system automatically searches for specific amounts in bank accounts, and has the function of updating the value of tax enforcement action and transferring the blocked amount to official accounts. |
|
RADAR |
Developed by the Minas Gerais Court of Justice to deal with the identification of repetitive demands. |
|
SAAJUS |
Developed and implemented by the Federal Justice of Rio Grande do Norte in partnership with the Federal University of Rio Grande do Norte to streamline the processing of legal proceedings. The system reads the petition for tax foreclosures and active debt certificates, captures all the data, prepares the initial order and moves the process for signature. |
Other systems have been adopted by Brazilian courts that implement intelligent models. For example, the Superior Court of Justice uses the ‘ATHOS system’, an AI-based tool created in 2019 with the main role of identifying, before case assignment, appeals that may fall under the ‘repetitive resources’ procedure, a mechanism used to resolve numerous cases involving the same legal issue efficiently. It clusters cases based on semantic similarity and flags those with convergent or divergent judicial positions.
In the Superior Court of Justice, ‘Projeto Sócrates’, through AI, seeks to reduce by 25% the time it takes to issue appellate judgments. The system analyses the appeals received by the Court from 300,000 resolved cases and groups cases that are similar, to decide them together. According to the Minister of Justice, Ricardo Villas, the aim is to implement this system to produce automated draft decisions based on previous judicial decisions, whilst retaining human review before a final decision is taken.
Legal research, analysis and drafting support
Judges in Brazil issue, on average, nine final judgments per working day: the overall number of rulings rendered daily is estimated to be around 100 to 150. Therefore, courts in Brazil have explored the potential use of AI systems for legal research, analysis and drafting support, and decision-making support (see below).
The ‘APOIA system’ (Assistente Pessoal Operada por Inteligência Artificial) is a generative AI assistant, integrating multiple AI tools (including ChatGPT and Gemini) implemented into the national Digital Platform of the Brazilian Judiciary (PDPJ-Br). It supports tasks such as drafting reports, summarising case files, and identifying applicable law. This tool was developed initially by the Federal Regional Court of the Second Region, and then incorporated into the PDPJ-Br and made available to Brazilian courts. APOIA is a secure, institutional alternative to ad-hoc private tools, emphasising responsible, ethically governed AI use and data protection. APOIA also includes a collaborative ‘prompt bank’ for reusing effective instructions across courts.
Brazilian courts have also adopted their own systems to assist with legal research and drafting support:
|
ASSIS |
At the Tribunal de Justiça do Rio de Janeiro, the Assistente de Inteligência Artificial Generativa (ASSIS) system generates drafts of judicial decisions, sentences, and reports using GPT-4 based generative models. The system tailors’ output to each judge’s writing style and judicial record, drawing from their prior decisions and reports, and also enables judges to ask questions about case documents and access relevant information from electronic case files directly. The system operates securely, with data governance and audit trials. It does not reuse data for AI training. |
|
ATHOS |
As well as assisting with case assignment, the ‘ATHOS system’ in the Superior Court of Justice, discussed above, highlights key matters like the overruling of precedents or cases of notable relevance. |
|
Legal Intelligent Advisor (LEIA) and HORUS |
The ‘LEIA’ is a system used across various Courts of Justice and developed by Softplan to read case files in PDF format, identify cases that potentially match with prior legal precedents, and connect them with those legal precedents. Across the Federal District, the HORUS system also performs similar functions. |
|
Projeto Sócrates |
Used by the Superior Court of Justice, ‘Projeto Sócrates’ (mentioned above) performs semantic analysis of the procedural documents in a case, researching court judgments that can serve as a precedent for the process under examination. |
|
VICTOR |
At the Supreme Federal Court, the ‘VICTOR system’ is used by court officials and aims to analyse compliance with the constitutional requirements of admissibility, and accelerate analysis of cases that reach the Supreme Court by using document analysis and natural-language processing tools. |
|
SIGMA |
SIGMA is an AI-powered system designed to assist judges and clerks in drafting reports, decisions, and judgments within Brazil’s TRF3 Electronic Judicial Process system. It analyses pre-existing decision models for similar cases. |
Decision-making support
Brazilian courts are currently exploring the use of machine learning for sentences using historical data, potentially in criminal cases. Examples include a second phase of the SINAPSE platform (in the Rondônia Court of Justice/TJ-RO and sponsored by the CNJ for the development and large-scale availability of AI prototypes) and a system called ‘Jerimum/Clara’ (in the Rio Grande do Norte Court of Justice). These systems are not currently in use.
Alongside these exploratory initiatives, courts have begun increasingly adopting generative AI tools to support judicial work, without automating substantive decision-making. In February 2025, the Superior Court of Justice launched STJ Logos, an internally developed generative-AI engine integrated into the court’s case management system. STJ Logos is designed to increase efficiency by assisting with tasks such as drafting reports and analysing the admissibility of appeals. The Court has emphasised that the tool is strictly supportive and that full responsibility for decisions and reasoning remains with the judge.
In 2024, a Brazilian judge signed a draft decision prepared by a court clerk who had used ChatGPT to generate a sentence, without informing him. The judge relied on the caselaw presented by the clerk, which resulted in false facts and fabricated jurisprudence being included in the final judgment. As these false premises formed the basis of the judge’s decision, the defendant was wrongly convicted. After a complaint was filed with the Office of the Chief Inspector of the Federal Judiciary of the First Region, the CNJ deemed it necessary to investigate the case, aiming to rectify the situation and prevent similar occurrences in the future. Before the CNJ decision to investigate, a regional inspectorate had decided to archive the investigation, as it did not detect any ‘disciplinary infraction’ on the part of the judge or his assistant. The result of the final administrative review is confidential.
In 2025, the Office of the General Inspector of Justice opened an investigation into a judge over suspected improper use of AI in judicial decision-making after his monthly output reportedly jumped from approximately 80 decisions to 969 in August 2024. The inspectorate identified serious issues including decisions lacking reasoning, failure to analyse evidence, and citation of non-existent precedents, described as creating legal uncertainty.
In April 2026, the STJ addressed whether a criminal complaint may validly rely on a technical report produced using generative AI, in a case where the São Paulo Public Prosecutor’s Office relied on material generated with tools such as Gemini and Perplexity to support the accusation. The case raised the question whether such material could constitute a sufficiently reliable evidential basis for criminal proceedings. The defence argued that the report lacked the guarantees associated with formal expert evidence, including a transparent methodology, reproducibility, and proper technical validation, while the lower court had treated it as an investigative aid rather than a formal forensic report. It was understood that the problem was not in the legality of obtaining the report or in the alleged violation of the chain of custody of the evidence, but whether the AI tools were reliable to support criminal accusations.
Defence
Legal research, analysis and drafting support
The Brazilian government is developing strategies to integrate AI into Public Defender offices to improve access to justice and efficiency. AI tools, especially those based on large language models, are being explored for tasks such as streamlining case analysis, summarising judicial documents, and drafting procedural documents.
Victims
Under Brazil’s criminal procedure rules, victims are not formal parties to the prosecution in the same way as the Public Prosecutor’s Office, which has primary responsibility for conducting criminal proceedings. However, victims may participate as civil parties (assistentes de acusaçao), allowing them to support the prosecution, present evidence, request investigative measures, and appeal certain decisions. Victims may also pursue civil compensation claims arising from the criminal conduct, either within the criminal proceedings or through separate civil litigation.
In terms of procedural involvement, victims may participate in evidentiary procedures—including identification procedures and testimony—and are entitled to certain procedural protections. These include measures to safeguard their privacy and dignity, protection against revictimisation, restrictions on offensive questioning or evidence during hearings (e.g., under Article 400-A of the Código de Processo Penal), and the right to receive information about the progress of the case.
As at March 2026, there are no reported uses of AI tools by victims in criminal proceedings.
TRAINING
Training is available for specialised AI tools used by prosecutors. For example, when apoIA.MP was launched in October 2025, it was announced that virtual training on the use of the tool was to be made available to prosecutorial staff in Paraíba.
Judges in Brazil are receiving training on the use of AI. Courts are required to offer continuous education for judges and court staff on the risks of automation, algorithmic bias, and critical analysis of AI-generated outcomes. There are mandatory courses in AI use for judges, with specific practical training on tools such as ChatGPT and Claude. For the courts that have already adopted their own institutional system (such as ASSIS), there is also training in those systems.
Judicial AI training has evolved from introductory courses on the nature of AI and its risks to more hands-on programs that demonstrate how judges can employ generative tools—such as ChatGPT, Claude, or court-developed systems—to draft, summarise, review case materials, and support legal analysis.
For example, in January 2025, the Federal Regional Court of the 1st Region (TRF-1) reported that its initial training for 50 new federal judges included, for the first time, specialised instruction on AI and generative AI tools such as ChatGPT. TRF-1 also later offered a course for court personnel on prompt engineering for legal use of generative AI, and another course for appellate judges on the ethical, responsible and secure use of generative AI in the judiciary. In December 2025, the Court of Justice of the State of Rio de Janeiro (TJRJ) funded a 25-hour AI course in Milan for 23 appellate judges, covering comparative AI regulation, data protection, and ethical/social challenges of automation in the judiciary.
The CNJ has also started to publish specific AI courses. Those include an “Introduction to AI for the Judiciary” course and, in September 2025, a self-paced course on Apoia, a generative-AI assistant for judicial staff. In October 2025, the CNJ also launched a course on grouping legal texts and AI, aimed especially at technical judiciary staff working on digital-platform solutions.
There has also been AI training for prosecutors. The National Council of the Public Prosecutor’s Office, through MP Digital, publicized a 2024 course called “Demystifying Artificial Intelligence: Fundamentals and Practical Applications” for members and staff of the Public Prosecutor’s Office, and in 2025 it announced further AI-focused training as part of the MP Digital learning programme.
After establishing guidelines for deepfakes and AI in campaign advertisements for 2024, the Superior Electoral Court (TSE) intended to train electoral judges and staff on handling AI-related challenges during elections, including regarding deepfakes. However, there is no evidence that this training has actually taken place.
REGULATION
As at March 2026, there is no national law governing the use of AI in criminal proceedings, but the CNJ has issued detailed guidelines. The Brazilian Bar Association has likewise issued guidance for practitioners. In addition, Brazil is seeking to establish a comprehensive AI regulatory framework inspired by the EU’s AI Act.
Guidelines for practitioners
Judicial guidelines
The CNJ has issued and updated its detailed ‘guidelines for the development, use and governance of artificial intelligence solutions within the judiciary’ (Resolution 615/2025). The guidelines require courts to categorise AI solutions into low-risk, high‑risk, and excessive-risk. Rather than providing definitions of each category, the annex of the Resolution includes a list of ‘purposes and contexts’ to exemplify what falls under each category. Courts are required to evaluate the risk level of AI solutions based on this categorisation and factors such as ‘the potential impact on fundamental rights, model complexity, financial sustainability, intended and potential uses, and the amount of sensitive data used’. Excessive-risk AI solutions are prohibited while high-risk AI solutions are subject to specific requirements and safeguards, including continuous monitoring and algorithmic impact assessment prior to their deployment. The Resolution specifically addresses the use of Generative AI solutions which are subject to additional requirements.
Low-risk solutions: AI solutions that support judicial administration and case management, or assist with legal research, analysis, and drafting, are considered low risk, provided they are overseen by a human and do not replace human judgment and evaluation. This includes, for example, ‘detecting decision-making patterns’ to ensure consistent case law, ‘producing supporting texts to facilitate the drafting of judicial acts’, ‘transcribing audio and video to assist judges’, and ‘anonymising documents’.
High-risk solutions: The evaluation of evidence, especially when this ‘can directly influence judicial decisions’; ‘identification of profiles and behavioural patterns’; ‘investigation, evaluation, classification, and legal characterisation of facts as crimes’; ‘formulation of conclusive judgments’ based on the application of legal norms to specific facts; and the ‘performance of facial or biometric identification and authentication to monitor behaviour’ are generally considered high risk. High-risk solutions must undergo regular auditing and continuous monitoring ‘to mitigate potential risks to fundamental rights, privacy, and justice’. Before deploying high-risk models, courts must carry out an algorithmic-impact assessment with public participation ‘whenever possible’ and make the findings public. They also need to implement additional governance measures, including ‘to mitigate and prevent discriminatory biases’. Courts must ‘enable explainability’ of AI-generated outcomes ‘whenever technically feasible’, ‘while respecting copyright, intellectual property and industrial and commercial confidentiality’, and use training data that is adequate and representative.
Excessive risk solutions: The Resolution prohibits developing and using AI solutions that pose ‘excessive risks to information security, the fundamental rights of citizens, or the independence of judges’, including solutions that:
- do not allow human review of the data used and the results proposed or that create an absolute reliance on the proposed outcome by the user, without the possibility of modification or review;
- ‘assign value to personality traits, characteristics, or behaviours of individuals or groups’ to ‘evaluate or predict the commission of crimes or the likelihood of recidivism in the reasoning of judicial decisions’;
- ‘classify or rank individuals based on their behaviour, social status, or personal traits for the purpose of assessing the plausibility of their rights, legal merits, or testimonies’; and
- ‘identify or authenticate biometric patterns for emotion recognition’.
Generative AI systems: Generative-AI systems一given special focus in a dedicated chapter一may be used by judges and judiciary staff to support case management or assist decision-making. While such AI solutions should ‘preferably’ be provided and monitored by the courts, judges may also use commercial solutions they acquired through private subscriptions provided that they have undergone specific training; that the tool is only ‘supportive’ and not used for purposes classified as high risk or excessively risky; and that the company providing the generative AI system complies with data-protection and intellectual-property standards and does not use the data entered to train the AI system. Resolution No. 615/2025 expressly leaves itto the judge’s discretion whether to disclose that generative AI was used to assist drafting of judicial decisions; however, the court’s internal system must automatically register such use for statistical, monitoring, and audit purposes. Judges and court staff who use commercial generative AI solutions must periodically report their use to the local Judicial Oversight Office, which submits consolidated information to the National Judiciary Committee for Artificial Intelligence. The Committee, established by the Resolution to oversee and implement the guidelines, is tasked with drafting and updating a best-practice manual on the proper, ethical, and efficient use of generative AI.
Key principles and rules: The Resolution sets out principles for the development, deployment and use of AI solutions by the judiciary, without specifying the mechanisms for their implementation:
|
Respect for fundamental rights |
Courts must ensure compatibility with fundamental rights through compatibility assessments and monitoring mechanisms. If there are ‘reports or indications of violations of fundamental rights’, the Brazilian Bar Association, the Public Prosecutor’s Office, and ‘other legitimate entities’ must be granted access to the algorithmic-impact assessment. |
|
Due process and right to a full defence |
Courts must be guided by ‘due process, the right to a full defence, the principle of adversarial proceedings, the physical presence of the judge, and the reasonable duration of proceedings, ensuring full respect for the prerogatives and rights of stakeholders in the justice system’. |
|
Human oversight and risk-based supervision |
Human participation and oversight is required at ‘all stages of the development and implementation cycles’ with narrow exceptions. The level of human oversight may also depend on ‘the degree of risk involved’, and ‘the level of automation and impact’. Under no circumstances may the AI system restrict or replace the ‘final authority’ of the judge. |
|
Transparency, explainability, traceability and auditability |
Courts must ensure transparency regarding the use and governance of AI systems and publish reports on ‘the system’s functionality, purposes, the data processed, and supervision mechanisms’. The CNJ, which validates and audits AI solutions, must be notified about them through the ‘Sinapses platform’. The individual use of AI must be automatically recorded in the court’s internal system ‘for statistical, monitoring, and auditing purposes’. Judges are, however, under no obligation to disclose the individual use of AI in judicial decisions. AI models must ‘include explainability mechanisms, whenever technically feasible, ensuring that their decisions and operations are understandable and auditable by judicial operators’. The data used in the development of AI systems must be ‘representative’, ‘secure’, ‘traceable’, ‘auditable’, and ‘preferably from a governmental source’. |
|
Non-discrimination and bias prevention |
Courts are required to implement measures to mitigate the risk of discriminatory biases, promote plurality, and ensure ‘that AI systems assist in fair trials’ by ‘minimising the marginalisation of individuals and judgment errors arising from bias’. |
|
Data protection |
The protocol mandates compliance with data protection regulations, requiring anonymisation and encryption, and prohibits the use of judicial data to train commercial AI models. |
The Resolution provides that the National Judiciary Committee for Artificial Intelligence is responsible for monitoring compliance with the principles and rules established in the Resolution. The Resolution does not prescribe any sanctions or disciplinary measures in the case of non-compliance, but there are general laws governing the conduct of judges which prescribe sanctions for improper or erroneous conduct. In the absence of specific rules, these would apply to e.g. the misuse of AI. However, practitioners noted that monitoring compliance with the Resolution remains challenging.
Brazilian Bar Association
In November 2024, the Federal Council of the Brazilian Bar Association (OAB) approved a set of recommendations on the use of generative AI in legal practice. These guidelines require lawyers to comply with applicable law and the OAB’s Code of Ethics and Discipline. They emphasise the lawyer’s obligation to ensure confidentiality and privacy as well as human oversight of AI. Before deploying AI, lawyers must inform their clients in writing about its intended use as well as the potential risks and obtain their explicit, informed consent. This is a client-focused disclosure requirement; there is no general obligation to disclose the use of AI to courts, opposing parties, or other third parties, nor to provide detailed technical information about the tools used.
Criminal procedure rules
There is currently no specific recognition of a need to reform the criminal procedure framework in isolation. Rather, policymakers and regulators have identified a broader need to adapt the wider legal and regulatory framework to address the risks and specificities of AI systems. Even though the Brazilian Criminal Procedure Code does not specifically target AI, its general rules – such as those governing the admissibility of evidence – apply equally to AI-generated or AI-assisted evidence.
Criminal Recognition Procedures
In 2026, the Ministry of Justice and Public Security issued Ministerial Ordinance No. 1122/2026, instituting a National Protocol, which regulates the formal identification procedures used in criminal investigations or prosecutions. The Ordinance applies mandatorily to the Federal Police and the National Public Security Force, and operates as optional guidance for State Civil Police forces.
The Protocol’s objectives include “reducing the risk of wrongful convictions” and “preventing discriminatory practices”. The Protocol requires safeguards such as continuous audiovisual recording and chain-of-custody measures, the use of multiple “fillers” (i.e. other individuals or images included in a live or photographic line-up who are known not to be suspects) with similar physical characteristics to the suspect – such as age, skin tone, build, hairstyle, or distinguishing features – so that the suspect does not stand out. The Protocol also states that the position of the suspect within the line-up should not follow a predictable pattern, and the overall composition of the line-up should avoid reinforcing racial or other visible biases. It also prohibits “suggestive” recognition practices, including suspect-only “criminal albums” (including images sourced from social media). The Protocol permits the use of AI tools to generate images for photographic line-ups, as long as visual equivalence, traceability and integrity are ensured – AI-generated images should not stand out from other images used as fillers, it must be possible to reconstruct how the image was created, and the generated images must remain unaltered after creation. Where AI is used, the authority must document the tools and parameters used, and retain the generated files so the procedure can be reviewed (including by the defence).
Data protection legislation
The General Data Protection Law (Lei Geral de Proteção de Dados) – Law No. 13,709/2018 applies to the processing of personal data by both public and private entities. It may be argued that any AI using personal data must comply with the protections established by the law.
Human Rights
The use of AI in criminal proceedings must be consistent with procedural guarantees and fundamental rights included in the Brazilian Constitution, including due process, equality before the law, the right to a fair trial and to privacy. Fair trial and privacy guarantees under regional and international human rights treaties to which Brazil is a party, such as articles 8 and 11 of the Inter-American Human Rights Convention, articles 14 and 17 of the International Covenant on Civil and Political Rights or articles 16 and 40 of the Convention on the Rights of the Child, may also be relevant.
Outlook
In December 2024, the Federal Senate approved Bill No. 2338/2023 on the development, deployment, and use of AI systems in Brazil. The proposed bill aims to protect fundamental rights, promote responsible innovation, ensure the implementation of secure and reliable AI systems that benefit people, democracy, and technological and economic development. It proposes a risk-based model, categorising AI systems into ‘excessive risk’ (prohibited), ‘high risk’ (permitted under strict conditions), and ‘low/minimal risk’. Within the proposed framework, excessive-risk systems include those that assess personal traits, characteristics, or past behaviours for predicting crime or recidivism, enable social scoring or real-time biometric identification in public spaces (except in narrowly defined circumstances, such as criminal investigations with prior judicial authorisation). Systems used in the administration of justice (excluding those used for administrative tasks), criminal investigations, and public security are classified as high-risk, triggering obligations that include algorithmic-impact assessments, governance measures, transparency, bias mitigation, human oversight, and detailed documentation. The draft bill also enshrines individual rights such as access to information about the use of AI, explanations of AI-driven decisions, human intervention, non-discrimination, privacy, and contestability.
However, the draft bill remains subject to change. It must still be scrutinised and voted on in the House of Representatives before presidential approval. As at March 2026, no expected date for the next legislative developments.
The bill is under consideration by a Special Committee created to report on the bill and is listed as ‘awaiting the rapporteur’s opinion’ in that Committee. The Chamber’s official tracking page logs the most recent legislative action as taking place on 23 December 2025, when the Special Committee approved a request (made by Federal Deputy for the state of Rio de Janeiro, Soraya Santos) for a public hearing to debate the use of AI technologies in Brazilian basic education.
Brazilian Artificial Intelligence Plan (PBIA)
The Brazilian Artificial Intelligence Plan (PBIA), titled ‘AI for the Good of All’, finalised in June 2025, is a national strategy coordinated by the Ministry of Science, Technology and Innovation (MCTI) with support from the Center for Strategic Studies and Management (CGEE) to position Brazil at the forefront of responsible AI development and application. The PBIA outlines a multi-year (2024–2028) investment of approximately R$ 23 billion (£3.3 billion) from public and private sources to strengthen AI infrastructure, research, workforce training, public-sector innovation and business uptake. It is structured around five strategic axes: (i) infrastructure and AI development, (ii) diffusion, training and qualification in AI, (iii) AI for improving public service, (iv) AI for business innovation, and (v) support for the regulatory and governance process of AI. The plan – which includes the launch of SoberanIA – is intended to guide long-term interministerial actions and complement ongoing legislative and judicial AI governance efforts to position Brazil ‘at the forefront of AI development, serving as a global example of the use of this technology for the benefit of all society’.
Deepfakes and synthetic media
Law No. 15,123/2025 (April 2025) introduced a targeted criminal law response to harmful AI-manipulated content. The law amended Article 147-B of the Penal Code (regarding psychological violence against women) to increase the penalty where the offence is committed using “artificial intelligence or any other technological resource that alters the image or sound of the victim”, an approach expressly aimed at conduct involving AI-generated or manipulated images, videos or audio (deepfakes).
Official communications accompanying the law explain that it seeks to address the misuse of technologies capable of creating or altering audiovisual content to simulate real people, with particular concern for gender-based harm. The legislation does not establish a general obligation to label or disclose synthetic content, nor does it set out technical standards for authenticating or verifying suspected deepfakes in criminal proceedings. Questions of authenticity and reliability therefore continue to fall to general evidentiary rules, including expert examination and existing safeguards against evidence tampering.
Cases
Policing and identification cases
As at March 2026, Brazilian case law is more developed in relation to unreliable identification evidence than to facial-recognition technology as such, though courts have also shown willingness to intervene against large-scale biometric surveillance on privacy and proportionality grounds.
On identification evidence, two key STJ decisions—HC 598.886/SC and HC 712.781/RJ—treated defective photographic recognition as incapable of safely grounding a conviction.
On biometric surveillance, the São Paulo Smart Sampa litigation is illustrative. A trial judge initially suspended the city's procurement of facial-recognition cameras, noting in his decision that the technology presented considerable risks to fundamental rights, with particular attention to its potential role in reinforcing structural racism. On appeal, however, the rapporteur Paola Lorena concluded that there was no evidence to support claims that video surveillance would aggravate social or racial discrimination, and the programme is now operational with 40,000 cameras across São Paulo. Separately, Espírito Santos public-security authorities reportedly arrested 32 wanted individuals in the first two months of a pilot project using AI-enabled facial-recognition cameras in public streets, launched in November 2024.
Misuse of AI in court filings
Brazilian courts are increasingly responding to the misuse of AI in court filings with bad-faith sanctions, referrals to the Bar, or sharp judicial warnings.
For example, in a Santa Catarina Court of Justice (TJSC) case publicised on 18 February 2025, the court imposed a 10% bad-faith litigation fine after identifying authorities in a filing; the appellants lawyer had admitted using ChatGPT, and the court observed that fabricated case law could mislead both the judiciary and the opposing party. At the STF level, Justice Flávio Dino in RCL 78.890/BA (May 2025) found that the pleading relied on decisions that could not be located and contained false statements; he summarily rejected the claim, imposed costs-related sanctions, and ordered communications to the OAB.
Lower courts have followed the same pattern. In Silva Cordeiro v Município de Santana de Parnaíba (TJSP, 25 July 2025), the court refused to entertain the appeal, imposed a one-minimum-wage sanction for bad-faith litigation, and notified the Bar after concluding that the submission relied on fabricated authorities. In a Regional Labour Court matter (ATSum 0010525-47.2025.5.03.0037), the court stressed that AI may be a support tool but not a mechanism for inventing false precedents, after counsel admitted drafting the pleading with AI without verifying its contents.
Deepfakes
As at March 2026, cases involving deepfakes in Brazilian courts have centred on AI-manipulated evidence in the context of elections.
In that regard, Brazil’s Superior Electoral Court (Tribunal Superior Eleitoral, TSE) has been actively shaping the regulatory and enforcement response to deepfakes and AI-manipulated content in elections. In public hearings held in February 2026 on the rules for the 2026 elections, the TSE has discussed additional safeguards, including proposals to impose fines of up to R$ 30,000 for the use of AI-fabricated or AI-manipulated content.
Brazilian electoral courts have already addressed manipulated media in a growing body of electoral litigation, particularly where such content threatens electoral integrity. Alongside enforcement, the TSE has also invested in public-facing prevention measures. Ahead of upcoming elections, it launched a voter education initiative aimed at helping the public identify false content, including the web series “V de Verdade (T for Truth) – In a land of facts, fake news has no place,” promoted through social media and major media outlets.
In preparation for the last municipal elections in 2024, the TSE amended Resolution No. 23.610/2019 (electoral propaganda rules) to introduce provisions specifically addressing AI in electoral campaigning. The amendments (i) restrict the dissemination of fabricated or manipulated content (including deepfakes), (ii) require clear identification/labelling of synthetic or AI-generated content, and (iii) provide for liability relating to the creation or dissemination of disinformation using AI tools.
Electoral courts have applied these principles in practice. During the 2024 municipal election cycle, an electoral judge in Mato Grosso do Sul imposed a R$ 10,000 (£1,600) fine on a pre-candidate/ex-mayor for disseminating a manipulated video attributed to a political opponent, allegedly depicting the opponent making disparaging remarks about the local population. The decision treated the content as unlawful electoral misinformation and ordered sanctions. In Pernambuco, the Regional Electoral Court (TRE-PE) upheld a conviction relating to the dissemination of a manipulated video in Garanhuns, maintaining a fine reported as R$5,000 (£800), an appellate-level confirmation that materially edited or “deepfake” content may breach electoral norms when used to mislead voters.
At the same time, Brazilian electoral courts have emphasised the need to distinguish materially deceptive AI-manipulated content from obvious satire or meme-style political expression, indicating that not all manipulated imagery will necessarily be treated as unlawful “deepfake” content in electoral disputes.