AI in the Brazilian Judiciary: what CNJ Resolution 615/2025 requires of the courts

What CNJ Resolution 615/2025 requires for AI in Brazilian courts: high and low risk, bans, market generative AI and incident reporting within 72 hours.

Updated 14/nov/2026 · 13 min read

The short answer

The use of AI in Brazil’s Judiciary is governed by CNJ Resolution No. 615 of 11/mar/2025, published on 14/mar/2025. It repealed Resolution 332/2020 and took effect 120 days later, in July 2025. The current wording is the one compiled with Resolution 674/2026, which changed the makeup of the National Committee on Artificial Intelligence in the Judiciary.

The rule rests on four ideas. Classify the risk of every solution (high or low) and ban those of excessive risk. Keep a person in charge, with human oversight and the final word left to the judge. Register and show what is in use, in a national registry called Sinapses. And regulate market generative AI, ChatGPT and its relatives, which the rule allows but only with conditions and with the court knowing. It is the same logic as the AI governance asked of a company, written as articles.

This guide summarizes what the official text requires, article by article. It is not legal advice. For AI in the federal Executive Branch, which has a different set of rules, see AI in the federal government.

72 hours

is the deadline for Judiciary bodies to report any adverse event tied to the use of AI to the National Committee on Artificial Intelligence in the Judiciary, counted from the moment it is identified. The report includes a description of the incident, its causes and the measures taken.

Conselho Nacional de Justiça (CNJ), Resolução CNJ nº 615, de 11 de março de 2025: diretrizes para o desenvolvimento, utilização e governança de soluções com inteligência artificial no Poder Judiciário (texto compilado com a Resolução nº 674/2026), art. 42, § 2

Who the rule applies to

The resolution applies to the Judiciary: it covers the development, governance, auditing, monitoring and responsible use of solutions that use AI techniques in the courts (article 1). It respects each court’s autonomy to create local solutions, provided they follow the rule’s audit, monitoring and transparency standards (article 1, paragraph 1). Auditing is proportional to impact and does not require unrestricted access to source code (article 1, paragraph 2).

Oversight and guidance come from the National Committee on Artificial Intelligence in the Judiciary, created by the resolution itself (article 15). It sets the guidelines and criteria for risk assessment (article 9, paragraph 2) and can order a solution to be reclassified (article 9, paragraph 3). Private companies are not a direct target of the rule, but they feel its effect through contracts: see the section on generative AI and the FAQ.

Risk classification: high and low

Every court that develops or buys an AI solution must assess its level of risk (article 9), based on the impact on fundamental rights, the complexity of the model, the intended uses and the amount of sensitive data. The assessment should preferably happen during testing and approval or, for low-risk ones, at the start of internal production. The resolution’s Annex lists the cases.

RiskCodePurpose, in short
HighAR1Identifying profiles and behavioral patterns of people or groups
HighAR2Assessing the adequacy of evidence and weighing it in contentious proceedings
HighAR3Investigating, weighing and classifying facts as crimes, misdemeanors or juvenile offenses
HighAR4Reaching conclusive judgments on how a rule or precedent applies to concrete facts, including quantifying damages
HighAR5Facial or biometric identification and authentication to monitor people’s behavior
LowBR1Routine procedural acts and administrative support: classifying, grouping, transcribing and summarizing documents, with human supervision
LowBR2 and BR3Detecting decision patterns and precedents; reports and analyses to inform judges, without replacing human assessment
LowBR4Producing support text to help draft judicial acts, with the final version done by the judge
LowBR5 to BR8Improving a human activity without changing the result, statistics for judicial policy, audio and video transcription, document anonymization

High-risk solutions undergo regular auditing and continuous monitoring (article 11, paragraph 1) and, before going into production, must meet measures such as representative training data, a record of the level of human supervision, documentation in plain language, logs and explainability wherever technically possible (article 13). They also require a continuous algorithmic impact assessment (article 14), whose summary is made public in Sinapses (article 24, paragraph 1). Low-risk ones are monitored and periodically reviewed to make sure they stay low (article 11, paragraph 3). The high-risk list is reviewed at least once a year (article 11, paragraph 2).

What is banned

Article 10 bans the Judiciary from developing and using solutions because of excessive risk to information security, fundamental rights or the independence of judges. There are four:

  • No human review. Solutions that do not allow human review of results across training, development and use, or that create absolute dependence of the user on the result, with no way to change or revise it.
  • Profiling to predict crime. Solutions that weigh personality traits, characteristics or behaviors of people or groups to assess or predict crimes or reoffending in the grounds of decisions, and to ground labor decisions on personal profiles.
  • Ranking people. Solutions that classify or rank people by behavior, social situation or personality attributes to assess the plausibility of their rights, judicial merits or testimony.
  • Emotion recognition. The identification and authentication of biometric patterns to recognize emotions.

Courts must monitor continuously so that no solution falls into these bans unnoticed. If that happens during use, the solution is discontinued and the reasons and measures taken are recorded in Sinapses for the Committee’s review (article 10, paragraphs 1 and 2).

Human oversight and transparency

Human oversight appears from the foundations (article 2, V) and the principles (article 3, VII). Article 32 says the intelligent system must allow detailed review of the content generated and the data used, with no binding to the solution the AI presents, and that at no point may the AI restrict or replace the final authority of internal users.

For people who receive the service, article 33 requires telling them clearly, in plain language, that AI is part of the service and that its proposal is advisory and non-binding, always subject to the decision of a competent authority. Mentioning AI use in the text of a decision is at the signer’s discretion (article 33, paragraph 3). The rule also addresses bias: article 8 requires continuous validation of solutions and, if discriminatory bias cannot be eliminated, the solution is discontinued and its registration in Sinapses is canceled.

Generative AI: what judges and staff may use

This is the part closest to what many companies know as Shadow AI. Article 19 allows judges and staff to use large and small language models and other generative AI systems available on the internet as an aid to management or to support decisions. The preferred route is access enabled, provided and monitored by the court (paragraph 1).

If the court does not offer a corporate solution trained and tailored for the Judiciary, the judge, staff member or contractor may sign up for one by private subscription or registration (paragraph 2). In that case the rule sets five conditions (paragraph 3):

  • Training. Specific training on good practices, limitations, risks and ethical use, under a standardized digital literacy program.
  • Support use only. The tool supports the decision and never decides on its own. The judge remains fully responsible for the decisions and the information in them.
  • User data protected. The vendor may not process, use or share the data the Judiciary’s user provides, or data inferred from it, for training or improvement.
  • No confidential material without protection. The private tool may not be used with confidential documents or data, or data under judicial secrecy, unless anonymized at the source or unless mechanisms ensure effective protection.
  • Nothing high-risk. The private or external tool may not be used for purposes the rule classifies as excessive or high risk.

On top of the conditions there is a transparency duty: someone who buys a market solution on their own, or a manager with a team member who uses one, must report that use to the court (paragraph 7), and courts consolidate it and send it to the Committee (paragraph 8). When generative AI helps draft a judicial act, the use may be mentioned in the decision, at the judge’s discretion, but automatic recording in the court’s internal system is required (paragraph 6).

The lesson for a company is direct: the rule does not ban the tool, it requires the institution to know who uses what. That is what an AI acceptable use policy and an AI inventory solve. For a court that buys generative AI, article 20 lists what it must require of the vendor: compliance with Brazilian law, the LGPD and the resolution; use of the Judiciary’s data for training only on the LGPD’s legal bases; confidentiality and acceptance of external audit; privacy by design and by default; and a ban on using confidential data to train models, except after anonymization at the source.

The Sinapses registry and transparency

Sinapses is the CNJ platform that keeps the catalog of the Judiciary’s AI solutions, organized by risk category. Every solution, in development or in use, must be registered (article 24). The timing of the registration changes with the risk:

  • Low risk: before going into production, with minimal information such as the purpose, whether it is in-house or collaborative, whether it was bought or built internally and a description of the goals (article 24, paragraph 3).
  • High risk: after the preliminary studies but before development starts, updated at each new phase or relevant version (article 24, paragraphs 4 and 5).

The CNJ publishes on its website the list of applications in use (article 25). In electronic case systems that use AI, the main interface must show the models in use, the version, the Sinapses registration code and the date of the last update, reviewed at least every 12 months (article 21). Products generated automatically record the use of AI in the logs with identification labels (article 21, paragraph 2).

Monitoring, audits and incidents

The Committee monitors, without disciplinary effect, use that does not follow the rule and may call for an audit of inadequate practices, misuse of data and lack of transparency (article 40). It also defines the audit and monitoring protocol (article 41) and, when it finds a nonconformity, sets a deadline for correction according to its seriousness.

Article 42 requires bodies to report all adverse events, meaning incidents with a negative impact on the operation of the system, data security or service delivery, within 72 hours of identification. Article 39 completes the picture with accountability: the names of those responsible, costs, intended and achieved results and disclosure of the nature of the service, the techniques used, performance and the risk of errors, in plain language and on an official channel.

Deadlines and what changed from Resolution 332

ItemWhat the rule says
PublicationResolution dated 11/mar/2025, published in the CNJ official gazette on 14/mar/2025
Effect (article 47)120 days after publication, in July 2025
Adjustment (article 45)Also applies to projects already in development or deployed; courts had 12 months from publication to adjust, that is, until March 2026
Repeal (article 46)CNJ Resolution 332/2020 stopped applying when 615 took effect
Latest amendmentResolution 674/2026, of 25/mar/2026, which changed the makeup of the Committee (article 15)

Compared with 332, Resolution 615 added classification into high and low risk with its own Annex, the algorithmic impact assessment, an entire chapter on generative AI, mandatory registration in Sinapses, the explicit ban on four types of solution and incident reporting within 72 hours.

Where to start: eight steps

For a court, or for a company that serves one, a practical reading of the rule fits in eight steps:

  • Take inventory. Know which AI solutions exist, in development or in use, and register them (article 24). Without a list there is nothing to classify.
  • Classify. Set the risk level of each one using the Annex and the Committee’s criteria (article 9).
  • Name an owner. Appoint people or committees responsible for following security and transparency (article 12, III).
  • Write the rule for generative AI. Say what is allowed, what is not and how to ask (article 19). A template is in the AI acceptable use policy guide.
  • Train. Train judges and staff on risks, bias and critical analysis of what AI delivers (article 3, VIII, and article 19, paragraph 3, I).
  • Record use. Keep logs and labels of AI use and of the information given about market tools (article 19, paragraphs 6 and 7, and article 21, paragraph 2).
  • Prepare incident response. Have someone who reports, and a way to do it, to meet the 72 hours (article 42).
  • Review contracts. Include the clauses article 20 expects from an AI vendor.

Where Tangerin AI fits

Tangerin AI is the independent AI governance platform that discovers, assesses, governs and verifies AI use across the entire company, from any vendor.

For those who must follow Resolution 615 or prove compliance to a court, the platform covers the part about knowing what is in use: it discovers the AI tools people actually access, classifies each one by exposure level, applies the use policy with a recorded acknowledgment and keeps the evidence. The Compliance module includes CNJ Resolution 615/2025 among the frameworks it tracks, with controls tied to inventory, policy, classification, human oversight, rules for generative AI and incident reporting. Tangerin AI is a private company, is not part of the CNJ and does not replace registration in Sinapses or the court’s own decisions. See the plans and the 14-day trial, no card required.

Limits of this guide

  • It applies to Brazil only. This guide covers the Brazilian rule. Other countries have their own rules for AI in the judiciary, and they are not covered here.
  • It is not legal advice. It is a summary of the official text. To decide a specific case, consult the text of the rule and the legal team of the court or company.
  • It is the text as of 14/nov/2026. The text read is the one compiled with Resolution 674/2026. The CNJ may amend the rule, and the Committee still defines guidelines and criteria (such as those for risk assessment) that are not in the text of the resolution.
  • It covers the Judiciary. It does not deal with the Executive, the Legislature, the Public Prosecution Service or states and municipalities, which follow other rules.
  • It does not measure compliance. The guide says what the rule requires, not how many courts comply.

Frequently asked questions

What is CNJ Resolution 615/2025?

It is the rule of Brazil’s National Council of Justice, dated 11/mar/2025, that sets the guidelines for the development, governance, auditing, monitoring and use of artificial intelligence solutions in the Judiciary. It sorts solutions into high and low risk, bans four types of solution, regulates the use of generative AI by judges and staff, requires registration in the Sinapses platform and requires adverse events to be reported within 72 hours.

Is CNJ Resolution 332/2020 still in force?

No. Article 46 of Resolution 615/2025 repealed Resolution 332/2020 from the moment the new rule took effect, 120 days after publication (article 47), in July 2025. Anyone still citing 332 as the current rule is out of date.

Can judges and court staff use ChatGPT and other generative AI?

Yes, with conditions. Article 19 allows large and small language models and other generative AI systems available on the internet to be used as an aid to management or to support decisions, preferably through access enabled and monitored by the court. If the court does not offer a corporate solution, the judge or staff member may sign up for one on their own, provided they have done the required training, use the tool only as support, do not allow the vendor to train on the data, do not use it with confidential documents unless anonymized and avoid it in high-risk uses. They must also report the use to the court.

What does the rule treat as a high-risk solution?

The Annex lists five high-risk purposes: identifying profiles and behavioral patterns of people, assessing and weighing evidence, investigating and classifying facts as crimes, reaching conclusive judgments on how a rule applies to concrete facts (including quantifying damages) and using facial or biometric identification to monitor behavior. High-risk solutions require continuous audit and monitoring and an algorithmic impact assessment. Low-risk ones, such as transcription, document summaries and anonymization, are also monitored and periodically reviewed.

Does the rule apply to companies that provide services to the Judiciary?

Directly, it binds the Judiciary’s bodies, not companies. But a court that buys AI must require the vendor to follow Brazilian law, the LGPD and the resolution itself (article 20), with compliance secured by contract and monitored (article 22, paragraph 1), confidentiality of the information and the possibility of an external audit. Anyone selling to courts should expect those clauses.

Sources

  • National Council of Justice (CNJ), Brazil. CNJ Resolution No. 615 of 11 March 2025: guidelines for the development, use and governance of artificial intelligence solutions in the Judiciary (compiled text with Resolution No. 674/2026) (2025). https://atos.cnj.jus.br/atos/detalhar/6001

Found it useful? Share it

LinkedInX

Instagram does not open links from outside the app, so we copy it for you to paste in a story or bio.

Where does your company stand today?

The free assessment has 12 questions (3 about your company and 9 on maturity), about 3 minutes, with no signup. It returns your maturity level per dimension and a prioritized action plan, enough to know where to start without buying anything.