BRAZIL Trends and Developments Contributed by: Eduardo Perazza, Glaucia Coelho, Débora Chaves Martines Fernandes and Agnes Domingues, Machado Meyer
The limits of automation: complex cases that do not fit neatly into boxes The congestion in the higher courts is an indisputable problem and one that goes beyond mere adminis- trative overload. It directly affects access to justice. When thousands of cases are piling up, the time avail- able for each one shrinks, the adjudicator’s capacity to delve deeply into the matter diminishes and the quality of the judicial function is inevitably compro- mised. Delay harms everything: careful examination, reflection and substantive adversarial debate. It is in this context that the strength of AI becomes undeni- able because it delivers speed that no human team could replicate at the same scale. According to the STJ Presidency Report, the Advi- sory Office for Admissibility, Repetitive Appeals, and Relevance (ARP) – a unit created to prevent mani- festly inadmissible appeals from even reaching the justices’ chambers – has blocked more than 120,000 cases since August 2024. In the Second Section, for example, each justice was spared an average of 6,722 cases, equivalent to roughly 83% of the total (8,059). This is a number that, read from one perspective, is reassuring. The court is not and does not have to be open to every appeal, and congestion is being con- tained. Without this mechanism, the workload would simply be unworkable. Read from another perspec- tive, however, the same figure is troubling. Of the inadmissibility decisions issued by the ARP, 95.44% are upheld by the Panels. Fewer than 4% of deci- sions are reversed. In a system where access to the STJ is already, by definition, exceptional – reserved for questions of federal law and filtered through rigor- ous admissibility requirements – this means that the screening decision is, in most cases, the final deci- sion. There is no effective second chance. The appeal that does not pass through the AI filter will not be judged on the merits; it will simply be terminated. The gateway has, in practice, become the judgment itself and the AI operating at that gateway does not forgive imprecisions that a human rapporteur, in another con- text, might overlook. But what about when the dispute does not fit neatly into a box? Consider a case in which a party raises a novel constitutional theory, for example, the intersec-
judge. Human oversight is required at every stage of the decision-making process, the autonomous use of AI to issue judicial decisions is prohibited, and each system deployed must be classified by risk level and remain auditable and transparent. On the legislative front, Bill No 2,338/2023 proposes a comprehensive Legal Framework for AI in Brazil but contains no spe- cific provisions for the judiciary, leaving sectoral regu- lation to the courts’ own rulemaking and resulting, for now, in a patchwork where each court adopts its own policies with visible risks of inconsistency. When AI gets it wrong The pattern first came to international attention through Mata v Avianca (2023), in which lawyers in the United States submitted AI-generated briefs con- taining entirely fictitious judicial precedents and were sanctioned by the court. Brazil was not long in pro- ducing its own versions of the same episode. In ruling on Appeal 78,890/Bahia, STF rejected a peti- tion drafted with AI assistance that contained false precedents and incorrect content from binding prec- edent summaries. The conduct was classified as pro- cedural bad faith, the party was ordered to pay double costs, and the court ordered notification to the Federal Bar Council for appropriate disciplinary proceedings against the lawyer. The Superior Labour Court (TST) has imposed fines for bad-faith litigation involving AI-generated ficti- tious precedents, and the severity of sanctions has escalated. Courts have ordered several referrals to the Federal Prosecution Service for assessment of poten- tial criminal liability (for example, case No 0000284- 92.2024.5.06.0351) – signalling that the judiciary is already seeking possible frameworks in existing crimi- nal law, even in the absence of a specific offence for fraudulent use of AI in court filings. What emerges is a grey zone. The lawyer practising before the superior courts today operates on three fronts simultaneously: competing with firms that already use AI to produce more, faster and at lower cost; litigating before a judiciary that has automated the critical stages of the appellate path; and doing so in an environment where the rules governing the tool itself are still being written.
76 CHAMBERS.COM
Powered by FlippingBook