Detailed example of AI applied to law and justice.

   


Detailed real-world example: Brazil’s Supreme Federal Court (STF) using AI to triage and draft in constitutional appeals

A concrete case study is the STF’s use of VICTOR (machine learning + NLP) to screen “extraordinary appeals” and identify whether they meet the constitutional admissibility requirement of “general repercussion” (repercussão geral)—i.e., whether the issue has broader social relevance and should be heard by the STF.

What the “case” looks like in practice (end-to-end workflow)

1) Intake: an appeal arrives at the STF

An extraordinary appeal is filed (typically a high-volume category). Court staff must quickly decide whether it’s the kind of matter the STF should take forward.

2) Document processing + classification with VICTOR

VICTOR’s pilot purpose is to distinguish appeals that likely involve “general repercussion” from those that do not, using document analysis / NLP and ML components (the paper discussing VICTOR mentions techniques including image-sorting approaches and XGBoost in its stack). 

It classifies incoming appeals into categories such as:

  • appeals with general repercussion,

  • appeals already dismissed for lack of general repercussion,

  • appeals where there is insufficient information to determine it (often because the database/model is built from previously-addressed themes). 

3) Speed & scale benefits (triage time collapses)

This is where the efficiency gain happens: the OECD describes that a clerk may take ~44 minutes to evaluate whether an appeal meets the “move forward” conditions, while VICTOR does it in seconds. 

The same write-up on VICTOR reports screening performance around 84% (and potentially up to ~95%) for identifying general-repercussion topics (as reported in that discussion of the system). 

4) Human review (the decision is not automated)

The STF framing is that VICTOR supports clerks in triage rather than issuing binding judicial decisions. It’s used to accelerate staff work: identify likely themes, parties/object, and route cases more efficiently—then humans confirm and proceed. 

5) GenAI drafting support (a second layer: “MarIA” for reports)

Separately from VICTOR, the STF also experimented with a GenAI-based assistant (“MarIA”) that drafts clerks’ reports, which the clerk then reviews and edits. In a reported description, the STF tool used Google Gemini and OpenAI ChatGPT models under the hood. 

6) Outcomes—and the “side effect”

Reportedly, STF productivity improved: by June 2025, its backlog was said to have dropped to the lowest level since 1992 (citing an STF productivity report). 
But the same reporting also highlights a systemic side effect: as lawyers adopt GenAI, filings can increase—CNJ data cited there indicates tens of millions of new lawsuits and a large jump since 2020. 

The key risk (and why governance matters)

GenAI can “hallucinate” citations. Oxford’s Technology & Justice project describes a Brazilian incident where a judge signed a draft prepared with ChatGPT that included fabricated jurisprudence, contributing to a wrongful conviction, and the CNJ moved to investigate. 

What makes this a strong “AI-in-justice” example

It combines:

  • Classic AI/ML for case triage and routing (VICTOR), and

  • GenAI for drafting summaries/reports (MarIA),
    …with humans retaining formal decision authority—but with real operational and accountability consequences. 

Sources for this case study (VICTOR + GenAI in STF): 

https://restofworld.org/2025/brazil-ai-courts-lawsuits/

https://www.techandjustice.bsg.ox.ac.uk/research/brazil

https://www.oecd.org/en/publications/2025/06/governing-with-artificial-intelligence_398fa287/full-report/ai-in-justice-administration-and-access-to-justice_f0cbe651.html





Comments

Popular posts from this blog

IBM goes big to keep quantum computers cold

IBM - avances en mantener adecuadamente fríos computadores cuánticos.

Estructura multinivel de los sistemas de agentes de IA