When migrating last year from a court responsible for cases to a civil court, judge Roberto Luiz Corcioli Filho, 43, says he already made the change based on the assumption that he would incorporate the use of to deal with the number of cases.
“I would only go because I was willing to explore this world of AI. Because it is an absurdly large amount of work, with enormous repetition, and you have to complete it within a suitable period of time,” he says.
Corcioli reports that he spoke with other colleagues, started watching videos and reading about the topic and that, when he took on the new position, he spoke openly with his team to build a workflow that incorporated the use of artificial intelligence.
“It’s not a blind delegation, far from it. I’ve always been very focused on this issue, that we’re always checking everything”, says he, who is head of the 2nd Civil Court of the Santo Amaro Regional Court in the first instance, but who is temporarily working in the second instance of the special Public Finance Court.
If the image of a judge working alone was no longer reality before, given the work done together with his assistants, with the advancement of generative AI tools, this, control and safety in its use, but regarding ethical limits.
A Sheet spoke with judges and assistants from different areas about how this technology has been used.
While there are cases like Corcioli’s, there are also almost opposite situations, such as that of judge Fabiana Alves Rodrigues, 51, who works at the 1st Federal Court of the TRF-3 (Federal Regional Court of the 3rd Region) and encourages her team, in general, to avoid using AI.
She asked her assistants, however, to notify her if they use the technology for any task and for what purpose, whether it was used to make a summary or to correct a text, for example. “When people do it by hand, I know what each person’s weaknesses are and what I need to pay more attention to. If it comes from AI, sometimes it seems like it’s good, but that could be an invention”, he reflects.
Fabiana believes that the relevance of AI varies greatly depending on the work, adding that, in the court in which she works, despite being more complex cases than in courts with repetitive demands, the volume is not so large, making it easier to do what she classifies as more artisanal work.
Corcioli says that he agreed with the rest of the team that everyone would start using the Gemini tool and that he polished a prompt – jargon used for the commands that are entered into an AI tool when placing an order – which included aspects that should be analyzed in different types of processes and understandings that he usually adopts.
In general terms, the judge reports that, in more complex cases, he performs an initial analysis and then talks to the AI, generally already giving some type of direction. In more repetitive demands, he says that sometimes he just inserts the general prompt and that, from there, he and his team hold a conference based on the process and evaluate whether or not the path adopted was appropriate and what needs to be adjusted.
TJ-SP judge Guilherme Madeira Dezem, 50, also started to adopt AI in his office, including to help with drafting sentences. He currently hires the Claude tool, but believes that the best option would be a quality AI developed by the court.
Madeira Dezem reports that he reads the process, decides the case, and only then gives a command to the AI to formulate a draft based on one of its previous models. After that comes the review process. He says he is completely against having any dialogue with the AI about the merits of cases before deciding. “I think this gives the judge a bias.”
In the case of confidential processes, he reports using an anonymization tool made available by the court to remove personal injuries.
He also argues that it would be important to make it mandatory to inform the decision that AI was used, which is currently optional. “If one judge informs and the other doesn’t, it seems to me that it could cause some confusion”, he says.
Challenges in regulation and control
In February 2025, the (National Council of Justice), which became permitted, as long as human supervision was maintained and the decision-making process was not delegated to technology, among other rules.
A recent case showed the types of risks involved in the event of inappropriate use of this technology.
The news that lawyers are including hidden commands in processes to try to manipulate the tools used by judges, with the aim of obtaining favorable opinions, raised an alarm and caused different bodies to issue notes on the subject. Strategies include including metadata in attached images or commands in white — imperceptible to the naked eye, but readable by AI.
Last week, the CNJ approved a protocol in this regard. Among the measures are the adoption of safeguards in the commands and procedures adopted so that this type of manipulation is avoided, in addition to close supervision.
The challenge of controlling those who do not adopt these precautions, however, remains.
An advisor from the first state instance heard subject to reservation by the Sheet who works in an office where there is constant use of AI, encouraged by the judge, reports that one of his colleagues had a significant increase in productivity after adopting AI, but along with this came errors, without there being an attentive review by the superior.
This advisor has also come across a sentence in which the names of the witnesses cited were people who were never part of the case in question. It was necessary to correct the mistake after questioning from one of the parties.
Judge Rafael Niepce, heading the TJ-MG AI committee, the second largest in the country in terms of number of judges, says that there is an effort to train people on the ethical use of technology. The Minas Gerais court is one of those that currently has its own AI tool that allows it to generate minutes, summarize documents, and the court’s own processes are already linked to it. Niepce explains that, with this platform, there is the possibility of auditing, as it is recorded which judge used it, in which process and even the commands used.
However, reality shows that even the existence of this type of tool does not fill the gap regarding the difficulty of control. A judge’s assistant interviewed under reserve by the reporter who works in the interior of Minas says that, even after the launch of the TJ-MG assistant, the office in which he works continues to use private tools such as , without being informed.
According to the CNJ resolution, magistrates who hire private AI tools must inform the court of which they are part, which must then forward this data to the national committee. Currently, however, the courts are not yet providing this information.
CNJ advisor Rodrigo Badaró, who chairs the Judiciary’s National Artificial Intelligence Committee, tells Sheet be against the use of private tools by the judiciary and is even considering proposing that the CNJ vote on a rule in this regard. He says that this use was permitted by resolution last year, but that since then different institutional and board-audited tools have been developed.
According to data from the CNJ, from the Sinapses platform, which concentrates AI initiatives across almost 90 courts in the country, records of systems that generate text, such as orders, sentences and votes, were identified in 31 courts, some of which are in the development or testing phase. In addition, there are AI initiatives for various purposes, such as process screening, identifying bad faith litigation and prompt banks.
The highest court in the country is structuring the governance model for the use of AI in accordance with CNJ rules, says judge Tom Alexandre Brandão, advisor to the presidency. He reports that today Copilot is already available, and that the court is starting to hire a new, more robust AI tool that will be integrated into the court’s internal system, allowing the judge to access processes directly there, without the need to attach files.
It also adds that current internal systems are auditable, and that a platform to gather information about judges who are using private AI tools is under construction.
FGV Direito SP professor Marina Feferbaum, who coordinated research on AI in the legal universe, points out that it is necessary to advance both in the Judiciary and in law with greater governance and more detailed rules about what can and cannot be done.
She also highlights that there were reports during the research regarding the difficulty of measuring what the return generated by the use of AI would actually be. “Is it saving me time or money? Is it giving me precision?”, he reflects. “It’s also something interesting to start investigating in the Judiciary, what value are we bringing and what are we losing too?”
Federal judge Marcelo Lelis de Aguiar, 50, who works in a federal court in Sorocaba and expects to use the AI provided by the TRF-3 itself, in addition to providing training, says that the technology helps to save time and increase productivity. He argues, however, that it is necessary to continue reading the processes in full, albeit dynamically, so that human beings remain in control.
“We wouldn’t have an evolution in our jurisprudence if we just kept taking past data and replicating it, which is what AI does. The evolution of our jurisprudence was precisely breaking paradigms. Changing the status quo for something new”, he assesses. “The criterion of justice is a human criterion.”
