–Estefânia Maria de Queiroz Barboza, Constitutional Law Professor, Federal University of Paraná (UFPR); Gustavo Buss, PhD Candidate, Federal University of Paraná (UFPR)
Debates about judicial AI tend to fixate on whether algorithms, rather the judges, will end up writing judgments. This focus reflects a valid unease with the prospect of a ruling entirely authored by a machine with no human oversight. But it also narrows the conversation about AI in decision-making to the judgment document itself, while also implying that keeping humans, at least as arbitrators and proofreaders of AI suggestions, is enough to avoid it.
By doing so, we inadvertently let courts and regulators off the hook for everything that happens before the judgment is made public. These “accessory” decisions are different among themselves. Some are procedurally inclined. Formatting mandates, for example, can directly affect the syntax of the dispute. Others have a clearer impact on the outcome. Evidentiary rulings, for instance, can open or close doors for the litigants that may prove determinative. In contrast, algorithmic sorting mechanisms can steer a decision by shaping what the judge effectively sees. In fact, a growing share of what actually shapes a case – for example, which facts get read, which precedents get surfaced, which arguments reach the judge’s desk at all – happens in the backstage of decision-making.
That is why, we argue, the real contest over interpretive authority has quietly relocated. We make this argument through the lens of Brazil’s recent experience regulating judicial AI and slightly expand on it with comparative insights. Ultimately, we aim to start a long-lasting conversation.
Brazil’s judiciary has been under strain for some time, facing a growing caseload that continues to increase each year. In response, it has embraced AI systems in a functional, risk-tiered regulatory model. This means incentivizing AI tools for activities deemed merely instrumental to decision-making, mostly to aid triage of similar cases and precedent search at a national scale. This raises three difficulties. Precedent search is never neutral, since precedent must first be summarized and then compared. Risk-based regulation can fail to tell truly accessory functions apart from decisive ones. And inviting the parties most affected by these systems to comment is not the same as letting them decide. We take each of these difficulties in turn below, using Brazil’s regulatory experiment as the connecting thread.
Brazil’s vague definition of AI risks suggests that judicial AI systems warrant greater constitutional attention. The risks associated with AI in other fields – such as platform power concentration, algorithmic bias, and abuse of power – are equally pressing in the context of decision-making. At its core, the question remains: who holds the authority to assess competing arguments, and in whose eyes must decisions ultimately be justified?
A lot happens before judging
Sorting and precedent search are, formally, different tasks. However, functionally, they point to a common push to expedite “accessory” decision-making.
The pressure driving courts toward AI is often invoked as a key feature of Brazil’s judiciary. Overwhelmed by volume, it asks machines to sort, cluster, and triage all proceedings before a human even looks at the files. The country closed 2025 with 75.5 million pending cases and 40.9 million new filings that year alone. A volume that would strain any judiciary, however resourced. In response, the National Council of Justice (CNJ) rolled out a triage AI tool, Berna, to every court. The tool screened 30 million cases in 2025 and grouped 2.5 million of them into 353,000 clusters of similar claims.
Courts treat sorting as pre-legal housekeeping, confined to the backstage of decision-making. In other words, they must clear the docket before moving on to the ruling. But sorting already presupposes an answer to a legal question about which differences are worth preserving. Once that premise is granted, calling the task “housekeeping” or “accessory” is contrary to the cognitive process it entails.
Compression is not new. Common-law headnotes and syllabi have done this job for over a century, on the assumption that a human reader could tell a good digest from a misleading one. What is new is who the targeted reader is. Brazil’s version, the ementa, was standardized nationwide in late 2024, in part so that AI systems, not just human readers, could search it. Hence, AI adds, on top of the human summarization, an algorithmic form of machine ranking and processing. This is a highly difficult task and implies several biases. In Brazil, for example, the Council has already found that the actual ratio decidendi of the country’s highest court typically lies scattered across individual justices’ reasoning, not inside the summary available for AI searching. Thus, choosing which precedents to present as viable in a given scenario is itself a cognitive decision.
This is not only a problem in Brazil. Wherever precedent search runs through a headnote-like intermediary, the same gap opens between what a machine can search for and what a judgment actually decided. This is why a judge who has handed down “five pertinent precedents” has not been given a sample of the law. She has inherited someone else’s editorial judgment about what counted, filtered a second time by whatever the ranking model rewarded.
Silence can speak as well
Brazil’s answer to this problem seems more conceptually ambitious than most comparable rules. Resolution 615/2025 classifies AI systems not by their degree of automation but by the function they perform within adjudication. A tool that organizes information is considered low risk. One that produces a juridical conclusion, such as profiling a litigant, weighing evidence, or deciding that a precedent governs a set of facts, is deemed high-risk.
This is, independently, close to the logic of the EU AI Act, which flags as high-risk any system assisting judicial authorities in researching and interpreting facts and law. Both regimes point to the same conclusion: risk should track cognitive function, because the most pressing risk of AI in decision-making comes from performing the actual decision-making. Tellingly, both also share the same instability at exactly the point that matters most for precedent search. Brazil treats detecting “pertinent” precedents as low-risk and reserves high-risk status for the conclusion that a precedent applies. The EU similarly carves out “preparatory” tasks from its own high-risk tier.
However, deeming something pertinent is hardly a neutral conclusion. Establishing that a precedent is applicable requires reconstructing its holding, choosing a level of generality for the comparison, and judging which facts would have changed the result. These are precisely the elements of conviction that, in appellate practice, can sometimes divide judges sitting on the very same panel. Any regulatory scheme built on a low-risk/high-risk axis defined by function will eventually meet this same wall: the step it calls preparatory is the step where the case is actually decided, or at least where conviction is shaped, just before anyone with the authority to decide it gets a look.
It is important to raise the issue because it might otherwise go unnoticed amid the more evident risks associated with judicial AI systems. In a recent example, a Brazilian court annulled an appellate judgment because the AI-assisted document cited nonexistent precedents. We often hear about such cases of AI hallucination. However, courts have tools to address this kind of failure. The most relevant is that a judge or opposing counsel will eventually read the document and confront the error, prompting correction.
A much more dangerous error happens when the system fails to surface something it should have. If an AI system returns five precedents it deems most suitable for a case, out of three hundred candidates, scrutiny of those five is insufficient to shed light on the possible reasons for the exclusion of the other two hundred and ninety-five. No register of the discarded arguments will exist, making oversight almost impossible.
In the US, similar debates about actuarial risk tools in sentencing have emerged. In State v. Loomis, the Wisconsin Supreme Court was confronted with the use of a proprietary algorithmic tool to predict a defendant’s risk of recidivism during sentencing (the tool is called COMPAS). The court held that using COMPAS does not violate due process as long as it is treated as one of many sentencing factors, rather than the decisive metric. It also mandated disclosure of the tool’s use. Loomis, in other words, guaranteed the right to contest a score already produced but not the right to know what was never surfaced in the first place.
Still, the backstage of AI systems leaves many decisions unchecked. Human efforts to assess AI output, or even the idea of human-mediated explainability as the master remedy for algorithmic opacity, may be insufficient. A recent comparative study of the EU, Brazil, and China found that what counts as an adequate explanation depends heavily on the function a system performs. The right to an explanation may depend on the judicial AI’s specific decision-making function. Due process standards, therefore, need to go further, to address the silences AI produces and the choices it makes to disregard and exclude arguments.
Human or non-human?
Thus, as AI retrieval systems reallocate interpretive authority in ways formal risk categories cannot fully capture, the question of who governs those systems becomes correspondingly more important. Brazil’s arrangement is worth exploring. A National Committee on Judicial AI was established to oversee and regulate the Brazilian judiciary’s use of AI tools. It includes members of the judiciary, the bar association, the prosecution office, and public defenders. Nonetheless, only the first group has voting rights. The rest participate with a voice but no vote on how these tools are classified, audited, and deployed.
AI governance bodies across sectors and jurisdictions tend to invite all actors most affected by AI systems to comment while reserving decision power only to the institution that will operate them. Inviting comments from the actors most affected by an AI system is a meaningful safeguard, but it is not the same safeguard as giving them a vote. Because a system that ranks and selects precedents is already exercising the kind of interpretive power that adversarial argument was designed to check.
Scholars in the field of digital constitutionalism have long raised the issue of accountability and authority. The perils of letting data, platforms, and cross-border digital infrastructure shape our lives in almost every form have been extensively documented. Judicial AI tools, intended to facilitate decision-making, are another chapter in this romance. A retrieval AI system that quietly decides which precedents a judge will see affects authority and challenges accountability because it takes a slice of the judge’s interpretive authority. And it does so without a clear explanation of the arguments that may have existed between the input and the output in any given interaction.
Our final point is not one against using AI. Using AI systems as tools to boost productivity is part of judicial decision-making, as it is in any other field. Yet we pointed to a sharper question than the one currently shaping the debate at the opening: who holds the authority to assess competing arguments, and in whose eyes must decisions ultimately be justified? It may be the case that, soon enough, nearly all elements of decision-making will be automatable. This does not mean we should let it be so. Judgments, however automatable, should be kept human and transparent as a constitutional mandate. For us – and we believe this extends to most constitutional systems – it is as important for an ethical and transparent use of AI that every element of the cognitive reasoning, especially the weighing and selecting of competing arguments, be fully explained and available for the parties to challenge.
Suggested citation: Estefânia Maria de Queiroz Barboza and Gustavo Buss, The Politics of Retrieval: Accessing the Backstage of Judicial AI, Int’l J. Const. L. Blog, Oct. 15, 2026, at: http://www.iconnectblog.com/the-politics-of-retrieval-accessing-the-backstage-of-judicial-ai/
Leave a Reply