To Keep Human Judgment, We Need AI Courts

The systems we use to enforce our rights and hold each other accountable are failing to keep up. Many legal cases in India take more than a decade to resolve, with a backlog now exceeding 54 million cases. Brazil’s backlog is even larger: 75 million pending cases, 40 million of which were filed last year.

Add AI into the mix and the problem of overwhelmed courts is liable to get much worse. When a plausible legal document costs almost nothing to write, courts will see more cases as well as more filings in each case. Economists Anand Shah and Joshua Levy offer credible evidence that this is already happening in U.S. federal courts. Before ChatGPT, about 0.1% of complaints contained AI-generated text. By early 2026, 18% did. And the share of cases brought by self-represented litigants jumped from a historical average of 11% to 17% in 2025.

To be clear, increasing access to justice is a good thing. Far too many people cannot afford basic legal assistance, and AI can and should be used to help bridge that gap. Yet as dockets grow, judges will either have less time to resolve cases, likely resulting in worse decisions, or take longer to decide, which means longer waits.

Overwhelmed court systems might turn to AI tools to increase productivity. Done properly, this can certainly help, but it also risks judges (or clerks) leaning on AI more than they mean to. Rather than simply hope judges keep up or pressure them to adopt AI, courts should create a parallel AI adjudication track.

This track, inspired by AI use in private arbitration but formally part of the court system like magistrate judges, would strengthen the role of humans in the judiciary. It would allow courts to be explicit about how human judgment is allocated, preserving human attention for the cases that need it. Many state supreme courts could launch a pilot now and start collecting evidence about where human judgment matters most.

How might AI-based adjudication work?

Arbitration is one example of a parallel dispute resolution process in the United States. Much like judges, arbitrators resolve disputes between parties. Yet unlike courts, they operate under rules the parties select, issue no binding precedent, and compete for parties’ business. Under the Federal Arbitration Act, private arbitrators can issue binding decisions that are enforceable in courts.

Because arbitration is a private dispute resolution process, there are few barriers to using AI. In fact, the American Arbitration Association (AAA) has already launched an AI-based arbitration process for documents-only cases related to the construction industry.

In this process, AI agents ask parties to explain their dispute and upload relevant documents as supporting evidence. An agent repeatedly clarifies its understanding of a party’s claim, prompting the user for more information where necessary. Once both parties are satisfied, the two AI agents present the parties’ claims and evidence to a different AI agent, customized with domain knowledge from the AAA. This agent functions as a clerk, drafting an award that is handed off to a human arbitrator to confirm or edit before issuing. The process is significantly faster and less expensive than traditional human-led arbitration, and the agents’ logs are auditable by humans.

While an AI arbitration process could save courts and parties time and resources, it is not a panacea for overburdened courts.

First, courts are limited in how they can funnel cases into arbitration. Judges can suggest arbitration, but both sides must consent for it to be binding. (State programs like Arizona’s, discussed below, can compel arbitration only because either party can demand a new trial afterward.) This makes it hard for courts to redirect enough disputes to meaningfully reduce caseloads. And the cases that do end up in the public court system might simply be those where people couldn’t afford private AI arbitration, rather than the ones where human judgment is needed most.

Second, private arbitration has a reputation for being exploitative because many of its agreements are forced upon consumers in terms of service they don’t read. Companies usually choose the arbitration provider, often outlined in those same terms, which incentivizes providers to favor companies over individuals. The arbitrators themselves face similar pressure: in one study, business-friendly arbitrators were about 40% more likely to be picked for the next case. Publicly funded court systems don’t have these same incentive problems.

And third, arbitration is sealed off from lawmaking. Plenty of legal forums don’t set precedent (small claims courts, magistrate judges), and the parallel track proposed here wouldn’t either. But in those forums hard questions can still reach judges. Arbitration lacks this path because awards are usually confidential, often come without reasons, and are only reviewed by courts for problems like fraud or bias. A court-run AI track could instead publish decisions and certify important legal questions to a human judge.

These problems all come from arbitration being private. A better path is to bring what works in AI arbitration (speed, experimentation, an auditable record) into public courts, though this will require answering a hard question that arbitration sidesteps: which cases need human judgment?

Which cases would be suitable for AI?

Identifying the cases where human judgment matters is easier said than done. There are two ways to go about it: a principled approach and an empirical one.

1. The principled approach: select cases with low stakes or that are less related to life, liberty, or fundamental rights.

A principled approach starts by asking: where is human judgment critical to a functioning justice system? Cases where someone’s life or liberty is at stake, for example, likely require a human judge. Death and incarceration are such severe punishments that the U.S. Constitution guarantees the right to a jury trial for nearly all defendants facing them. I doubt many people would want AI judges to decide serious criminal cases or sentencing lengths, and doing so would likely be unconstitutional anyway.

Next come cases related to fundamental rights and duties. Obergefell v. Hodges, the 2015 case that legalized gay marriage, and Dobbs v. Jackson, the 2022 case that eliminated the constitutional right to an abortion, both relate to deeply held values. In the private law context, United States v. Carroll Towing Co. altered the duties private parties owe to each other by proposing a new standard of care in negligence suits.

To turn these cases over to AI would mean turning over a core aspect of what it means to be self-governing.

This category of cases, though its boundaries are fuzzy, concerns fundamental rights. Who can we marry? Is the right to bodily autonomy absolute? What duties do private parties owe to each other? What role should the government play in all this? To turn these cases over to AI would mean turning over a core aspect of what it means to be self-governing.

Other cases might have less need for human judges. Take contract law, whose main purpose is to create stable rules so that people can make deals with confidence, not to make value judgments. Here, it matters less what the rules are than that they are clear and workable. In one famous case, a buyer and seller fought over whether “chicken” in their contract meant only young birds. Does it really matter whether a human judge decides which birds count as chicken under a contract?

A precedent for drawing such lines exists. Magistrate judges can try misdemeanors but not felonies. In England, low-value money claims are routed through an online process. In America, low-value money claims drive a lot of litigation — 75% of state civil judgments were under $5,200 in 2013 — which may be another promising category for an alternative process, though we’d need to address the risk that poorer parties are systematically routed into what might seem like second-tier justice.

In the abstract, many lawyers and judges agree human judgment matters. The problem is that everyone thinks their preferred area of law is where human judges are needed most.

2. The empirical approach: identify case categories with high consent rates or low appeal and reversal on appeal rates.

At first, the main goal of the AI track should be simply to decide enough cases to see how it performs — effectively A/B testing AI and human judges.

At this stage, exactly how cases reach the AI track would matter less. Courts might assign them at random or have a human judge refer them. And to reduce the risk for both parties, courts could guarantee a right to appeal to a human judge, or only send cases where both parties agree to the AI track.

Over time, the AI track would show which kinds of cases are well suited to AI adjudication, perhaps because parties prefer it or because they’re satisfied with its decisions.

Patterns of consent can reveal where parties accept AI judges. If most parties in small claims cases choose the AI track, that tells courts the parties themselves are often willing to trade a human judge for a quicker decision. Lawyers and judges might view each case as unique and entitled to an individual judge’s attention, but if the people bringing those cases think otherwise, that is valuable information for designing a better justice system. Based on that evidence, a legislature might empower judges to require, rather than simply recommend, that similar cases use the AI track.

Similarly, how often losing parties appeal, and how often a human judge overturns the AI’s decision on appeal, can be good measures of quality. In general, parties only appeal decisions when they are dissatisfied with the outcome or the process, so the less often losing parties appeal, the more confident we can be that AI judges are performing well. As suggested above, appeal rates aren’t a perfect signal, since appeals will take time and money that some parties won’t have. But reviewing judges do not simply rubber-stamp cases on appeal, so low reversal rates would be excellent evidence of the quality of AI decisions.

As these measures build public and judicial confidence in AI judges, courts might give greater weight to AI decisions. If reversals drop below a certain threshold, a court might treat AI decisions the way it treats a magistrate judge’s report and recommendation today. If reversals become rarer still, the option to appeal to a human judge might be removed altogether. That would be a significant step, but it may be necessary to reduce caseloads. Otherwise, losing parties might appeal by reflex and cancel out the time saved.

Courts could then use the data from these cases to train classifiers that predict which new cases will need human judges. Cases where parties are unlikely to consent to the AI track, or are likely to appeal its decision, could be automatically sent to a human judge.

If appeal and reversal rates get very low, they stop being a useful check on the AI’s work. Courts could audit a random sample of cases, so they can compare the kinds of cases each track handles and how it decides them. This would help courts catch problems, such as human judges approving AI decisions without properly scrutinizing them, and monitor for undesirable drift in the common law.

How can courts mitigate the risks?

Of course, no approach is perfect. Even a combination of the two approaches will need to address concerns that AI judges are biased and that moving cases away from human judges will slow the common law’s development or push it in a more mechanical and formalist direction. Courts might not be able to fully resolve either concern, but they can mitigate them.

For example, to reduce the risk of bias, parties could agree in advance which AI model will decide their case. Courts could offer a council of AI models (perhaps instructed to mirror differing judicial philosophies) or the government could develop a custom model for adjudication. Importantly, the risk would be limited to the cases decided by AI judges. Human judges would still dedicate their full attention to the cases where we decide their judgment is needed most.

Whether AI adjudication would distort the common law is an important but untested question. The common law has always changed in response to new technologies and social shifts, and we don’t have great empirical methods for assessing its evolution over time.

And the AI track could be designed with this risk in mind. Because it would absorb much of the growth in new cases, human judges would still decide about as many cases as they do today, leaving plenty of room for judges to keep developing the common law. An AI judge could also be instructed to pass along specific legal questions it thinks would benefit from human judgment, much as federal courts can send unsettled questions of state law to a state’s highest court. That way, courts don’t lose valuable test cases for doctrinal development.

How does an AI track compare with the alternatives?

One obvious response to rising caseloads is to hire more judges. Yet despite growth in both the population and the number of cases, Congress hasn’t passed a major bill creating new federal judgeships since 1990. New judges are expensive, adding them is politically contested, and hiring is unlikely to scale as AI drives caseloads higher.

A second option is to make it harder to file lawsuits and motions — perhaps by tightening the rules for what a filing must contain before a court will consider it. But this wouldn’t hold up for long either, since increasingly capable AI could meet stricter requirements too. And more fundamentally, it’s a blunt tool that would likely shut out people who can’t afford lawyers, undermining the very access to justice that AI could provide.

Third, courts could rely on reputation to discourage weak cases. Many litigators, for example, are repeat players with a strong incentive to protect their reputation and credibility with a judge. Unfortunately, this proposal for reducing caseloads would be least effective where the problem is greatest: self-represented litigants, who are the most likely to file AI-generated briefs and the least likely to appear before the same judge twice.

Finally, the most popular proposal I’ve heard is to fight AI with AI: give judges AI tools so they can work through more cases. This is already happening in Brazil and there’s definitely merit to the idea.

It seems near-impossible for AI tools to improve judges’ productivity enough to handle that many cases without them delegating some degree of judgment to AI systems.

AI tools can help judges be more productive, and judges should experiment with AI so they can build their intuitions about how transformative the technology will be. But relying on AI tools alone risks under-resourced courts relying on them too heavily. Recall that Brazil saw 40 million cases filed last year. It seems near-impossible for AI tools to improve judges’ productivity enough to handle that many cases without them delegating some degree of judgment to AI systems. Indeed, we’ve already seen signs of this: public databases of AI errors in legal documents include judges’ rulings that contain hallucinated citations to cases that don’t exist.

Ultimately, the real choice for legal systems is not between AI judges and human judges with unlimited time. It’s a choice among a set of imperfect options: we can do nothing, add more judges, make it harder to go to court, or use AI in the courts. As caseloads rise, judges will have a difficult time giving each case the attention it deserves. An AI track has an advantage the alternatives lack: it would let courts gather evidence and determine where human decision-makers are truly essential to the judicial process.

How could courts get started?

The federal judiciary’s interim guidance warns judges against delegating “decision-making or case adjudication” to AI, while state courts in California, Illinois, Arizona, and Delaware have similar policies. Yet courts can revise these policies, and some could begin piloting an AI track without any legislative involvement.

Arizona is a prime candidate. Its state law has a well-established system for funneling cases into arbitration. Today, state law requires trial courts to set a dollar threshold, no higher than $65,000, below which civil claims go to an arbitrator appointed by the court rather than to a judge. Maricopa County, for example, sends most claims under $50,000 to an initial round of non-binding arbitration.

A pilot program could work much like the AAA’s process today: an AI agent reviews each side’s claims and evidence and drafts a decision, which a human arbitrator confirms or edits. Because a human still signs off, this could fit within existing law. To implement this, the Arizona Supreme Court would need to (1) add the AI agents to its list of approved AI products, (2) issue an administrative order launching the pilot in one county, and (3) bring in an outside evaluator to measure how often parties accept the AI track and how often they appeal.

An AI track has an advantage the alternatives lack: it would let courts gather evidence and determine where human decision-makers are truly essential to the judicial process.

Arizona has done this before. In 2017, the state’s chief justice signed an administrative order launching a pilot in Pima County, which includes Tucson. The order lowered the county’s arbitration threshold to claims under $1,000, which in effect made arbitration optional for most claims: plaintiffs with claims up to $50,000 could now choose between arbitration and a trial. Since 2017, the Arizona Supreme Court has upheld the program’s legality in the face of litigation and extended the pilot three times, eventually making it permanent in 2025. This required no legislative approval, and the National Center for State Courts evaluated the results of the pilot.

Federal courts could run a similar pilot, but I expect it would be an uphill battle. They face less pressure to change because they handle far fewer cases: roughly 300,000 civil cases last year, compared to roughly 65 million cases of all kinds in state courts in 2022. There’s also more process involved with rule changes, which slows them down. A federal rule change needs approval from the Judicial Conference, the Supreme Court, and Congress.

Once a pilot is running and courts have evidence of its success, a legislature could expand the AI track to a wider range of cases. The way federal courts use magistrate judges offers a useful model.

Today, whenever a magistrate judge proposes a ruling, the parties get 14 days to file specific objections. The district judge then decides only the parts under dispute. If no one objects, the magistrate judge’s ruling stands. Courts can send cases to magistrate judges without the parties’ consent, though these decisions are only final if the parties do consent. The Supreme Court has approved this kind of arrangement, concluding that someone who isn’t a life-appointed federal judge may propose findings, as long as a judge reviews them independently.

Unlike the pilot, expanding the AI track this way would require new legislation because no human would be the official decision-maker. Arizona’s arbitration law currently requires courts to appoint “qualified persons” and requires arbitrators to sign their awards. Federal alternative dispute resolution statutes similarly reference a “person” serving as a neutral decision-maker and “individuals” who take an oath to serve as arbitrators. A powerful way to convince legislators to make this change would be a pilot that already works.

The cost of waiting

Nothing described here is without risk. AI judges may be biased in ways that are more difficult to detect than I’ve suggested. Parties may opt for the faster and sometimes easier AI track without fully understanding what they’re giving up. That’s why courts should start with a careful pilot to find out whether these risks are too severe to overcome.

But doing nothing is no longer an option either. Without a clear and deliberate plan, the possibility that judges will quietly lean on AI tools with no way to assess the results increases alongside the caseloads they are expected to consider. At worst, justice stalls, much as it appears to have done in India and Brazil.

Courts don’t need to wait for legislators to get started. A state supreme court could act this year: approve the tools, select a pilot county, and bring in an evaluator to measure the outcome. The systems that guarantee the law applies equally to everyone are under immense pressure. To ensure courts can continue to resolve difficult disputes, protect citizens’ rights, and uphold the rule of law, state supreme courts should seize the opportunity for experimentation and reform created by AI.