Key Takeaways
- 3 AI justices reached rulings matching their human counterparts
- The real focus was designing input materials, not the ruling itself
- Use AI first to anticipate questions, not to make decisions
- Write down the prep documents from your most recent deliberation meeting
In a mock trial at the University of North Carolina School of Law, 3 AI justices issued rulings matching those of human justices, but Associate Dean for AI Nicole Downing concluded that AI cannot replace judges. According to a Bloomberg report on October 29, 2026 (local time), students and faculty at the law school conducted the mock trial during an AI-themed conference held on campus on October 22. The experiment comes as legal tech companies continue to develop generative AI tools for judicial analysis and case outcome prediction.
How Was the Experiment Designed?
The subject matter was a case actually heard by the U.S. Supreme Court in 2025. The issue was whether, under a system requiring a person in removal proceedings to voluntarily depart within 60 days, the deadline would extend to the following Monday if the 60th day fell on a Saturday.
The organizers designated 3 of the 9 justice seats as AI justices, with each AI representing a different judicial philosophy: textualism (prioritizing the literal text of the law), pragmatism, and a progressive approach. The AI justices were given the following materials:
- Procedural history and briefs from the actual case
- Relevant key precedents, statutes, rules, and agency guidance
- Opinions written by human justices who held the approach each AI was assigned to represent
What Was the Same, and What Was Different in the Outcomes?
Among the 6 human justices and 3 AI justices, the result was 5-4. The majority held that a voluntary departure deadline falling on a weekend or holiday extends to the next business day, with Justices Clarence Thomas and Samuel Alito dissenting. Among the AI justices, the progressive and pragmatist AIs ruled that the deadline should be extended, while the textualist-leaning AI dissented.
The process, however, differed. The AI justices posed questions to the law students arguing the case, but unlike the human justices, they did not interrupt or talk over the advocates. The law school had conducted a similar experiment last year using an AI jury, and this experiment produced comparable results.
Associate Dean Downing said, "We were not asking the question of whether AI can replace a justice." She explained, "We were interested in whether AI could reason in a manner similar to Supreme Court opinions," concluding that AI is useful only as a tool for attorneys to anticipate judges' questions. She added, "AI reasons in a human-like way, but it does not yet act in a human-like way."
Implications for AX at Korean Companies
What companies should take from this experiment is not whether the ruling was correct, but what was given to the AI. The three AI justices received the same case materials, but each was separately given the perspective it was to represent along with the opinions of human justices who held that view. Organizations testing a similar setup for internal deliberations would do well to start by replicating this design: assigning a distinct set of source documents to each perspective.
What changes when Korean organizations adapt this setup is the input material. In this experiment, the AI justices were given U.S. Supreme Court precedents, statutes, and agency guidance. For a Korean company, that role would be filled by internal regulations, past deliberation records, and domestic statutes and regulations, so the starting point is whether such materials are organized in one place.
The conclusion about use cases also translates directly. Associate Dean Downing's conclusion was not to delegate judgment to AI, but to use it to help attorneys anticipate a judge's questions. For functions like legal affairs, contracts, and investment review—where judgment responsibility remains with people—introducing generative AI should begin with testing it as a preparation tool that surfaces likely questions in advance, rather than having it make decisions outright.
There is one thing worth checking this week. Pick the most recent deliberation or approval meeting your organization held, and list out the documents that were needed to answer the questions raised there. That list becomes the draft of the input materials you would need to hand over to AI.
Source: "We Tried AI Justices... 'Reasoning Is Human-Like, But Behavior Isn't Yet'" [Park Si-jin's Column...