Three artificial-intelligence “justices” built on ChatGPT reached the same outcome as the US Supreme Court during a mock argument held on October 22 at the University of North Carolina School of Law. The result matched; the courtroom give-and-take did not, Reuters reported on October 28.

A 2-1 vote mirroring a 5-4 decision

At the event, titled “May It Please the Machine,” third-year students Elijah Baetz and Reda Ansar argued Monsalvo Velázquez v. Bondi, a 2025 immigration case, according to UNC Law. The question before the panel: if a 60-day voluntary-departure deadline fell on a Saturday, did it extend to Monday? That deadline marks when someone permitted to leave the United States on their own must depart to avoid penalties.

Data librarian Rolando Rodriguez developed the three AI justices using a single ChatGPT model, with each configured around a different approach to interpreting the law. The textualist adhered strictly to the literal meaning of the statute; the pragmatist weighed practical, real-life results; and the progressive took the widest view, factoring in changing social values and modern legal principles. To ensure fairness and neutrality, all three worked only from anonymized legal records and remained offline throughout.

The pragmatist and the progressive backed the extension in a 2-1 vote, with the textualist dissenting. In the actual Supreme Court decision, the justices split 5-4, with Justice Neil Gorsuch writing for the majority and Justices Clarence Thomas and Samuel Alito in dissent.

The machines reasoned but never interrupted

The AI justices posed questions and floated hypothetical scenarios, yet they never talked over the advocates. Nicole Downing, assistant dean for AI at UNC Law, said the point of the exercise was to test AI's capacity to reason as a judge does in a written opinion, not to replace judges. Clinical professor Annie Scardulla explained that oral argument is really a “professional conversation” — a natural back-and-forth the AI justices lacked. That distinction matters because matching a case outcome does not show that a system can reproduce the questioning, responsiveness and interpersonal judgment involved in an actual hearing.

Law professors rated AI tutors above their peers

AI looks more promising as a teaching tool for law students than as a judicial decision-maker. In a blinded study, 16 contracts professors from 14 US law schools reviewed 2,918 comparisons and preferred AI-written answers 75.33% of the time. Only 3.53% of AI responses were flagged as harmful, compared with 12.06% of human-written answers. Still, helping students learn the law is one thing; deciding cases is another.

Consumers draw a line at AI decisions

Public skepticism runs deep as well. Legal software company Clio found in its 2026 survey that only 26% of US consumers were willing to let AI make decisions. Where a matter was highly personal, that figure fell to 6%. In addition, 82% of respondents wanted their lawyers to tell them whether AI was being used.

What the trust gap means for legal AI vendors

Grand View Research values the legal AI market at $1.4 billion in 2024 and projects $3.9 billion by 2030, a 17.3% annual growth rate for 2025–2030. North America accounted for 46% of 2024 revenue. Demand centers on research, document review and compliance — not automated judging.

Thomson Reuters found that 77% of clients consider AI-enabled quality improvements very important or essential, but only 5% said most providers deliver them. Meanwhile, 34% of law firm professionals have used unapproved tools.

Accuracy remains uneven. On a multi-state statutory-research benchmark, Westlaw AI scored 58%, Lexis+ AI 64% and the custom system STARA 83% — figures that are not general legal-accuracy ratings.

As Cryptopolitan reported, Google has launched Gemini Enterprise for Legal in preview, emphasizing traceable citations and client-data protections.

UNC's lesson is clear: AI can help lawyers reason, but humans must still judge.