Before running our own experiments, we surveyed every serious attempt we could find to put AI in a judging role — academic, governmental, and commercial. Here is what exists, what it found, and the gap none of it fills.
Published 31 August 2026 · sources linked throughout · corrections welcome, prominently applied
In 2026, legal scholar Eric Posner and Shivam Saran at the University of Chicago re-ran an experiment originally conducted on 31 sitting U.S. federal judges — a simulated appeal in a war crimes case, where researchers varied how sympathetic the defendant appeared and whether the correct legal outcome followed precedent. GPT-4o sat the same test.
The result: the AI applied precedent correctly and disregarded the legally irrelevant sympathy factor. The professional human judges, by contrast, were influenced by how sympathetic the defendant appeared. On this one measure, in this one study, the machine behaved more like the law demands a judge behave than the judges did.
One study is one study — it tested a narrow task, not a full trial. But it is precisely the effect our research programme exists to test at full-case scale.
In November 2025, the University of North Carolina School of Law staged a mock criminal trial in which three AI systems deliberated together as a jury and returned a not-guilty verdict. It drew national attention and heated debate about bias, accuracy, and whether machines should ever hold that role. It was a single festival event — no methodology published for reuse, no ongoing programme — but it demonstrated two things: the experiment is technically possible, and the public appetite for it is enormous.
Since 2017, China has operated online courts (beginning in Hangzhou) that handle high-volume internet disputes with heavy AI assistance in processing and drafting. The scale is real; the transparency of the reasoning is not — the systems are assistive infrastructure inside a court, not published, checkable judgments.
Estonia was widely reported around 2019 to be piloting automated resolution of small claims — the famous "robot judge" headlines. Its Ministry of Justice later downplayed the reports. We cite it as what it is: a signal of governmental interest in exactly the low-stakes, high-volume arena our research ladder points to, not a deployed AI judge.
Tools like COMPAS have been used for years in U.S. courts to score defendants' reoffending risk at bail and sentencing — and became the field's chief warning: journalists and researchers found racial disparities in error rates, and the reasoning was proprietary and unauditable. A closed-box algorithm inside a courtroom is the opposite of what this research argues for. Every criticism of COMPAS is an argument for the glass box.
Beyond that, the landscape thins quickly: small online tools that will "judge" a dispute between two people for entertainment; legal-tech products that help lawyers prepare and test their cases; a self-published book applying AI to a famous trial without a stated methodology. All signals of interest; none of it research.
If we have missed a project doing this work, tell us — this page carries a date because it is a snapshot, not a claim of omniscience, and corrections will be applied prominently.