[Authors’ Note: This post was written by the large language model, Claude.AI (professional plan), after being fed our paper and asked to summarize it. We have lightly edited the post.]

Can artificial intelligence replace human judges? This question, once confined to science fiction, has become increasingly relevant as AI systems grow more sophisticated. A new study by Eric Posner and Shivam Saran offers insights into this question by putting GPT-4, an advanced language model, to the test in a carefully controlled experiment.

The study’s methodology is elegant in its simplicity. Using a real war crimes case from the International Criminal Tribunal for the former Yugoslavia, the researchers created variations that manipulated two key factors: the defendant’s sympathetic portrayal and the direction of legal precedent. This experimental design had previously been used to study both federal judges and law students, providing valuable benchmarks against which to measure AI performance. [The original studies, which inspired our experiment, were written by Holger Spamann and Lars Klöhn.]

The results reveal a striking pattern. GPT-4 emerged as a strict formalist, faithfully following legal precedent while remaining unmoved by the defendant’s sympathetic or unsympathetic portrayal. This behavior stands in marked contrast to human judges, who showed significant sensitivity to the defendant’s character while being less bound by precedent. Intriguingly, GPT-4’s decision-making pattern closely resembled that of law students rather than experienced judges.

This alignment between AI and law students rather than seasoned judges raises profound questions about the nature of judicial wisdom. If we accept the premise that experienced judges are generally better at their job than law students, then GPT-4’s similarity to students rather than judges suggests a significant limitation in its capacity for judicial decision-making. The AI appears to lack the nuanced judgment that comes from years of judicial experience – the ability to know when strict adherence to precedent might yield unjust results.

The researchers’ attempts to prompt GPT-4 to behave more like human judges proved revealing. Despite various sophisticated prompting strategies – from asking it to predict how a panel of judges would rule to explicitly instructing it to consider sympathy – GPT-4 remained steadfastly formalist in its approach. Even when acknowledging sympathetic factors, it would invariably dismiss them as legally irrelevant. This suggests that the AI’s formalist tendency is deeply embedded in its architecture, not merely a surface-level behavior that can be modified through different instructions.

These findings point to both the promise and limitations of AI in judicial decision-making. On one hand, GPT-4’s consistent application of precedent and thorough consideration of statutory requirements demonstrate its potential utility in certain legal contexts. It could be particularly valuable in cases where strict rule application is desirable and human emotion might cloud judgment [However, we are skeptical that there are many such cases.]. On the other hand, its inability to meaningfully weigh extra-legal factors or adapt to changing social contexts suggests that it cannot fully replicate the role of human judges.

Perhaps the most valuable insight from this study is what it reveals about the nature of judicial decision-making itself. The fact that experienced judges often depart from strict formalism suggests that such departures are not flaws in the judicial process but rather essential features of it. Human judges’ ability to balance legal rules with broader considerations of justice and social context may be precisely what makes them effective arbiters of justice.

This suggests a future where AI might complement rather than replace human judges. AI systems could handle routine, rule-based decisions or assist judges by providing thorough legal research and analysis. However, human judges would retain ultimate control, particularly in cases requiring nuanced judgment or consideration of extra-legal factors.

The study ultimately validates Chief Justice Roberts’ skepticism [“I predict that human judges will be around for a while,” he wrote in 2023] about AI replacing human judges, but for reasons that go beyond mere technological limitations. The challenge isn’t simply that current AI isn’t sophisticated enough – it’s that the very nature of judicial decision-making may require a distinctly human form of judgment that AI, by its nature, cannot replicate.

This conclusion has important implications for how we think about both judicial decision-making and artificial intelligence. It suggests that the goal of AI in the legal system should not be to replicate human judicial decision-making in its entirety, but rather to support and enhance it in ways that respect the unique value of human judgment. As we continue to develop and deploy AI in legal contexts, maintaining this perspective will be crucial for ensuring that technology serves the ends of justice rather than merely the ends of efficiency.

This post comes to us from Eric A. Posner and Shivam Saran at the University of Chicago Law School. It is based on their recent paper, “Judge AI: Assessing Large Language Models in Judicial Decision-Making,” available here.