Silicon Formalism: Rules, Standards, and Judge AI

This paper replicates a judicial experiment originally conducted on 61 U.S. federal judges, this time with GPT-5 as the decision-maker. The experiment involves a choice-of-law dispute in a hypothetical automobile accident case, where three variables are manipulated: (1) whether the applicable doctrine is a rule or a standard, (2) whether the plaintiff or defendant is portrayed more sympathetically, and (3) the location of the accident, which affects the legal outcome under different states' choice-of-law rules. Consistent with our prior work, we find that the LLM adheres to the legally correct outcome significantly more often than human judges. In fact, the LLM makes no errors at all.