GPT-5 Outperforms Judges in Rule Application

University of Chicago law professors Eric Posner and Shivam Saran report in a 2025 paper that GPT-5 applied applicable state-law rules correctly in their replication study of federal-judge decisions. The researchers found GPT-5 (and Google Gemini 3 Pro) reached legally correct outcomes in 100 percent of tested car-accident choice-of-law scenarios, compared with judges' 52 percent compliance. The result raises questions about delegating consequential judicial decisions to rule-bound models.