How do judges use large language models? Evidence from Shenzhen

This article reports on the systematic use of a large language model by a court in China to generate judicial opinions—arguably the first instance of this in the world. Based on this case study, we outline the interaction pattern between judges and generative artificial intelligence (AI) in real-world scenarios, namely: 1) judges make initial decisions; 2) the large language model generates reasoning based on the judges’ decisions; and 3) judges revise the reasoning generated by AI to make the final judgment. We contend that this pattern is typical and will remain stable irrespective of advances in AI technologies, given that judicial accountability ultimately rests with judges rather than machines. Drawing on extensive research in behavioral sciences, we propose that this interaction process between judges and AI may amplify errors and biases in judicial decision-making by reinforcing judges’ prior beliefs. An experiment with real judges provides mixed evidence.