Courts of Tomorrow: Evidence from a Nationwide Rollout of Generative AI

We present the first large-scale field experiment evaluating the integration of generative AI into a national justice system. In partnership with Pakistan's judiciary, we built JudgeGPT, a custom generative AI assistant designed for Pakistan's trial courts, and randomized 1,559 judges serving across 118 courts into one of three arms: (i) access to the assistant with targeted training tailored to its use; (ii) access to AI with generic training on technology and law; and (iii) generic training without access to AI. Judges who received AI access together with targeted training on the use of the tool were more likely to adopt it, use it more intensively, and continue using it over time. Their attitudes toward AI also shift: they expect the tool and the targeted training to increase their productivity. Administrative records move in the same direction: districts with greater exposure to treated judges resolve more cases. At median-district exposure, introducing AI with targeted training corresponds to 1,848 additional cases resolved per year, a 6.3 percent increase over the mean. This increase does not appear to come at the expense of reduced quality, as measured by case appeals (which slightly decline) or text-based measures of judicial writing (which slightly improve). Chat-log-based usage patterns suggest that judges use AI mainly to clarify legal concepts and support drafting. Targeted training appears to shift use toward tasks where language models are likely to be more useful, such as text improvement, and away from more open-ended legal queries where responses are more costly to verify. Overall, the results suggest that generative AI can raise public-sector productivity, but that these gains may depend on targeted training that directs use toward tasks for which the tool is better suited.