Pakistan's JudgeGPT cut backlogs 6% — but only with targeted training
a study of 1,559 pakistani judges using gpt-4 found that districts using judgegpt resolved 1,848 more cases annually with a $38.50 return per dollar spent, showing targeted ai training is critical for bureaucratic productivity.
a field experiment with 1,559 pakistani judges tested judgegpt, a gpt-4 tool with rag access to 129,000 legal documents, to address a backlog of 2.26 million pending cases. judges receiving targeted training resolved 1,848 more cases per district annually compared to a control group, with appeal rates dropping slightly and ruling quality improving in 59% of comparisons. the success depended heavily on training; untrained judges barely used the tool, while trained users shifted to reliable editing tasks rather than risky broad queries. the findings suggest high-impact government ai deployments require specific implementation strategies, not just access to models.