Source: Facebook

Pakistan’s overloaded trial courts cleared about 1,848 extra cases a year at median exposure to JudgeGPT, an effect equivalent to a 6.3% increase in an average district’s annual output after judges received a bespoke generative-AI assistant plus practical, tool-specific training. The experiment covered 1,559 judges across 118 courts — roughly half the country’s trial judiciary — and the available data showed no corresponding increase in appeals. The observed productivity difference derived from targeted training rather than mere access: judges who received tool-specific instruction logged in about 56 more times and submitted about 212 more prompts than the control group, whereas judges given JudgeGPT alongside a generic technology-and-law course recorded only about 10 additional logins and 25 prompts.

JudgeGPT was built around Pakistani law by linking GPT-4-family models to 129,235 legal documents, including over 128,000 judicial opinions and 943 statutes; judges primarily used it to explain legal concepts, improve writing, find authorities and support drafting. Targeted training steered users toward bounded tasks such as editing and summarisation and away from broad legal questions where errors (including hallucinated citations) are harder to detect. Quality measures are encouraging but less conclusive than productivity: treated judges produced opinions with more text classified as AI-generated without detected declines in readability, length or legal argumentation, and an automated assessment tended to prefer post-treatment opinions from the targeted-training group. The study cautions that judges were randomised individually but court output was only available at district level and only through 2024, leaving a single post-treatment year; therefore the link between higher district exposure to JudgeGPT-trained judges and higher case disposal cannot yet show whether gains persist, whether higher courts later reverse more decisions, or whether the tool could substitute for future judicial hiring. Researchers estimate about $38.50 in salary-equivalent judicial capacity for every dollar spent on recurring infrastructure, while noting that this calculation excludes development, maintenance and support costs and represents freed capacity rather than direct budgetary savings.