Pakistan trial courts
Pakistan randomises JudgeGPT across 1,559 trial judges and districts resolve more cases
Researchers and Pakistan's judiciary gave an AI assistant, JudgeGPT, to trial judges in a randomised experiment across 118 courts. Districts with more trained judges resolved more cases, about 6.3 percent above the mean at median exposure, with appeals not rising.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
1,848
Additional cases resolved per district per year at median-district exposure to trained judges1
roughly 616
Additional cases resolved per district-year at the 25th percentile of exposure1
1,559
Judges randomised in the experiment1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Pakistan has very few judges for its population, and millions of cases wait for a decision. The Decoder reports 2.26 million pending cases at the end of 2024, most in trial courts. Judges work with little technology and no support staff.
What they built
JudgeGPT is a chatbot built on OpenAI's GPT-4 family of models. It was tailored to Pakistan with a custom instruction and a library of court rulings and laws that it searches before answering, so replies cite real documents. This approach is called retrieval-augmented generation.
How they did it
Researchers from ETH Zurich, Imperial College London and the New Economic School worked with the judiciary. They randomly assigned 1,559 trial judges in 118 courts to one of three groups: AI access with targeted training, AI access with generic training on technology and law, or generic training only. Training for the first group was staged by cohort, starting in March 2024. Chat logs showed how judges used the tool.
Results
Judges with AI and targeted training adopted the tool more, used it more and kept using it. Districts with more exposure to these judges resolved more cases: about 1,848 more per district per year at median exposure, a 6.3 percent increase over the mean. Appeals per resolved case did not rise and, if anything, fell slightly. Training shifted use toward editing and drafting support and away from broad legal research questions.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Additional cases resolved per district per year at median-district exposure to trained judges | 1,8481 | |
| Increase in cases resolved over the mean at median-district exposure | 6.3 percent1 | |
| Additional cases resolved per district-year at the 25th percentile of exposure | roughly 6161 | |
| Judges randomised in the experiment | 1,5591 | |
| Courts covered | 1181 |
Lessons for your project
Pair access with training. In the study, judges who got targeted training used the tool more and kept using it; access alone mattered much less.
Steer people toward tasks where the tool is reliable, such as improving text, and away from open-ended questions that are costly to verify.
Measure output with administrative records, not only user surveys, and track quality (appeal rates) alongside volume.
Randomise rollout where you can. It gave this evidence a credibility that most vendor case studies lack.
Caveats
- The figures are research results: the authors' estimates of how district output changes with the share of trained judges, not a count of cases the tool itself resolved.
- The paper is a working paper dated August 9, 2026 and may be revised. The Decoder article (July 21, 2026) summarises it.
- The paper gives a back-of-the-envelope cost benchmark of about $41.90 of salary-equivalent capacity per dollar spent, and at least $10 under conservative assumptions. The Decoder reports $38.50. Because it is a modelled estimate, it is not shown as a result here.
- The paper names a GPT-4 family model; no exact version is given. The first training cohort began in March 2024, used here as the start date.
- Output data used in the analysis end at year-end 2024, so only the earliest cohort contributes.
Sources
- Courts of Tomorrow: Evidence from a Nationwide Rollout of Generative AI
Mehmood, Goessmann, Ash (working paper) · 2026-08-09 · Research paper · elliottash.com
- An AI system helped Pakistani judges clear massive backlogs at $38.50 return per dollar invested
The Decoder · 2026-07-21 · News · the-decoder.com
Summarized in our own words from the sources above. Spotted an error? See corrections.