Pakistan's court backlog is not a minor administrative headache. Over 2.2 million cases are pending, 82 percent of them at trial court level. A civil suit can take roughly 20 years to reach a verdict. The country has fewer than two judges per 100,000 people, while the EU has 22 and England and Wales have 30. Hiring more judges is the obvious fix, but the government has repeatedly said it cannot afford to.
So researchers from three universities built a different kind of trial. They created JudgeGPT, an AI assistant sitting on top of a database of 128,292 court rulings and 943 Pakistani laws. When a judge types a question, it finds the ten most relevant passages and generates a cited answer. Think of it as a very well-read legal researcher available at any hour, who never takes a day off.
The experiment covered 1,559 judges, roughly half of all Pakistani trial court judges, split into three groups. One group got the AI tool plus six weeks of hands-on training after court hours. A second got the AI tool but only a general technology seminar. A third got the seminar with no AI access at all.
AI access alone did almost nothing. Judges who only got the generic seminar averaged about 20 logins and fewer than 50 prompts over 40 weeks. The trained group averaged nearly 60 logins and over 200 prompts. The difference was not the technology. It was knowing which tasks the tool handles well and which it does not.
Trained judges used JudgeGPT mainly for editing, summarizing, and looking up procedures: tasks where AI is reliable. They asked fewer broad open-ended legal questions, where AI is more likely to produce confident-sounding but wrong answers. The training, in effect, shaped how people used the tool rather than just handing it over and hoping for the best.
The output gains were real. Districts with higher concentrations of trained judges resolved around 1,848 more cases per year. Even the bottom quarter of districts still cleared around 616 extra cases. Ruling quality held steady or improved slightly: appeal rates did not rise, readability did not fall, and an independent quality review rated trained judges' rulings better in 59 percent of side-by-side comparisons versus 42 percent in the control group. Critically, the study found no increase in gender or religious bias in the language of rulings.
The $38.50 return per dollar figure is calculated by comparing program costs to what it would cost to hire enough additional judges to produce the same output. That is a useful benchmark, not a marketing figure: the researchers also published a conservative estimate of at least $10 per dollar.
The Pakistan Supreme Court has formally endorsed JudgeGPT in a published ruling. That matters because it moves this from a pilot experiment to something closer to institutional policy. The tool ran on GPT-4, which was state-of-the-art at the time of the experiment. Newer AI models produce fewer errors and handle complex reasoning better, which means the gains documented here are likely a floor rather than a ceiling.
This is not a court story. The pattern it reveals applies to any institution where backlogs grow faster than headcount budgets: insurance claims processing, permit applications, customs clearance, public procurement reviews. The lesson is specific. Giving people an AI tool without telling them where it is reliable and where it is not produces almost no benefit. Pairing the tool with targeted training on its limits and best uses is what drives the outcome. That training does not have to be long. In this trial, it was six 90-minute sessions. What matters is that it is task-specific, not generic.