JudgeGPT помог пакистанским судьям разгрузить завалы дел: +6,3% и ROI $38,5 на доллар

Researchers conducted a field experiment involving 1559 Pakistani judges and found that the AI assistant JudgeGPT increased the share of resolved cases by 6.3%. The effect appeared only among judges who completed in-person training on working with the system, while the return on investment reached up to $38.50 for every dollar invested.

What the experiment showed

The field experiment covered 1559 Pakistani judges who were given access to the AI assistant JudgeGPT for working on court cases. According to the study, among judges who mastered the tool, the share of resolved cases rose by 6.3% — a tangible shift for an overloaded system where case backlogs drag on for years.

A field experiment is a study conducted under real-world conditions rather than in a laboratory, which is why its results reflect an actual causal relationship rather than a simple correlation.

Key figures from the study:

• 1559 judges — participants in the field experiment
• +6.3% — increase in the share of resolved cases among trained judges
• up to $38.50 — return for every dollar invested in the project
• JudgeGPT — the AI assistant that helped judges handle cases
• without in-person training, the gain almost completely disappeared

Why only training made a difference

The 6.3% increase was produced exclusively by judges who received hands-on, in-person training in working with JudgeGPT. Among those who simply got access to the system without in-person training, the effect nearly vanished entirely — the tool alone did not deliver results.

This is the study's central finding: the digital assistant pays off not because of the technology itself, but because people were taught to use it. The gap between "given access" and "taught to work with it" turned out to be decisive — it was precisely what separated the judges who cleared their backlogs from those whose situation didn't change at all.

"The 6.3% increase came only from judges who completed in-person training; without it, the effect virtually disappeared," according to the study's key finding, as reported by

The Decoder.

How much the investment paid off

The project's return, according to the researchers' estimate, reached $38.50 for every dollar invested. Even allowing for the fact that this is the upper bound of the estimate, the cost-to-result ratio for a government project is rare: technology rollouts in the public sector usually pay off slowly and unpredictably.

This kind of economics stems from how cheap the intervention was relative to its effect. Judges' processes weren't overhauled entirely — they were given an AI tool and training, and the result was a measurable rise in case resolution rates within a system plagued by chronic backlogs.

What this means

The JudgeGPT experiment shows that the benefit of AI in the public sector is determined not just by the model, but by user training: without in-person coaching, even a working tool delivers no effect. For courts, healthcare, and other overloaded public services, this is a direct signal — the budget for AI rollout should be planned alongside a budget for training people.

Frequently Asked Questions

What is JudgeGPT?

JudgeGPT is an AI assistant for judges that helps them handle court cases. In a field experiment with 1559 Pakistani judges, it increased the share of resolved cases by 6.3% among those who completed training.

Why did AI help only some of the judges?

The increase appeared only among judges who completed hands-on, in-person training in working with the system. Among judges without training who received only access to JudgeGPT, the effect nearly disappeared entirely.

How much did the investment pay off?

According to the researchers' estimate, the project's return reached $38.50 for every dollar invested — thanks to the low cost of the intervention relative to the increase in the number of resolved cases.