JudgeGPT: What Pakistan's Large-Scale Experiment Teaches Us About AI Agents in High-Stakes Regulated Domains

When AI Passes Judgment – But Gets It Right: The JudgeGPT Experiment from Pakistan

Judges secretly using ChatGPT and citing fabricated rulings – headlines like these have significantly shaken public trust in AI within the justice system. All the more remarkable, then, is a large-scale field experiment from Pakistan, now described in the journal IEEE Spectrum: under real-world conditions, with real judges and real cases, a specially developed AI tool increased the case completion rate by 6.3 percent – without any measurable decline in judgment quality. For companies looking to deploy AI in sensitive business processes, this experiment delivers a number of highly relevant insights.

A System at Its Limits – and a Purpose-Built AI Tool

Pakistan's judiciary is under enormous pressure: 2.26 million unresolved cases, fewer than two judges per 100,000 inhabitants (compared to 22 in the EU and eight in Brazil). Economics professor Sultan Mehmood of the New Economic School in Moscow, together with Elliott Ash of ETH Zurich and other researchers, developed a tool called JudgeGPT in close collaboration with the Pakistani judiciary.

The technical foundation: OpenAI's GPT-4, combined with a structured knowledge base comprising approximately 128,292 Pakistani court rulings and 943 laws. The decisive architectural principle behind it is Retrieval-Augmented Generation (RAG) – an approach in which the language model does not hallucinate from memory, but actively queries a verified database and provides responses with linked source citations.

"It turns out that the way to fix hallucinations is not simply smarter models – but coupling the models to a tool that can search and verify sources." – Elliott Ash, ETH Zurich

This design principle is not a Pakistan-specific feature. It is the universal core of functional AI systems in regulated environments.

What RAG Does to Hallucinations – and Why This Matters for Companies

Many companies fail in their use of AI not because of the technology itself, but because of the wrong deployment pattern: a generic Large Language Model (LLM) is turned loose on company-specific questions – and produces answers that sound plausible but are factually incorrect. The Pakistan experiment demonstrates vividly that even in the highly specialised field of law, commercial chatbots "frequently hallucinated case law" before a domain-specific RAG system was implemented.

Dr. Maik Bunzel, founder and CEO of mabucon.eu, regularly emphasises this point in his consulting practice: AI agents intended to work reliably within business processes require a clean, structured knowledge base – and an architecture that compels the model to draw from verified sources rather than generating freely. This applies equally to legal documents, technical manuals, compliance requirements, and product databases.

The Underestimated Variable: Training and Adoption

Particularly revealing is the finding on intensity of use. The researchers divided the 1,559 participating judges into three groups:

• Group 1: Specific JudgeGPT training (6 sessions of 90 minutes each, topics: LLM functionality, limitations, hallucination risks, verification obligations)
• Group 2: General training on AI in legal practice
• Group 3: No training

The results are unambiguous: judges with specific training logged in an average of 56 times and submitted 212 prompts – compared to 10 logins and 25 prompts for the generally trained group. Judges without training used the tool for about one month before adoption collapsed. The 6.3-percent improvement therefore did not arise from mere tool availability, but from a combination of the right instrument and targeted competence development.

For organizations, this means: AI adoption is not a deployment problem – it is a change management problem. Anyone who rolls out an AI system and expects the workforce to use it instinctively and effectively will be disappointed.

Quality assurance through AI – methodologically interesting, but open to debate

How do you measure whether rulings improve or deteriorate after AI is introduced? The researchers turned to a remarkable approach: GPT-4o-mini evaluated pairs of rulings from the same judge – before and after training. The model preferred post-training rulings in 59 percent of cases. Two experienced Pakistani lawyers agreed with the model in 70.6 percent of cases and with each other in 73 percent – a remarkably high level of agreement for an LLM-based evaluation method.

In addition, appeal rates declined slightly, which can be interpreted as an indirect quality indicator: faster rulings apparently did not lead to less accurate rulings.

The ROI aspect is equally noteworthy: for every dollar spent on operating the tool, the researchers calculated approximately $38.50 in saved judicial costs. This is not a theoretical figure, but the outcome of nine months of active operation.

What this experiment means for enterprise automation

Pakistan is not an isolated case – it is proof of concept. Brazil and India are already rolling out similar systems, and in the private sector the use of AI agents in complex, document-intensive processes is accelerating rapidly: contract management, regulatory compliance, customer communication, internal knowledge queries.

The JudgeGPT experiment distils the most important success factors into a clear formula:

• Domain-specific knowledge base instead of a generic LLM: The narrower the context, the higher the reliability.
• RAG instead of pure generation: Source grounding is not optional – it is a prerequisite in regulated domains.
• Structured training instead of mere provisioning: Adoption does not follow automatically from accessibility.
• Iterative deployment: The system improves over time – both the data foundation and the model were refined during live operation.
• Quality measurement from the outset: If you don't measure, you can't manage.

Dr. Maik Bunzel, founder and CEO of mabucon.eu, sees field studies like this as an important foundation of legitimacy for responsible AI deployment: "What Pakistan demonstrates applies to every area of business where errors have consequences – which is essentially every area. The difference between an AI tool that increases productivity and one that destroys trust lies not in the model itself. It lies in the system architecture, the data foundation, and the training program."

Outlook: AI as a Reliable Process Partner – But Only with the Right Foundation

The JudgeGPT experiment is not an argument for blindly integrating AI into critical processes. It is an argument for doing it right. The judiciary is one of the most conservative and formality-driven institutions in the world – and even there, a well-conceived AI architecture can deliver significant efficiency gains without sacrificing quality.

For companies looking to integrate AI agents into their workflows, the message is clear: the technological maturity for reliable, domain-specific AI applications exists today. What is often lacking is not a better model, but rather clearer system design, a well-maintained knowledge base, and a thoughtfully developed onboarding concept for users. Those who bring these three building blocks together can – as Pakistan demonstrates – achieve measurable results even in highly regulated environments.