The bench, the bar and the algorithm

Pakistan's justice crisis is not difficult to diagnose. It is difficult to live with. The World Justice Project's 2025 Rule of Law Index ranks Pakistan 130th of 143 countries overall; on its justice factors, Pakistan stands 129th for civil justice and 101st for criminal justice. At the end of 2025, the Law and Justice Commission of Pakistan reported 2.22 million cases pending across the superior and district judiciary, about 1.9 million of them in district courts. More judges and courtrooms matter, but delay is also produced by thousands of smaller frictions: defective filings, missing documents, service of notices, repeated adjournments, record retrieval, cause-list preparation and routine research. This is where artificial intelligence can help, if Pakistan avoids the temptation to create an "AI judge". The National Judicial (Policy Making) Committee has already opened that door. On April 29, 2026, it issued National Guidelines for the Use of AI in Judicial Institutions, approved at its 57th meeting after a consultative process led by the National Judicial Automation Committee involving all High Courts and expert institutions. The guidelines are sensibly human-centred: AI should assist, not replace, judicial decision-making; privacy, transparency, bias and accountability must be addressed; and applications may include case management, legal research and document processing. However, the harder question is implementation. At district courts, where most pending litigation sits, the larger opportunity is agentic AI. Unlike a chatbot that merely answers a question, a constrained AI agent can perform a sequence of authorised tasks. It could scrutinise an e-filing, flag missing documents, classify the matter, propose a hearing slot, prepare notices, track service, identify repeated adjournments, and alert staff when a case has stalled. Court staff would approve consequential actions, but the routine movement of a file would no longer depend on someone manually checking every step. At the High Courts and Supreme Court, where precedent and constitutional interpretation carry wider consequences, the boundary should be different. AI can turn a 2,000-page record into a chronology, retrieve relevant precedents, compare conflicting judgments, translate material, identify citations, and summarise the parties' arguments. Judicial readers and research associates could use a retrieval system grounded only in authenticated Pakistani statutes and reported judgments, with every proposition linked back to its source. But weighing evidence, assessing credibility, granting bail, sentencing, or deciding constitutional rights must remain human judicial functions. The efficiency case is not speculative. Brazil's Supreme Federal Court reported that its VICTOR system could perform a classification task that took staff an average of 44 minutes in under five seconds. A 2026 randomised study of modern legal AI found productivity gains of 50% to 130% on five of six tested legal tasks. Those results do not mean judges will suddenly work twice as fast, but they show how much clerical and research time can potentially be recovered. Brazil has moved well beyond one experiment. Its National Council of Justice identified 178 AI projects in the judiciary in 2024; 58 of 92 participating courts and councils had developed AI projects, while 45.8% reported using generative AI. Singapore offers a different lesson. Its judiciary uses Pair Search for AI-assisted research of Supreme Court judgments, legislation and parliamentary debates, and has introduced generative-AI tools for translation and case summarisation in the Small Claims Tribunals. Yet Singapore's judges have explicitly said that AI is not, for now, deployed to weigh evidence, apply law to facts or exercise judicial discretion. China has taken the broader "smart courts" route. Its Supreme People's Court has directed AI integration across adjudication, enforcement, litigation services and court management, but its own rules state that AI may not substitute for a judge's decision. That distinction should matter to Pakistan too. The risks are not theoretical. In July 2026, India's Supreme Court dealt with lower decisions that had relied on six AI-generated citations that were non-existent or contained fabricated passages. AI can accelerate legal work; it can also accelerate error with impressive confidence. Pakistan now needs institutional architecture behind its April guidelines. The NJPMC should establish a permanent Judicial AI Unit under the National Judicial Automation Committee and begin with controlled pilots in selected district courts, one High Court, and the Supreme Court registry. Success should be measured publicly: filing-to-scrutiny time, service time, adjournment frequency, case age, and staff hours saved. A sovereign judicial data environment is equally important. Sensitive records should not casually pass through public AI services. Pakistan needs a secure national judicial data layer, common metadata standards, and an authenticated digital corpus of statutes, rules, and precedents. High-risk systems should face independent accuracy, bias, cybersecurity and red-team testing. Every AI action should be logged; every generated citation should be verifiable; and litigants should be told when AI materially assists a process affecting them. Judicial academies must train judges, readers and staff not merely to "use ChatGPT", but to challenge AI output, detect hallucinated authorities, protect confidential data and recognise when automation is inappropriate. Pakistan's courts need technology that removes the clerical drag surrounding those constitutionally entrusted to decide. If AI can return hours to judges, days to registries and months to litigants without surrendering judicial responsibility, it would do something far more valuable than imitate a judge.