This month, the Supreme Court of Nepal introduced a new policy permitting the use of artificial intelligence (AI) in the judiciary. One particular provision deserves close attention. The policy explicitly warns that AI systems may cite laws that do not exist. The Chief Registrar acknowledged this risk, and the Joint Secretary involved in drafting the rules echoed the same concern. In other words, the judiciary is allowing AI into the courtroom while simultaneously acknowledging that it may fabricate the very laws the courts are meant to interpret. That cautionary note is perhaps the policy's most candid and important provision.
At face value, the policy is both reasonable and pragmatic. It allows AI to assist with legal research, case management, and routine administrative work while prohibiting it from drafting judgments, writ petitions, or performing tasks that require judicial discretion. AI may assist judges with research, but the responsibility for making decisions and writing judgments remains with human judges. It is a sensible boundary, and anyone familiar with current AI technology would likely draw it in the same place.
The judiciary is right to approach AI cautiously because today's AI models do not actually understand the law; they merely predict text based on patterns in data. When asked a legal question, an AI system may produce a confident answer complete with legal citations. Sometimes those citations are accurate. At other times, they are entirely fabricated. AI may refer to statutory provisions that have never existed or quote judicial decisions that were never issued. This phenomenon, commonly known as "hallucination," remains an inherent limitation of current generative AI models. Lawyers in several countries have already faced disciplinary action for submitting court filings that cited completely fictitious cases generated by AI.
Artificial Intelligence Conference: AI welcomes guests and part...
However, the line between acceptable and unacceptable use of AI is not as clear as it initially appears. No competent judge would simply ask AI to write a judgment and sign it without review. The greater risk is more subtle. A court clerk may ask AI to conduct legal research, and the system may produce a polished summary supported by several precedents, one of which is entirely fictitious but presented so convincingly that no one notices. The clerk incorporates that summary into a briefing note, the judge reads the note, and the final judgment unknowingly incorporates the fabricated citation. The policy rightly focuses on preventing AI from making judicial decisions, but it says little about the information reaching judges before those decisions are made.
This challenge may be even greater in Nepal than in the countries from which the concept has been borrowed. Current AI models have been trained primarily on legal materials from jurisdictions such as the United States and India. By contrast, Nepali statutes, the Muluki Codes, and decisions of Nepali courts are only minimally represented in the datasets on which these systems are trained. The less relevant legal material an AI model has access to, the more likely it is to generate inaccurate or fabricated information. The policy itself recognizes this limitation by proposing the digitization of court records so AI systems can rely on authentic Nepali legal sources. Until such comprehensive databases exist, however, AI applications dealing with Nepali law will inevitably rely on educated guesses presented with unwarranted confidence.
The policy also correctly states that AI systems should be deployed only after they have been evaluated for reliability and bias. Yet the policy simultaneously acknowledges that Nepal currently lacks the technical expertise and institutional capacity to conduct such assessments. Evaluating AI systems is a specialized task requiring both skilled personnel and robust infrastructure—resources that remain limited.
Despite these constraints, pressure to adopt AI will continue to grow because Nepal's courts face an enormous workload. The Supreme Court alone has more than 27,000 pending cases. Last year, it disposed of fewer than one-third of them, partly because the Gen Z protests damaged court buildings and records. Even under normal circumstances, however, the judicial backlog remains overwhelming. Judges reviewing hundreds of case files each week may simply lack the time to verify every AI-generated citation, increasing the likelihood that fabricated references will slip unnoticed into official decisions.
The Supreme Court has taken the important step of permitting AI in judicial administration. The more difficult task—ensuring its safe and responsible use—has only just begun. If AI is to improve judicial efficiency without compromising the integrity of court records, substantial groundwork must come first. Nepali laws, regulations, and judicial precedents must be comprehensively digitized, organized, and verified so AI systems rely on authentic legal sources rather than statistical guesswork.
Every AI-generated output should be treated like a junior clerk's memorandum—potentially useful but never authoritative until every citation and legal proposition has been independently verified. Judges will also require training, not merely on how to use AI tools but, more importantly, on how to recognize their limitations and detect their mistakes. Before any AI-generated claim finds its way into a judicial decision, someone must verify the original statute or precedent rather than relying solely on the machine's summary.
At the very least, Nepal's judiciary has acknowledged the risks associated with AI, which is more than many governments have done. A technology capable of inventing laws has now been admitted into the institution responsible for interpreting and enforcing them. AI will continue to generate convincing but occasionally false information—that limitation is unlikely to disappear anytime soon. The real question is whether judges, lawyers, and court officials will continue to verify every AI-generated claim, even when workloads are overwhelming and deadlines are pressing. The day they stop checking is the day a fictitious law could become part of a very real judicial decision.
The author is a PhD student in Machine Learning and Bioinformatics at the University of Georgia, Athens, USA.
