Case law search — Nepal Kanoon Patrika and Supreme Court precedent

The realistic use, drawn narrowly

There is a legitimate use of AI in case-law research, and it is narrower than the brochures suggest. The model is reasonably good at generating a list of hypotheses — “what are the leading Nepali Supreme Court decisions on the question of whether a guarantor is discharged when the principal contract is materially varied without the guarantor’s consent?” — and a junior researcher then takes that list and verifies, case by case, against Nepal Kanoon Patrika.

The hypothesis is the AI’s contribution. The authority is yours.

For each item the model produces — typically a case name, sometimes a year, sometimes a Nepal Kanoon Patrika (NKP) citation, sometimes a one-sentence statement of the holding — the verification sequence is:

1. Look it up. Open Nepal Kanoon Patrika, or the Supreme Court of Nepal website (supremecourt.gov.np), or the search index at the Supreme Court library if you have access. Search by parties, by year, by NKP volume and decision number.

2. If it does not exist, discard it silently and move on. Do not “give the model the benefit of the doubt.” The benefit of the doubt is what gets briefs sanctioned. A case that you cannot find in NKP or on the Supreme Court site does not exist for the purposes of citation, regardless of how confidently the model produced it.

3. If it exists, read the actual judgment. Confirm that the holding is what the model said it was. Models routinely cite real cases for propositions the case does not stand for, summarise judgments in the direction of the question that was asked, and flatten dissents and qualifications into apparent ratios. Reading the actual decision is the only way to catch this.

4. Only after steps 2 and 3 does the case enter your working list. Then you cite it as if the AI had never been involved — because, for the purposes of the brief, it had not.

Why the fabrication rate is so high in Nepali case law

Three structural reasons, and recognising them is half the defence.

The public Nepali case-law corpus is small and uneven. Nepal Kanoon Patrika has been published since 2015 BS, and the bound volumes are a serious body of work — but only a fraction of it is online in a form a model could have ingested in pretraining. The Supreme Court has improved digital access in recent years; older volumes remain partly offline, partly scanned, partly paywalled, partly behind a search interface the model never used. The result is that the model has read a thin and patchy slice of Nepali precedent and is confidently filling in the gaps.

Case names follow a pattern the model can fake. “X v. The Government of Nepal,” “Y v. The Office of the Company Registrar,” “Z v. Nepal Rastra Bank” — these are formats the model has seen thousands of times. Generating a plausible-sounding case name in the right shape is a few tokens of work. The model is not lying in any cognitive sense; it is producing the kind of string the question seemed to call for.

The model has no incentive to say “I don’t know.” When the training data is thin, the architecture still rewards producing a confident, well-formatted answer. Saying “no leading Nepali Supreme Court decision on this point is known to me” is not how the model was rewarded during training. So it produces something, and the something is, often, a fabrication wearing the costume of a precedent.

The authoritative sources, named explicitly

There is exactly one set of sources for Nepali case authority, and the AI model is not on the list.

Nepal Kanoon Patrika (NKP) — the official law reports of the Supreme Court of Nepal — is the authoritative source for reported decisions. Bound volumes are in any serious law office and in the Supreme Court library; an increasing share is available digitally through the Court’s own portal and through commercial legal-research subscriptions. When you cite a case, the NKP citation is what makes it a citation.

The Supreme Court of Nepal website (supremecourt.gov.np) publishes daily cause lists, judgments in many recent matters, and a search interface for decisions. It is not exhaustive — older matters and many unreported decisions are not there — but it is authoritative for what it contains.

What is not on the list: AI model summaries, blog posts citing AI summaries, LinkedIn threads, Facebook law groups, third-party “case digests” of unclear provenance, and screenshots in WhatsApp. Any of these may, in a given instance, be correct; none of them is a source. If a proposition matters enough to put in a brief, it matters enough to find in NKP or on the Court’s site.

What has already happened, and what it means

The pattern is now international and unambiguous.

In Mata v. Avianca (United States, 2023) a federal judge sanctioned two lawyers who had submitted a brief containing six judicial decisions invented by ChatGPT. The cases were formatted correctly. They had docket numbers. They did not exist. The lawyers were fined and publicly censured. In the months and years that followed, courts in the United Kingdom, Canada, Australia, Brazil, South Africa, and Israel reported substantively similar incidents. The pattern is the same in every jurisdiction: a hurried filing, a chatbot used as if it were a case-law database, no verification step, a bench that does the verification the lawyer did not.

Nepal has seen at least one publicised instance of an AI-fabricated citation surfacing in proceedings, and the practising bar should assume there have been more that did not make the press. The cost is asymmetric: the time saved by skipping verification is minutes; the time and reputation lost when a fabricated citation is caught is measured in years.

The defensive conclusion is uncomfortable but plain. The duty of competence owed to the Nepal Bar Council, the duty of candour owed to the court, and the duty of care owed to the client all sit with the lawyer. None of them transfers to the tool. If a brief filed under your name cites a case that does not exist, the disciplinary inquiry does not pause to consider that the model is the one that invented it.

A practical workflow that works

The verification step does not have to be slow. For a research note with twelve candidate cases from a model, an hour of NKP and Supreme Court website checks usually leaves three or four real authorities, which is more than enough for most propositions. The remaining eight are discarded without ceremony. The discipline is to do the checks every time, not to do them once and then skip them when the work is urgent. The urgent matters are the ones the disciplinary committee remembers.

Check your understanding

Quick check

—

A model produces a confident, perfectly-formatted citation to a 2073 Supreme Court of Nepal decision, including parties, NKP volume, decision number, and page, together with a one-sentence summary of the holding that is exactly on point for your writ petition. What does this tell you about whether the case exists?

1. AIt almost certainly exists — the citation is too detailed and well-formatted to be fabricated.
2. BIt probably exists — the model is unlikely to invent a citation this specific.
3. CIt tells you nothing about whether the case exists. The fabrication rate for Nepali case-law citations from general-purpose AI models is high, and the fabrications are confident and well-formatted. The case is a hypothesis until you find it in Nepal Kanoon Patrika or on the Supreme Court website; if you cannot find it there, it does not exist for the purposes of citation, regardless of how persuasive the model’s output looks.
4. DIt depends on which AI model — the newer ones do not fabricate.

Check answer→

What comes next

Knowing that the model fabricates is half of the defence. The other half is understanding why it fabricates — what in the architecture and training process produces confident invented citations — and codifying the verification habit into your daily practice so the discipline survives a 10 p.m. filing deadline. That is the next section.