Chapter 03 · Section I · 16 min read
Statute lookup — Nepal Acts, Rules, and gazetted notices
How to use AI to find the right provision in the right Act fast — and the thirty-second source check that catches the section numbers and quoted wording the model quietly invents.
The fastest legitimate use of AI in Nepali legal practice is also the most dangerous if you stop one step early. You have a client question — “can the borrower offset this amount against the loan?”, “does the company need a board resolution for this guarantee?”, “what is the limitation period for filing this claim?” — and you want a candidate provision in the relevant Act in under a minute, not after twenty minutes of paging through a PDF. The model gives you exactly that: an Act, a section, a sentence-long summary of what it says. The trap is that the Act may be right, the section number may be wrong, and the sentence-long summary may be a paraphrase of a provision that exists only in the model’s head. The verification step is not optional. It is the work.
The workflow, honestly described
Used properly, statute lookup with AI has four steps, and the last two are non-negotiable.
1. Ask the question with enough context. A loose prompt — “what does Nepali law say about specific performance?” — produces a loose answer. A tight prompt — “under the National Civil Code 2074, what are the grounds on which a court may decline specific performance of a contract for sale of immovable property?” — narrows the model to a smaller patch of law where its training data is denser and its fabrications easier to spot.
2. Read the model’s answer as a candidate, not a citation. Treat the Act name, the section number, and the paraphrased rule as three independent hypotheses, each of which can be wrong. The model’s confidence tone is the same whether all three are right or all three are invented.
3. Open the actual text. Go to lawcommission.gov.np — the Nepal Law Commission’s Acts portal — and open the Act itself. For the post-2074 codes (Civil Code, Criminal Code, Civil Procedure Code, Criminal Procedure Code), the consolidated bilingual PDFs are there. For older statutes (Companies Act 2063, Income Tax Act 2058, Banks and Financial Institutions Act, Labour Act 2074, Contract Act 2056 where still relevant, Foreign Exchange Regulation Act 2019), the portal has the consolidated text plus the amendment history. Read the section the model cited. If the section number is wrong, search the Act’s table of contents for the actual provision.
4. Read the section yourself before you quote, advise, or rely. A paraphrase from a model is not a citation. A paraphrase from you, having read the section, is.
Why models are particularly bad at Nepali statutes
The fabrication rate is not the same across jurisdictions, and Nepal sits in the worst band. Three reasons.
The training corpus is foreign. A large language model has read perhaps several terabytes of English-language legal material — US federal statutes, Indian Acts and the All India Reporter, English and Welsh case law, EU regulations. It has read a much smaller volume of Nepali legal text, much of it scanned PDFs the OCR mangled. When you ask it about the Civil Code 2074, it does not have the Civil Code memorised the way it has the Indian Contract Act 1872 memorised. It has fragments, summaries, blog posts, and the silhouette of what a Nepali code “ought to” look like.
The model extrapolates from analogues. When the model is asked about a Nepali provision it does not actually know, it does not stop. It produces what a comparable provision in Indian or English law would look like, dresses it in a Nepali Act name, and assigns it a plausible-sounding section number. The output reads correctly because it is correct law — for a different country. The Civil Code 2074 borrows from many sources, so the analogue is often close. But “close” is exactly the failure that puts the wrong number on the right idea, or the right principle on a section that does not exist.
Section numbers and amendments shift. Nepali statutes are amended often, sometimes substantively. A model trained on a 2022 snapshot of the Companies Act may produce section numbers that were correct then but have since moved, or may quote language the latest amendment deleted. Even when the section exists, the wording in the model’s head may be the pre-amendment text.
Concrete examples of the failure mode
A junior associate, drafting a memo on directors’ duties, asks a chatbot for the relevant provision of the Companies Act 2063. The model produces a confident answer: “section 99 of the Companies Act 2063 imposes a fiduciary duty on directors to act in the best interests of the company, including a duty to disclose personal interests in any contract.” The section number sounds right. The paraphrase sounds right. The actual section 99 of the Companies Act 2063 — when you open the PDF — is about something else, and the duties the model described are split across sections 99, 175, and 176 with materially different wording. A thirty-second source check would have caught it. The associate, in a hurry at 11 p.m., did not. The memo went up to the partner with a citation that, had it been filed, would have embarrassed the firm.
A second example, more dangerous. A model produces a confident reference to “section 47 of the National Civil Code 2074” in support of a claim that a particular contract is voidable on grounds of mistake. The phrasing of the rule is correct contract-law doctrine. The section number is wrong — the mistake provisions sit elsewhere in the Code. If the lawyer pastes “section 47” into a writ petition, the bench will check, and what looked like a research shortcut becomes a competence problem in front of a judge.
A third pattern, particularly with banking and FX work. A model summarising the Foreign Exchange Regulation Act 2019 or an NRB directive produces a clean rule with no exceptions. The actual provision has three exceptions and a proviso, and your client’s transaction falls squarely inside one of them. The model’s omission is not malicious — the exception was in a sub-clause two pages later in the original, and the model’s summary smoothed over it. The cost is borne by the client.
Amendments and gazette notices
A specific habit, because it catches a specific class of error.
The Act on screen — even on lawcommission.gov.np — has a date. Read that date. If the consolidated version is from 2078 and you are advising in 2082, there have been four budget seasons of possible amendments. The Income Tax Act 2058 in particular is amended every year by the Finance Act. The Companies Act has been amended several times since 2063. Any AI summary of these statutes is at best a snapshot, and possibly an old one.
The discipline is to check the Nepal Gazette (rajpatra.dop.gov.np) for amendments to the specific Act since the consolidated date. Statutory amendments are gazetted; rules and directives are gazetted; many notifications that change the operative rule are gazetted and never make it into a consolidated PDF for months. For tax and banking work especially, the gazette is the source of last resort, and AI has read almost none of it.
A second habit: when the rule that matters to your matter is a rule or regulation (नियमावली) under an Act, find the rule separately. Models conflate Acts and their rules constantly. The Companies Act 2063 says one thing; the Companies Regulation says the operational detail. Both are on the lawcommission portal. Both must be read.
Check your understanding
Quick check
—A junior associate asks a chatbot about the legal effect of a particular contract clause. The model replies, confidently, that ‘section 312 of the National Civil Code 2074 provides that…’ followed by a one-sentence paraphrase. What is the correct next step?
Quick check
—Why are general-purpose AI models particularly unreliable when asked about specific Nepali statutory provisions, compared with their performance on US or Indian statutes?
What comes next
Statutes are one half of legal research; case law is the other, and it is where the fabrication problem turns from embarrassing to career-ending. The next section is about using AI to surface candidate Supreme Court authorities — Nepal Kanoon Patrika citations, leading cases on a point — and the verification habit that catches the fabricated ones before they reach a bench.