The verification workflow and your professional liability

The rule, stated once and never softened

There is one default rule that has to survive the busiest week of your year, the most exhausted partner in the firm, and the most confident-sounding chatbot output you will ever see.

Every AI-produced legal authority — every statute reference, every case citation, every regulation, every gazette notice — is unverified until you have opened the original source and read it. No exceptions. Not for “obvious” authorities. Not for “famous” cases. Not for statutes you think you know by heart. Not when the model has cited the case correctly nine times in a row this morning. Not when you are late, tired, or behind on filings. The rule is a default precisely because the moments when you most want to skip it are the moments it most matters.

The verification has to be against the original source, not a secondary one. The model’s own restatement of the case does not count. A blog post quoting the case does not count. A previous file note in your firm that cites the case does not count, unless someone in your firm verified it then and you trust that verification today. The standard is: a human in your office has read the actual judgment, the actual statute, the actual notice, recently enough to know it still says what you are about to say it says.

The tiered approach to actually doing this

In Nepali practice in 2026, the verification path differs slightly by source type, and it pays to know the routes by heart.

For Nepali statutes. Open the Nepal Law Commission portal or the official gazette PDF. The Acts are there in Nepali, and the section numbers are stable. Do not trust the model’s English summary of a Nepali statute — translation drift between the model and the original is a frequent source of error, especially on procedural sections where a single word changes a deadline. Read the Nepali section yourself.

For Nepali case law. The Nepal Kanoon Patrika, the Supreme Court website, and the High Court repositories where they exist. Coverage is uneven and search is poor; this is part of the reason models hallucinate so freely here. If you cannot find the case in the official source within fifteen minutes, the working assumption is that the case does not exist as cited. The next step is not “trust the model” — it is “find the actual authority for the proposition, or rewrite the argument so it does not depend on a missing case.”

For Nepali regulations and gazette notices. The official Nepal Gazette and the issuing agency’s website. Regulations change quietly and frequently; a model trained eighteen months ago does not know that the threshold was revised last quarter.

For foreign authorities. The original law report, the official court website, or a paid database (Westlaw, LexisNexis, ManupatraIndia where relevant). A model’s summary of an Indian Supreme Court case, an English High Court decision, or a US federal ruling is exactly as unreliable as its summary of a Nepali case — sometimes more so, because the volume of plausible-sounding material in the training data makes the hallucinations smoother.

Why “obvious” is the most dangerous word in this whole topic

The most common professional failure with AI in law in 2026 is not a junior pasting in something exotic. It is a senior accepting something “obvious” without checking.

Models hallucinate most confidently exactly where the data is sparse. For widely-documented topics — the US Constitution, the major Indian constitutional rulings, English contract law — the model has read enough that the citations are usually right and the summaries roughly accurate. For Nepali statute and case law, the model has read very little; the confident output is a mask over guesswork. “Of course there is a Section X of the Act dealing with this” is precisely the sentence at which a model will helpfully invent a Section X that does not exist, with a number that fits, and a wording that sounds right.

The same pattern applies to procedure. “The limitation period for this kind of writ is thirty-five days” — said with confidence by a model — is the moment to open the actual rule, not the moment to relax. The most dangerous AI outputs are not the ones that are wrong in obviously strange ways. They are the ones that are wrong in the way a slightly junior, slightly overconfident, slightly under-prepared advocate would be wrong — because that is exactly the distribution the model is mimicking.

The concrete habits that protect you

Five habits, internalised across the firm, will absorb the great majority of the risk.

Never paste a model-quoted clause directly into a pleading. If the model claims “Section 17(2) of the Contract Act provides that…”, open the Act, find the section, copy the actual wording, and use that. The model’s restatement is for your thinking, not for the document the court will read.

Never quote a model’s case summary directly to a client. If you are advising a client that “the Supreme Court held in X v. Y that…”, you have read X v. Y. Not the model’s paragraph about X v. Y. The actual judgment, in the actual report. The client is making a decision on the strength of your advice; the model is not making that decision, and cannot be sued for getting it wrong.

Never accept “the model says the deadline is X” without checking the source. Limitation periods, filing deadlines, service requirements, gazette notification periods — these are the highest-consequence procedural facts in your practice. A wrong deadline ends a case. The check takes two minutes; the failure ends your insurance year.

Document the verification, however briefly. A one-line file note — “Section 17(2) wording verified against gazette PDF dated [date]; case X v. Y verified against NKP volume Y page Z” — converts your verification from a memory into a record. In an inquiry two years from now, the record is what protects you.

When verification is too expensive, that is a signal not to use AI there. If the only way to use AI on a particular task is to skip verification — because the source is too hard to access, or the volume of citations is too large to check each one — the conclusion is not “verify less.” It is “do not use AI for this task, or restructure the task so verification is feasible.” The shortcut that skips verification is the shortcut that ends in a Bar Council letter.

The liability dimension nobody likes to think about

Professional liability is not a hypothetical for Nepali lawyers in 2026. Bar Council complaints, malpractice claims, and client disputes increasingly turn on documentation — what you did, when you did it, what you checked, what you did not check. AI use sits inside that documentation layer whether you put it there deliberately or not.

If a client complains that your pleading cited a non-existent case, the question that decides the matter is not “did you use AI” — by 2026 the assumption is that you did, somewhere in the workflow. The question is what your verification workflow was, and whether you followed it. A documented workflow, followed in the file in front of the Council, is a strong defence. The absence of any workflow — “we just trusted the output” — is, in practice, an admission of incompetence. The difference between the two is a Sunday-evening habit of writing the verification line into the file note, every time, until it is automatic.

The same applies in malpractice. Insurers in 2026 increasingly ask about AI usage at renewal, and increasingly differentiate premiums based on whether the firm has a documented AI verification protocol. The firms with a protocol pay less. The firms without one will, over the next three years, either adopt one or find their cover narrowing in ways they will not like.

And in the worst case — a client says the advice you gave, based on an unverified AI summary, cost them a sum of money — the protection is the same: a record, contemporaneous, showing that you treated AI output as a draft and verified before you advised. The lawyers who already work this way will be unaffected by the change. The ones who do not will find that “I used a chatbot and did not check” is, in 2027, not a defence anywhere.

Check your understanding

Quick check

—

A model returns a confident citation to ‘Section 17(2) of the Contract Act 2056’ with a quoted clause. What is the correct default verification rule?

1. AIf the citation sounds plausible and the model is usually right, you can quote it directly in the pleading.
2. BYou should check the section only if you have time, and skip the check if the deadline is tight.
3. CEvery AI-produced legal authority — statute, case, regulation — is unverified until you have opened the original source and read it. No exceptions for ‘obvious’ or ‘famous’ authorities, and no exceptions for tight deadlines. If verification is too expensive, that is a signal not to use AI on that task, not a signal to skip verification.
4. DVerification is only required for foreign authorities; Nepali statutes can be trusted to the model because they are simpler.

Check answer→

What comes next

Confidentiality decides what goes in. Verification decides what comes out. The final section of this chapter, and of the course, is the harder commercial question: once you are using AI competently and safely, how do you bill for the work, what do you disclose to the client, and what does the small Nepali firm look like five years from now?