Hallucination and the duty of competence · AI for Lawyers

There is a particular failure mode of generative AI that the legal profession is, by its nature, the worst-positioned to absorb. The model produces text that is confident, well-formatted, structurally correct, and silently wrong. In most professions the silent wrongness gets caught downstream — by a calculator, by a reviewer, by reality the next morning. In law, the silent wrongness frequently goes all the way into a pleading, all the way into a submission, and is caught for the first time by a judge with the bound NKP volume open on the bench. By then it is too late.

The dangerous middle

It would be easier if AI failed loudly. A model that produced obvious gibberish would be a tool a lawyer could not misuse: nobody puts gibberish in a writ petition. The actual failure mode is more dangerous than that. The model produces a paragraph that reads exactly like a competent junior associate’s research note. The case name has the right shape — a surname versus a Ministry, the right decade. The NKP citation has the right format — volume number, year, decision number, page. The proposition is reasonable, sometimes even correct as a general matter of Nepali law. The judges named are real Justices who have sat on the Supreme Court. Every surface signal says: this is real research.

It is not. The case has never existed. The citation points to a page that, in the actual volume, deals with a partition suit from a different district. The proposition is the model’s own paraphrase of something it half-remembers from an Indian decision, dressed in Nepali clothing. None of this is visible in the output. The model is not lying — it is doing what it always does, which is to produce the most plausible-looking continuation of the prompt, with no internal flag that distinguishes “I read this” from “I am guessing.” That distinction does not exist inside the model.

This is the failure mode known, politely, as hallucination. The accurate description is that the model is a fluent guesser that has no idea when it is guessing, and is at its most confident precisely where it is least informed. Nepali legal authority — NKP cases, the bare text of the Acts, NRB unified directives, Insurance Authority circulars, ICAN pronouncements with legal effect, decisions of the various tribunals — is exactly such a place. Most of this material has never been on the open internet in a form the model could have read. The model knows it does not know, and then it does not tell you.

What has already happened

By 2025 the international record is no longer ambiguous. United States federal judges have sanctioned at least three sets of lawyers for filing briefs containing AI-fabricated authorities, with the most-cited instance being the New York personal-injury matter in which the court found six non-existent decisions in a single submission and imposed monetary penalties. Courts in the United Kingdom, Canada, Australia, Colombia, and at least one bench of the Delhi High Court in India have issued comparable warnings or formal admonishments. Bar associations in multiple jurisdictions have followed with formal guidance — sometimes mandatory disclosure of AI use, sometimes outright prohibition on AI-drafted citations without independent verification.

Nepal is not outside this picture. By 2025 at least one documented instance has surfaced of a writ petition before a Nepali court containing a citation to a Supreme Court decision that did not exist, traced to an associate’s use of an unverified chatbot output. The lawyer’s defence — that the tool produced the citation — did not move the court, and the matter generated discussion at the Bar Council about competence and the use of AI. The specific facts will continue to evolve, and the case may not be the last; but the principle the bench articulated is already settled in every comparable jurisdiction, and there is no reason to expect Nepal will be the exception.

The principle is this: the lawyer’s signature on a document is a representation that a competent lawyer has read what is in the document and stands behind it. If the document contains a fabricated authority, the representation is false, regardless of which tool produced the paragraph. The tool has no standing before the court, no licence to practise, and no professional responsibility. The lawyer has all three. The consequence falls on the lawyer.

The duty of competence does not transfer

The Nepal Bar Council’s professional standards, like the standards of every other functioning bar in the world, place on every advocate a duty of competence — owed to the client, who is paying for skill they cannot themselves judge; to the court, which is entitled to assume that what is filed represents a professional view honestly held; and to the profession itself, which depends for its credibility on the assumption that any advocate before any bench has met a basic standard of care. This duty is not a piece of paperwork to be discharged at admission. It is the continuing condition of being permitted to practise.

The duty of competence does not transfer to a tool. This is not a controversial proposition; it follows from the structure of the duty. A lawyer cannot delegate the duty to a junior associate (the partner remains responsible for the work that goes out under the firm’s name). A lawyer cannot delegate the duty to a paralegal (the lawyer is responsible for what the paralegal produces). The lawyer certainly cannot delegate it to a piece of statistical software that has no licence, no oath, no insurance, and no ability to be disbarred. “The model said so” is not a defence. It is, in fact, an aggravating admission, because it concedes that the lawyer signed a document containing material they had not personally verified.

What this means in practice

Three practical rules follow directly, and they are not negotiable for a working Nepali advocate.

1. Every citation, every time, opened in the source. Every NKP citation, every High Court decision, every section of the Civil Code 2074 or Criminal Code 2074 or Companies Act 2063 or Labour Act 2074, every NRB directive, every Insurance Authority circular that goes into a document with your signature must be opened in the actual source — the Supreme Court repository, the Nepal Law Commission’s bare-Act page, the regulator’s own website, or the bound volume — and read in its actual context before the document leaves your hands. If you do not have time to verify it, you do not have time to file it.

2. AI output is a draft, not a source. A summary the model produces is a route to the source, not a substitute for it. A clause comparison the model produces is a checklist for your review, not an opinion on the contract. A list of candidate arguments the model produces is brainstorming, not advice. The mental model is: AI sits one rung below a junior associate in the trust hierarchy — useful, fast, and to be verified.

3. The verification step is the professional act. In the workflow described in the previous section, the saved hours come from drafting and structuring. The protected hours — opening the source, reading it in context, forming the view — are where the practice of law actually happens. A firm that uses AI to compress those hours has not become more productive; it has become less competent, and the consequence will arrive on some future morning in front of a bench that has read the cited decision and noticed it is not what the submission claims.

The opportunity here is real. A Nepali advocate who treats AI as a force-multiplier on the draftable parts of the week, and who treats the verification of authority as the place where her professional identity lives, will be both faster and more careful than the lawyer of five years ago. A Nepali advocate who treats AI as a shortcut around verification will, in due course, find out what the duty of competence means by having it enforced against her. The choice is structural, not moral, and it is made early — in the first month of using these tools, not in the moment of filing.

Check your understanding

Quick check

—

An advocate files a written submission containing three citations to Supreme Court decisions, two of which are later found to be entirely fabricated by the AI tool the advocate used. Which statement most accurately reflects the position under the Nepal Bar Council’s duty of competence?

The professional responsibility shifts to the AI vendor, since the advocate relied in good faith on the tool’s output.
The duty of competence remains with the advocate; the lawyer’s signature represents that a competent lawyer has verified what is in the document, and “the model said so” is not a defence.
The advocate is responsible only if the firm did not have a paid subscription to the tool.
The Bar Council has no jurisdiction unless the client formally complains.

What comes next

This chapter has mapped where AI helps in a Nepali legal practice, where it does not, and where the duty of competence keeps the professional load on the lawyer rather than the tool. The next chapter takes the largest single category where AI genuinely compresses time — contracts — and works through review, drafting, and negotiation in detail: the prompts that work, the templates worth keeping, and the verification steps that turn a faster workflow into a safer one.