Court submissions and the candor obligation

The duty, restated for 2026

The advocate is an officer of the court. The Bar Council Rules, the Nepal Bar Council Act 2050, and a long line of judicial pronouncements impose, in different words, the same essential obligation: do not knowingly assert as fact what you know to be false; do not cite as authority what you know does not exist; do not stand by silently while the court relies on a misstatement you can correct. The duty is owed to the institution of the court, not to the client, and where the two are in tension, the court wins.

This duty is independent of how the submission was produced. If a junior typed a fabricated case into the brief, the duty is breached. If a senior dictated it, the duty is breached. If the lawyer pasted a question into a chatbot, accepted the response, and filed it, the duty is breached. The court does not weigh the typing instrument; it weighs the signature.

The pattern of documented incidents

By mid-2026, the documented incidents of lawyers sanctioned for filing AI-generated fabricated citations have moved from “novelty” to “weekly news.” The pattern is depressingly consistent. The earliest widely reported cases — Mata v. Avianca in New York in 2023, followed by sanctions matters in Texas, Colorado, the United Kingdom, Canada, Australia, India and Singapore — share the same shape. A lawyer under time pressure asks a model for “relevant cases.” The model produces names, citations, and apparent holdings. The lawyer files them without opening any of the cited decisions. Opposing counsel cannot find the cases because the cases do not exist. The court asks. The lawyer first explains, then apologises, then is sanctioned, fined, and in several jurisdictions referred to the bar for disciplinary proceedings.

Nepal has not been spared. By 2025 there were at least two documented matters where district court advocates had to withdraw filings after opposing counsel demonstrated, in open court, that cited NKP decisions did not exist — and the bar had begun discussing whether to issue formal guidance. The pattern that has emerged is that the cost of one such incident — to the lawyer’s standing, to the firm’s reputation, to the client’s case — dwarfs any productivity gain the use of AI might have produced across a career of careful practice. There is no economic case for letting an unverified citation enter a filing. The math is one-sided.

The pre-filing checklist

Before any document goes from the office to the registry, the bench memo system, or opposing counsel, a short discipline runs. It is not optional and it does not get skipped because the deadline is tight — if the deadline does not allow it, the deadline is wrong and an adjournment is the correct response.

1. Every authority verified against its source. Open the NKP. Open the bare Act. If the citation is a foreign decision, find the reported text. If you cannot find the decision, the citation comes out — there is no third option. The verification can be delegated; the responsibility cannot.

2. Every factual assertion traceable to evidence. For each numbered paragraph in the facts, the file should contain the document, the witness statement, or the affidavit that supports it. A paragraph the model produced that you cannot trace to a source comes out, even if it sounds true.

3. Every quotation cross-checked. Models paraphrase quotations subtly; a quoted sentence in a brief that does not exactly match the source is a misrepresentation, even if the paraphrase is fair. Quotations are either verbatim with page references or they are recast as paraphrases without quotation marks.

4. Every cross-reference checked. “As argued in paragraph 7 above” must, in fact, be argued in paragraph 7. Models renumber and reorganise; the cross-references do not always travel correctly.

This four-step pass takes about an hour for a typical brief and about twenty minutes for a short submission. It is non-negotiable.

The defensive workpaper

A small habit that protects the lawyer when, months later, a question arises about how a particular submission was put together: keep a short workpaper, in the matter file, recording AI use.

The workpaper does not need to be elaborate. Five or six lines is enough. Tool used: Claude Sonnet, accessed via firm enterprise account on 14 June 2026. Inputs provided: verified fact statement (attached), section X of the Y Act, NKP citations [list]. Outputs accepted: structural draft of paragraphs 4 to 11. Outputs rejected: model proposed citation to “Sharma v. Government of Nepal” which could not be verified and was removed. Reviewed by: [senior name], [date]. That is the entire artefact. It lives in the matter file alongside the draft history and the partner’s mark-up.

What the workpaper provides is straightforward: if the brief is later challenged, the lawyer can demonstrate exactly what was done — what the model produced, what was accepted, what was rejected, who reviewed. It converts a “I used AI” admission from a vague liability into a documented professional process. It is also a forcing function: knowing that the workpaper will record what you rejected makes you read the output more carefully.

When does AI use become disclosable?

A question that increasingly comes up: at what point does AI use in a submission become material enough that it should be disclosed — to the client, to the court, or to opposing counsel?

The answer is moving, and the direction of travel is clear. To the client, increasingly: clients want to know, retainer letters increasingly require it, and the firms that have led on transparency have been rewarded. To the court, the position depends on jurisdiction. By 2026, several US courts and at least one UK court require an explicit disclosure that AI was used in the preparation of a filing, along with a certification that the citations have been verified. The Indian Supreme Court has issued guidance in the same direction. The Nepali courts have not yet formalised a disclosure rule, but the bar is discussing it and individual benches have begun asking the question informally when fabricated citations have surfaced.

The safer practice, in advance of any formal rule, is to be ready to answer the question if asked. If the court inquires whether AI was used in preparing a submission, the lawyer should be able to answer honestly, immediately, and with reference to the workpaper. The lawyer who can say “yes, the model was used to draft the structural skeleton, the citations were verified by me on the dates recorded in the matter file, and I take full professional responsibility for the submission” is in a much stronger position than the lawyer who has to improvise an answer.

Check your understanding

Quick check

—

A lawyer files a brief containing a fabricated case citation generated by a language model. When the citation is challenged in court, the lawyer explains that the model produced it. Does the use of AI change the candor obligation?

1. AYes — the duty is reduced because the lawyer did not personally invent the citation.
2. BYes — the duty does not apply when the typing was done by a machine.
3. CNo — the duty of candor attaches to the lawyer’s signature on the filed submission, and the use of AI as the drafting instrument is professionally irrelevant to the obligation not to mislead the court.
4. DNo, but only if the lawyer disclosed AI use in advance.

Check answer→

What comes next

Pleadings, memos, and court submissions cover the documents lawyers produce. The next chapter steps back to the underlying conditions that make AI use safe or unsafe in legal practice — client confidentiality and privilege, the ethical duties that frame the profession, and the shape that practising law in Nepal in 2026 is taking as the technology and the regulation around it both move.