There is a principle that has been present in professional life long before AI existed. It is simple enough that it rarely needs stating: you are responsible for what you put your name to. The document you file, the letter you send, the report you submit — these belong to you. The tool you used to produce them does not change that.
New Zealand's Supreme Court has now said this in terms that should settle any remaining ambiguity. In Jones v Family Court at Whangārei [2026] NZSC 1, decided in February 2026, Justices France, Kós, and Miller observed that misuse of AI in legal proceedings has serious implications for the administration of justice and public confidence in the justice system. They confirmed that reliance on false citations — including the unverified outputs of AI applications — may in serious cases amount to obstruction of justice or contempt of court.
The case involved a self-represented litigant. The Court was explicit that its observations apply with equal — and greater — force to lawyers. By extension, the same principle applies to anyone using AI to produce output that is then relied on in a consequential process.
That includes employers.
What happened, and what the Court said
The underlying proceedings were a protracted family law dispute. What brought the case to the Supreme Court's attention was not the merits of the application, but the quality of the submissions filed in support of it. Those submissions cited several legal authorities that, in the Court's words, "appear to have been hallucinated by an Artificial Intelligence (AI) application."
The examples the Court identified are instructive. Citations such as "Teddy v Police [2015] NZSC 62" and "Baird v R [2013] NZSC 120" combined real case names with incorrect citations. The genuine cases bearing those names were of no direct relevance to the application. In other words, the AI had constructed plausible-looking authorities from real names and fabricated references — and those fabrications made it into the Court.
The Court's response was measured but clear. Persons filing submissions must ensure all authorities referred to are genuine and correctly cited. The responsibility for accuracy rests with the person filing, not the tool. And in serious cases, the consequence of getting this wrong is not merely a professional embarrassment — it is potential criminal liability under the Crimes Act 1961 or the Contempt of Court Act 2019.
The Court was not criticising AI. It was confirming something more fundamental: that using a tool does not transfer responsibility for the output. That principle does not stop at the courthouse door.
What this means for lawyers
For lawyers, the implications are sharpest. The Court was explicit that its observations apply with greater force to legally qualified practitioners. The reason is Rule 13.1 of the Lawyers and Conveyancers Act (Lawyers: Conduct and Client Care) Rules 2008, which requires lawyers to act with competence, diligence, and in accordance with their duties to the court.
A self-represented litigant who submits AI-hallucinated authorities faces the risk of contempt proceedings. A lawyer who does the same faces that risk, plus a professional conduct inquiry, plus the question of whether they have breached their duties to the client who trusted them with the matter.
The practical question is not whether lawyers are using AI — many are, and there is nothing wrong with that. The question is whether the verification step is being treated as optional.
It is not. The Courts of New Zealand GenAI Guidelines state plainly that AI chatbots "can make up fake cases, citations and quotes, or refer to legislation, articles or legal texts that do not exist." The Supreme Court's decision in Jones is a confirmation that the Guidelines' underlying concern is not theoretical. It is already reaching proceedings.
The verification obligation is not new. What is new is the consequence of ignoring it when AI is involved. A lawyer who relies on a junior's research without checking it is arguably negligent. A lawyer who relies on AI output without checking it, and files fabricated authorities as a result, is in a materially more difficult position.
Seven cases of AI-hallucinated citations have now been documented in New Zealand proceedings, all identified since 2024. All involved self-represented litigants — so far. The profession should not read that pattern as reassurance.
What this means for employers
The courthouse is not the only place where AI output is relied on in consequential processes. Employers rely on it in hiring, performance management, disciplinary decisions, restructuring analysis, and workplace investigations. The accountability principle the Supreme Court articulated applies in all of those contexts.
Consider the disciplinary process. An employer uses an AI tool to help draft the letter putting allegations to an employee. The letter is reviewed by a manager, approved, and sent. The employee challenges the process at the Employment Relations Authority. The Authority examines the letter and finds that the factual assertions it contains are inaccurate — not because of managerial carelessness, but because the AI generated them and no one checked.
The employer owns that letter. The fact that AI drafted it is not a defence. The s103A justification test under the Employment Relations Act 2000 does not ask how the document was produced. It asks whether the employer's process was fair and reasonable. An inaccurate allegation letter does not satisfy that test, regardless of its origin.
The same logic applies to AI-assisted performance assessments, redundancy reports, investigation summaries, and any other document that forms part of an employment process. If the AI-generated content is wrong — factually, legally, or contextually — and that error is not caught before the document is used, the employer has a problem that the tool cannot solve.
The ERA does not care what generated the letter. It cares whether the process was fair. An AI-generated error that was not verified before being acted on is the employer's error.
The verification gap
The Jones decision is useful not only for what it decides, but for what it reveals about how AI failure actually happens. The litigant presumably did not intend to mislead the Court. The AI produced plausible-looking output. The litigant accepted it. Nobody checked.
That sequence — plausible output, uncritical acceptance, consequential reliance — is the standard pattern of AI failure in professional settings. It is not a story about bad faith. It is a story about misplaced confidence in a tool that presents fabrications with the same confidence it presents accurate information.
This is what makes AI output different from most other forms of professional assistance. A junior lawyer who cites the wrong case will usually know they have done so. An AI tool that hallucinates a citation presents it without qualification, without doubt, and without any signal that the reader should treat it differently from verified output — and will need someone to supervise the output. Because AI, as useful as it is, makes mistakes.
The verification gap is the space between what the AI produces and what is actually true. Closing that gap is the professional responsibility that Jones confirms belongs to the person using the tool.
What responsible use looks like
None of this means AI should not be used. The Courts of New Zealand GenAI Guidelines, the Law Commission's ongoing review of automated decision-making, and the emerging body of case law all point toward the same destination: AI as a capable instrument, governed by human judgment and verified before reliance.
For lawyers, responsible use means:
- Verifying every citation independently before filing — not trusting the tool's confidence as a substitute for checking the primary source.
- Treating AI-drafted submissions as a starting point, not a finished product. The same scrutiny applied to a junior's work applies here, and then some.
- Retaining the record of what AI produced and what human review occurred, so that process can be demonstrated if challenged.
- Requiring human review of any AI-generated content before it is used in an employment process — disciplinary, investigative, or otherwise.
- Checking AI output against the underlying facts. Not accepting that because the output reads well, it is accurate.
- Documenting that review contemporaneously. The ability to show that a human checked and confirmed the AI's output is the difference between a defensible process and an exposed one.
Jones v Family Court at Whangārei is the strongest NZ authority to date on AI hallucination risk in professional practice. It will be cited. It will be used by counsel in proceedings where AI misuse is alleged. It may eventually inform the Law Commission's framework for automated decision-making.
What it already does, right now, is settle a question that some practitioners and employers have been treating as open: whether the accountability for AI output rests with the tool or the person who used it.
It rests with the person. It always did. The Supreme Court has simply confirmed it.