A Citation You Can Defend Is Not a Decision You Can Defend

On 11 February 2026 the Supreme Court declined to hear an appeal in a long-running family dispute. Its reasons took six paragraphs. Then it added a seventh, under a heading of its own, about something else.

The applicant, who was representing himself, had filed submissions citing authorities that the Court said appeared to have been hallucinated by an AI application. Some paired real case names with citations that belonged to other cases. Four more were genuine cases, wrongly cited and relied on for propositions they do not support. The Court said that anyone filing submissions must make sure every authority is genuine and correctly cited, quoted the courts’ guideline for non-lawyers, and noted that relying on false citations may, in serious cases, amount to obstruction of justice or contempt of court.

That was Jones v Family Court at Whangārei. It is the most prominent of a small run of New Zealand decisions on the same problem, and it shares one feature with every other we have read so far.

The pattern so far

In March 2025 the Employment Court dealt with a plaintiff, appearing for herself, who had asked for more time to pay security for costs. In support she cited Fowlie v Stonex (No 2) for the proposition that the Court takes financial constraints into account in procedural rulings. No such case exists. Judge King reminded her that generative AI output ought to be checked before it is relied on — and then, on conditions, gave her the extension anyway. That was LMN v STC (No 2). The defendant had counsel.

In July 2025, at an Authority investigation meeting, Mr O’Driscoll — also appearing for himself — confirmed he had used AI to help present his case. The Member reminded him that checking what he relied on was his responsibility, and asked for copies of two cases he had cited, which he provided. The later costs determination recorded that his oral submissions had referred to cases that did not exist or could not be found, and that his final written submissions then cited more of them. He lost: the Authority found he had been a director and shareholder working for sweat equity, not an employee. The respondents, who had counsel, sought costs. The Authority took its daily tariff of $4,500 as the starting point and lifted the award to $6,500 in O’Driscoll v Rehab Co Mobile Ltd. The uplift covered several things — delays, extra steps the respondents had to take, witnesses who never came — and the citations were among them. The Member called it a modest uplift, given that he was a litigant in person. The Member recorded that the guideline had been drawn to his attention at the meeting, and that the non-existent cases came back regardless.

Add the Court of Appeal in Wikeley v Kea Investments Ltd, the Employment Court again in Cunningham v healthAlliance NZ Ltd, and the Authority again in QTR v BXD, and the picture holds. In every New Zealand decision of this kind that we have read, the party relying on the phantom authority was appearing in person. That is not proof that no New Zealand lawyer has ever done it. Overseas, plenty have, and a decision we have not found is still a decision. But the decided cases here sit, so far, with people doing their own advocacy.

Small employers are in that group too

None of those parties was an employer. Mr O’Driscoll, as it happens, was a company director — but he was the one bringing the claim.

Still, the Authority is built to be workable without a lawyer, and small employers use it that way. A sole director answers a grievance from the kitchen table, with whatever help is at hand, and more and more often the help at hand is a chatbot. The courts’ guideline is written for non-lawyers, not for employees. It applies to the director as squarely as to the person who was dismissed. And nothing in the Authority’s reasoning in O’Driscoll suggests the costs consequence belongs only to applicants; it followed the conduct, not the side of the table.

So the first lesson is simple, and the courts have already stated it. Check every authority before it goes in. And check more than whether it exists — Jones shows the subtler failure, the real case cited for something it never said. A checker that only confirms a case exists would not catch that.

The smaller lesson

That is where most commentary on these decisions stops. It is the smaller of the two lessons.

Research tools are getting better at tracing their answers back to sources a person can check. That is exactly what every one of these litigants lacked, and it is worth having. But look at where all of these failures happened. In submissions. After the event. In front of the person deciding the case.

For an employer, the decision under review was made long before anyone filed anything. Section 103A of the Employment Relations Act 2000 asks whether the employer’s actions, and how it acted, were what a fair and reasonable employer could have done in all the circumstances at the time the dismissal or action occurred. The test is fixed at that moment. Nothing filed afterward, however well sourced, changes what the employer did then.

There are two layers here, and they answer different questions. The research layer asks what the law requires. It can be answered from sources, and anyone with access to those sources can check the answer. The decision layer asks whether this employer did what the law requires — whether it sufficiently investigated, having regard to its resources, raised its concerns before acting, gave the employee a reasonable opportunity to respond, and genuinely considered what the employee said. Those are questions of fact about one employer and one employee, and the only evidence of the answer is the record of what happened.

A manager can ask a chatbot whether a particular kind of misconduct justifies dismissal. The answer can come back anchored in genuine, correctly cited authorities. The manager can check every one of them, and still dismiss unjustifiably — because the answer described the standard, and nobody met it. The citation was defensible. The decision was not.

O’Driscoll makes the same point from the other side of the table. Better citations would not have rescued Mr O’Driscoll’s claim. It turned on who controlled the money, who did the hiring, and what the parties had intended when the business began. Facts, not authorities.

Two checks, two questions

For an employer using AI anywhere near an employment decision, that means running two different checks and not mistaking one for the other.

The research check comes first, and it is the one the courts have been talking about. Does each authority exist? Does it say what it is said to say? Is it still the law? That is the verification layer, and it can be done carefully and well.

The decision check is the one the Authority actually applies. Can we show what we did, when we did it, and why? The notes of the meetings. The letter that put the concerns out. The employee’s response, and how it was weighed. And if AI drafted any of it, can we show that a human read it before it went out? In O’Driscoll, the Authority recorded the use of AI before it had even set out the issues for determination.

The research check decides whether your submissions survive the afternoon. The decision check decides whether your dismissal does.

Lex Praxis advises New Zealand employers on employment law and AI governance. This article is general commentary and does not constitute legal advice. For advice specific to your circumstances, contact us directly.