It is eleven o'clock on a Wednesday. The brief is done, the citations are in, and the associate who drafted it has gone home. A partner is reading the final version before it goes out, running her eye down the list of authorities. They look right. They sound like real cases. She moves on.
That is the moment New Zealand's Supreme Court is now thinking about.
According to a report published by Newsroom NZ on 25 May 2026, the Court has flagged explicitly that submitting AI-generated citations that turn out to be fabricated could constitute obstruction of justice. Not a professional conduct matter. Contempt. The Court has moved from advisory to enforcement framing, and it has done so at the apex of a common law system that Australian courts cite regularly and take seriously.
The partner reviewing the brief is not careless. Neither is the associate who drafted it. Both are operating inside a workflow that was never designed to catch the structural flaw that contempt proceedings now attach to.
That flaw is worth naming precisely, because the profession has mostly been addressing the wrong version of it.
The framing most firms have accepted is a verification problem: check the output before it files. That advice is correct as far as it goes, but it undersells the real exposure. If your tool produces a citation and you then have to separately confirm that the citation is real, you are not using a legal research tool. You are using a drafting assistant with a hallucination risk baked into its design, and you are managing that risk manually, at the end of the workflow, under time pressure, as a final step before filing. That is exactly where verification fails. Not because lawyers are careless, but because the architecture puts the burden in the worst possible place.
Australian and New Zealand courts share more than geography. They share common law foundations, cite each other's decisions regularly, and have been moving in close parallel on AI guidance. The Federal Court's GPN-AI, issued in April 2026, drew a direct line between verification obligations and the professional duties that already exist. The NZ Supreme Court has now drawn the same line more sharply, and attached a consequence with real teeth. What one serious court in this tradition has said out loud, every other court in this tradition is thinking.
We would resist the conclusion that the lesson from Wellington is to slow down. Research that used to take a full morning can now be completed in minutes, and that return is genuine. Surrendering it in the name of caution is the wrong read on what the Court actually did. The Court did not criticise speed. It criticised sourcing.
The question the partner reviewing the brief needs to be able to answer is specific: was each citation retrieved from a real, locatable source, or was it generated by a model that produces fluent legal-sounding text and sometimes invents the cases it cites? In most firms, right now, the honest answer is: it depends on what the associate used, and nobody checked.
The only durable answer is a tool where the citation is the output, not an afterthought. Habeas's Search Engine scans over 300,000 Australian cases and pieces of legislation, drawing results from a closed dataset of legitimate Australian legal sources. Barristers running brief review and authority-gathering through Habeas are saving hours at the stage where the NZ risk lives, with citations traceable back to the primary source. The citation does not need to be checked separately because it was never generated from a token-prediction model. It came from the source. It points back to the source. You can open it.
That is not a feature added to a general-purpose tool. It is the architecture. And in a world where the NZ Supreme Court is reaching for the contempt power, the architecture is the point.
The partner reading the brief at eleven o'clock deserves to know the answer without having to check. If you would like to see how that works in practice for your firm, book a demo at habeas.ai.
Hero image: Leon Seibert on Unsplash
