In short. The legal profession's alarm over Artificial Intelligence is, in large part, the defence of an old business model dressed as an ethical stand. New Zealand's courts have not banned the technology: the published guidance, and decisions such as LMN v STC, ask only that a practitioner verify what they file — the oldest duty in the book. Yet the same jurisdiction that treats a machine's error as a scandal tolerates the opaque, subjective judgements of human managers every working day, and in MW v Spiga Limited it has already recognised that Artificial Intelligence in the hands of employers can lock workers out of justice. The honest question, under the Employment Relations Act 2000, is never whether a tool was used, but whether the human who used it did their job.
From the New Zealand Law Society to the chambers of the Employment Relations Authority, much of the profession speaks of Generative Artificial Intelligence as a contagion: hallucinating models, eroded ethics, careers ended over a fabricated citation. Overseas, the alarm has hardened into rule. In the United States the well-publicised matter of Mata v Avianca Inc ended in a five thousand (5,000) United States dollar sanction after lawyers filed, and then defended, citations that a chatbot had invented; some judges responded with standing orders requiring litigants to certify whether Artificial Intelligence had touched a filing. It is tempting to read that as the shape of things to come everywhere.
It is not the shape of things here. New Zealand's judiciary did not ban the technology or demand a certificate with every document. The Guidelines for use of generative artificial intelligence in Courts and Tribunals, issued on Thursday the 7th of December 2023 and adopted across the courts and tribunals (the Employment Relations Authority among them), impose no routine duty to disclose that Generative Artificial Intelligence was used. They say something far older and far harder to argue with: your existing professional obligations still apply, and you are responsible for the accuracy of everything you put before the decision-maker. The panic, in other words, has run well ahead of the rules. What the rules require is competence. That is not a new burden invented for a new machine; it is the burden every advocate has always carried.
There is also less novelty here than the alarm suggests. Strip away the word "intelligent" and a large language model is, at bottom, a very large autocomplete: a statistical engine that predicts the next word from the patterns in everything it has read. That is the same lineage as the predictive text that finishes a sentence in an email and the search box that guesses a query from three letters — and it sits on the same continuum of everyday automation as the autofill that completes a form or a login, none of which the profession has ever called a menace.
The point sharpens at the very desk where the research is done. The platforms a practitioner logs into — LexisNexis and Thomson Reuters — have built generative artificial intelligence directly into their tools: LexisNexis now ships its assistant as Lexis+ with Protégé, and Thomson Reuters has embedded CoCounsel and Deep Research inside Westlaw and Practical Law. The technology the profession warns against is not waiting at the gates; it is already inside the trusted databases lawyers pay to use, dressed as a feature rather than a threat. A profession cannot coherently treat the same predictive machine as indispensable infrastructure in one window and an existential danger in the next.
Consider the everyday substance of an employment dispute: selection for redundancy. The Employment Relations Authority and the Employment Court scrutinise selection criteria that are routinely subjective, undocumented, and opaque — and they do so without anyone suggesting the tool of human judgement should be outlawed.
In Mateiwai and ors v Sensation New Zealand Ltd [2004] NZERA 416, the people placed on the redundancy list were chosen because supervisors' "current perceptions of skill levels, attitude and teamwork" were treated as the primary determinants. In Stevenson v Chief Executive of The Auckland University of Technology [2001] NZEmpC 184, a manager allocating points to staff candidly accepted that, alongside the documented record, he drew on his own "personal 'insider' knowledge" of the individuals. These are human algorithms — weightings applied inside a person's head, on grounds nobody can fully reconstruct. The institutions test them, remedy them where they are unfair, and move on. No one calls for the manager to be struck off for using discretion.
The instructive case is the one where opacity itself was the fault. In Gilbert v Transfield Services (New Zealand) Limited [2013] NZEmpC 71, the employer ran affected staff through an online psychometric assessment and could not explain to the worker, or to the Employment Court, how it worked or why it produced the scores it did.
The Employment Court found that an employer's decision to use an assessment tool that was "incapable of meaningful explanation" made it impossible to comply with the duty in section 4(1A) of the Employment Relations Act 2000 to give the worker access to the information on which the decision turned. The flaw was not that a tool had been used; it was that the result could not be justified or examined.
Gilbert v Transfield Services (New Zealand) Limited [2013] NZEmpC 71 at [113]
That holding is the whole point in miniature. A poorly understood algorithm fails because its output cannot be explained or tested against the duty of good faith in section 4 of the Employment Relations Act 2000 and the test of justification in section 103A of the Employment Relations Act 2000 — not because it is a computer. A manager's unconscious bias, hidden behind "insider knowledge", is at least as opaque, and far harder to audit. The profession's terror of machine error, set against its calm acceptance of human error, is not a principle. It is a preference for the familiar.
The strongest evidence against the panic is the case the panic likes to cite. In LMN v STC [2025] NZEmpC 46, a self-represented litigant relied on a precedent that did not exist; the Employment Court recorded that no such case could be found and reminded her that material from Generative Artificial Intelligence "ought to be checked before being relied on" in documents filed in court. Note what the Court did not do. It did not ban the technology, did not threaten the litigant's standing, and did not treat the tool as the wrong. It restated the duty to verify.
None of this is to pretend the risks are imaginary. They are real, and the same body of law shows why. A chatbot can fabricate an authority; entering a client's confidential or suppressed material into a public tool can breach privilege, privacy, or a non-publication order; an unexamined model can smuggle bias into a decision. Those are serious harms. But each of them is a failure of a person — of diligence, of confidentiality, of competence — that the existing rules already name and already punish. The harm is the unchecked filing, not the keystroke that produced a first draft.
The privacy argument deserves to be taken at its word — and then tested for consistency. The New Zealand Law Society's own generative artificial intelligence guidance for lawyers rests squarely on confidentiality and on Information Privacy Principle 12, in section 22 of the Privacy Act 2020 (disclosure of personal information outside New Zealand): it warns that anything typed into an overseas tool may be stored, seen, or reused beyond the practitioner's control, and that private, confidential, or privileged client information should be kept out of such tools altogether. That concern is entirely sound. The difficulty is that the very same principle is pressed hard against one tool and quietly waived for every other.
Consider what passes without a murmur. Privileged client files are routinely drafted in, and synchronised through, the cloud services of overseas providers — Google, Microsoft, Adobe — whose servers and parent companies sit outside New Zealand and whose lengthy user agreements almost no one reads in full before clicking "agree". The very Information Privacy Principle 12 invoked against a chatbot applies, word for word, to a privileged document stored on an overseas server by one of those companies. Confidential material is still carried between offices on unencrypted portable Universal Serial Bus (USB) drives and sent as plain, unencrypted electronic mail attachments — a far more common and more careless exposure than anything a language model has been blamed for.
There is a harder edge still. Information held by a United States company is reachable by United States legal process. The Clarifying Lawful Overseas Use of Data Act of 2018 (the CLOUD Act) added section 2713 to title 18 of the United States Code — the Stored Communications Act — requiring a United States provider to preserve, back up, or disclose the contents and records within its "possession, custody, or control" regardless of whether the data sits inside or outside the United States. A single subpoena issued in an American court, directed at the ordinary Google, Microsoft, and Adobe accounts a practice runs on, could surface a remarkable amount about a person and their affairs — and would lay bare, in passing, just how much privileged material already sits within reach of a foreign court while the profession trains its alarm on Artificial Intelligence alone.
The point is not that practitioners should abandon the cloud, the office suite, or the design tools they depend on. It is that a privacy principle is not a principle at all if it is enforced against one disfavoured tool and excused for the convenient ones. Either confidentiality and overseas disclosure are taken seriously across the board — encrypted media, user agreements actually read, deliberate choices about where client data is allowed to live — or the privacy objection to Artificial Intelligence is revealed as something other than a concern for privacy.
And if confidentiality really is the worry, the remedy is plain, and it is not prohibition. A practice that is genuinely anxious about client data leaving its control can build and run an internal, offline model — hosted on its own hardware, behind its own firewall, with nothing transmitted to an overseas provider and nothing crossing a border. That answers the principle completely, because no disclosure outside New Zealand ever occurs; the tool can be made as private as a locked filing cabinet. A profession that reached for that solution would be taking confidentiality seriously. A profession that reaches instead for a blanket suspicion of the technology, while its privileged files sit on overseas servers and travel on unencrypted drives, is doing something else.
Strip away the ethical language and a plainer anxiety appears. The traditional fee model rewards time spent, and a great deal of billable time has historically been spent on tasks — combing databases, summarising precedent, drafting a first statement of problem — that software can now do in minutes. Technology of that kind does not make a good advocate obsolete; a lawyer's real worth lies in strategy, judgement, advocacy and care for the client, none of which a model supplies. What it threatens is the price of the routine. To insist that any use of Artificial Intelligence is inherently perilous is, conveniently, to keep clients paying artisanal rates for work a machine performs at a fraction of the cost. That is the protection of a revenue stream, not the protection of the public.
The deepest irony sits in the jurisdiction's own leading authority on privacy. The Employment Relations Authority was built, under section 157 of the Employment Relations Act 2000, as a specialist investigative body so that working people would not always need expensive representation to be heard. Yet real barriers remain. In MW v Spiga Limited [2024] NZEmpC 147, a full bench of the Employment Court considered a worker who feared that being named in a decision would see prospective employers draw "adverse conclusions about their suitability" — and the Court squarely confronted the prospect that employers, using algorithms and Artificial Intelligence to screen applicants against their litigation histories, could quietly blacklist anyone who had ever enforced a right.
Hold the two fears side by side. The profession frets that Artificial Intelligence might put a wrong word in a lawyer's mouth. The Employment Court, in the very same period, frets that Artificial Intelligence might be used to punish a worker for the rest of their career for daring to bring a claim. Only one of those is treated as an existential crisis, and it is not the one that falls on the worker. The same technology that could lower the gate — helping an unrepresented person frame a coherent claim, understand the test of justification, or navigate a process — is the technology the establishment is most anxious to keep at arm's length. The fear is not that Artificial Intelligence will get it wrong. It is that it will get it right often enough, and cheaply enough, to loosen the profession's grip on the gate.
The logical error at the heart of the panic is to regulate the tool rather than the work. An advocate is accountable to their client and to the decision-maker for the integrity of every document they lodge, whatever its source. Whether a submission was drafted by a junior, lifted from a precedent, or generated by a model, the duty to read it, check it, and stand behind it is identical. An order requiring a practitioner to certify that a machine's work was "checked by a human being" is very nearly a tautology: that checking is the definition of the job. A practitioner who files fabricated authorities should answer for it — not because they used Artificial Intelligence, but because they failed the duty of competence that the Lawyers and Conveyancers Act 2006 and the conduct rules already impose. The instrument is beside the point.
No one proposes to ban the word processor that introduces a typographical error, or the database that carries an outdated headnote. The selective demonising of one tool, and only one, betrays the absence of any coherent principle behind it. The institutions established under Part 10 of the Employment Relations Act 2000 — the Mediation Service, the Employment Relations Authority, and the Employment Court — exist to deliver justice that ordinary people can actually reach. The challenge for the profession is not to fight the future, but to harness it: to insist, as the rules already do, on competence and accountability, and to let the tool be a tool.
Legislation cited is linked to the current consolidated text on the New Zealand Parliamentary Counsel Office website at legislation.govt.nz. Case authorities are linked to the Working for Workers case law and reference repository at caselawandreference.workingforworkers.nz. Stevenson v Chief Executive of The Auckland University of Technology [2001] NZEmpC 184 is held in the repository as an image-only scan; the passage attributed to it has been confirmed against the body of the judgment. The Guidelines for use of generative artificial intelligence in Courts and Tribunals are linked to the Courts of New Zealand website at courtsofnz.govt.nz. Mata v Avianca Inc is a United States authority and is not held in the repository; it is referenced here only as a widely reported overseas illustration and its details should be confirmed against the official report before being relied upon in any filed document. The Clarifying Lawful Overseas Use of Data Act of 2018 is United States federal legislation; its text is linked to, and confirmed against, the copy published by the United States Department of Justice. The LexisNexis and Thomson Reuters product names (Lexis+ with Protégé, CoCounsel, and Deep Research) are current as at Monday the 22nd of June 2026 and are used for identification only. This article is opinion and commentary; it is not legal advice and does not establish a lawyer–client or advocate–client relationship.
Whether the decision came from a manager's undocumented judgement, an automated screening system, or any public body exercising a statutory power, the question is the same: was the process fair, and can it be explained? Working for Workers helps people stand up to decisions they were never allowed to see inside.
Working for Workers stands up for people facing unfair and unreasonable process, across four areas of practice:
Fairness in Law · Accident Compensation Corporation (ACC) Law · Employment Law · Incorporated Societies Evaluation
Whatever the decision-maker, if you are up against an unfair or biased process, get in touch and we will help you work out where you stand.