This is a Tautoru view, not a settled position.
Using an AI tool does not change your professional obligations; it gives you a new way to engage them. For lawyers the touchstones are the Conduct and Client Care Rules — competence and diligence, confidentiality (chapter 8), the duty not to mislead the court, and proper supervision — read alongside the New Zealand Law Society's generative-AI guidance and the courts' own practice notes. For public-sector users, the Government Chief Digital Officer's and Public Service guidance points the same way. This page is about how to comply with them.
Competence, and the duty not to mislead the court, mean you must understand the tool's limits and check its output. These models produce fluent, confident text that can be wrong — invented cases, citations in the right format that do not exist, a statute section that says something it does not. Filing that, or relying on it, is your responsibility, not the tool's: verify anything that matters against the source. The risk is not hypothetical.
Confidentiality and privacy are obligations in their own right, dealt with on their own pages here: keep client and other confidential material out of tools that do not keep it confidential, obtain authority where the obligation requires it, and run the Privacy Act analysis before any personal information goes to an overseas tool. Where a matter is before a court, check that court's expectations — some practice notes ask that the use of generative AI be disclosed, and all of them expect suppressed or protected material to stay out of these tools.
Diligence and supervision extend to how your firm or organisation uses these tools: have a clear policy on which tools and plans are approved, and for what; train the people who use them; and supervise juniors and staff as you would on any other delegated work. The duty to provide competent, careful services does not soften because a machine drafted the first version.
Our view
The professional guidance is, on the whole, sound, and the precautionary line on public chatbots is right. Where the broadest reading of it goes too far is in treating every tool and plan alike. The worry underneath a blanket "keep it all out of AI" is that typing something in "feeds it into the model" — that your words become part of what the model knows, or train the model others use. On a plan that contractually does not train on your inputs, that is not what happens: your text is processed to produce a reply and held under the retention terms, not added to the model's training data. Complying with the rules does not require avoiding these tools; it requires using the right ones, in the right way, and standing behind the work. How that applies to your matter is your judgment, on your facts.