A practitioner reads the news on 15 July. The Albanese government has committed to legislated AI standards, a new AI office, a national framework that takes the technology seriously. They file it somewhere in the back of their mind, feel a mild sense of resolution, and go back to preparing submissions.
Two months later, at a directions hearing before the Federal Court, a judge asks what authority supports the proposition in paragraph seven of those submissions. The proposition came from an AI-generated research memo. The memo looked right. The citation looked right. The practitioner opens the case on their laptop.
The case does not say what the memo said it said. And the practitioner has signed a document warranting that it does.
That moment was not waiting for Canberra to arrive. It was already live, governed by an instrument that came into force in April, months before the government's announcement. The regulatory clock the practitioner should have been watching had already run.
On 16 April 2026, Chief Justice Mortimer signed the Federal Court's Generative AI Practice Note. Victoria's Supreme Court, the County Court, and the Federal Circuit and Family Court issued their own AI practice notes across the same period. These are binding instruments, operative now, with professional consequences attached.
Each instrument shares a consistent obligation: verify every authority your AI generates before it reaches the court, and accept full professional responsibility for everything you file.
What "verify" means here is worth sitting with, because it is two obligations, not one. The Federal Court's practice note requires that cited legal authorities "support the stated proposition." A real case, misread or mischaracterised by an AI, does not satisfy that requirement. The practitioner signing the document warrants both that the authority is real and that it says what they say it says. Confirming the citation resolves only one of those handles. The other requires you to open the document.
At that directions hearing, the practitioner's problem is not that they failed to check whether the case existed. It is that the tool they used produced a plausible characterisation sourced from nothing in a closed corpus, from no real document at all. There is no source to open and defend. The failure is not a verification failure. The workflow produced nothing to verify.
The practice notes have a further dimension practitioners tend to underestimate. Entering compulsory process documents, privileged materials, or information subject to suppression orders into a standard open AI tool may breach the implied undertaking that governs such documents in litigation, even where sharing was not intended. The Federal Court's instrument distinguishes between open tools, where information may become accessible to third parties, and closed or ringfenced systems. For any practitioner whose AI workflow touches a client matter with sensitive materials, that distinction is now a professional risk with a named source.
The practical question the judge's question puts squarely is this: was the tool used to form this legal view built to produce an auditable chain from output back to source document? If it was not, the practitioner cannot answer the question.
The government's announcement is worth taking seriously on its own terms. Legislating AI standards at a national level is overdue. Clarity on data sovereignty, minimum standards for AI systems operating in regulated industries, and how Australian-generated content can be used in AI training are legitimate policy concerns. Questions about where Australia's AI infrastructure is built and on what terms are not trivial.
But those are upstream, industrial-scale questions. The framework's scope is data centres, energy capacity, and copyright. That is not the practitioner-level problem. The practitioner's problem is whether the tool used to form a legal view can be checked, sourced, and defended in proceedings.
A national framework designed to protect Australian intellectual property in AI training will not make a general-purpose model more accurate about how a High Court decision interacts with state legislation. A standard can require that an AI system perform at a certain level. No standard creates, retroactively, the closed corpus and jurisdictional architecture that produces verifiable Australian legal research in a tool not built that way from the start.
This is the structural point the Canberra announcement cannot reach. Accuracy on Australian primary law is a build problem, not a standards problem. The underlying corpus has to exist, be maintained in primary Australian sources, and be the thing the model reasons from. The practitioner sitting in chambers in September is not eighteen months away from a solution. They have a directions hearing next week.
Return to that hearing. The judge asks about paragraph seven. The practitioner whose research was grounded in a closed corpus of verified Australian legal sources has a different experience. Every output traces to an identifiable document. Each citation resolves to actual decided law, a document they can open, read, and show the court. The question does not produce a problem. It produces an answer.
That is the design the Federal Court's practice note contemplates. A practitioner whose workflow is auditable at the matter level, who can account for what tool was used, on what sources, to what end, has already satisfied the court's disclosure expectations. The practice note does not name a particular platform. It describes a set of capabilities, and the capabilities describe something built from the ground up for this purpose, not a general-purpose tool pressed into service.
Barristers who run their brief workflow on Habeas work from a search engine that scans more than 300,000 Australian cases and pieces of legislation, grounded in a closed dataset of legitimate Australian legal sources, results traceable and verifiable, sourced from no latent weight that cannot be checked. Foundational research that used to take a full morning takes minutes. When the judge asks what authority underpins the proposition in the submissions, the answer is in the documents, documented, not reconstructed from memory.
That is the answer paragraph seven's question requires. A workflow built on something else cannot give it.
The period between now and early 2027 will not be quiet. The court practice notes are the operative standard, enforced now, increasingly refined as courts gather data on how AI is being used in real proceedings. The Federal Court has signalled a symposium later this year at which those standards are expected to develop further. The direction of travel is toward greater specificity.
Firms and practitioners who treat those practice notes as their actual regulatory environment are not being conservative. They are reading the room accurately. Whatever eventually arrives from Canberra will sit on top of what the courts have already established, not replace it.
Reading the Prime Minister's announcement alongside the Chief Justice's practice note is clarifying. One is about infrastructure and copyright and governance at scale. The other is about whether the citation in your submissions resolves, and whether it says what you said it said. A practitioner who reads the first document and decides their AI question is on its way to being answered has confused two different regulatory exercises.
The AI rules that apply to their work were written in April. They are already in force. The question is whether the workflow is built to meet them.
Sign up at habeas.ai or book a demo to see how Habeas works in a research context.
If you want to try for yourself or get in contact, book a demo with us here. We also offer the capacity for self-serve individuals to sign up, and subscribe or register a free trial at app.habeas.ai.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Erik Mclean on Pexels
