Greens senator David Shoebridge has called the state of AI rules across Australia's courts and tribunals an "unholy mess." Each jurisdiction keeps setting its own practice notes and guidance, each with its own disclosure threshold, its own definition of what triggers a reporting obligation, its own idea of what a lawyer is supposed to have done before filing something an AI helped produce. Shoebridge's diagnosis is accurate. The patchwork is real, and anyone trying to maintain a coherent multi-jurisdictional practice has felt it.
A single national standard, the cure being implied, would not fix the thing that is broken.
The Victorian Law Reform Commission considered whether legislative reform was needed to govern AI in courts and tribunals. Its conclusion, in the report "Artificial Intelligence in Victoria's Courts and Tribunals," was firm: "we are not proposing legislative reform at this stage." The Commission found that "courts appear to have sufficient powers to establish rules, practice directions and practice notes governing use of AI."
That finding matters here because the Commission did what a senator's press comments cannot. It consulted, it examined the existing statutory and inherent powers of courts, and it tested whether a consolidated framework would add anything operationally. Its answer was that courts already have the tools. What they lack is a shared target to regulate toward, and that target is uniformity of substance, not uniformity of form.
The Commission also observed that the field is moving too quickly for legislative fixes to keep pace. Practice notes can be updated in weeks. Legislation takes months or years, by which point the technology and the professional practices around it have already shifted. That is a practical argument, not a principled one, but it has force. Courts issuing practice notes are responding to real problems in real time. A legislative process would freeze a snapshot of today's assumptions into statutory language that outlasts its relevance.
This archive alone has tracked separate AI instruments from the Federal Court, NSW, Victoria, Tasmania, and the Fair Work Commission. They vary in disclosure thresholds. Some require proactive disclosure of AI use. Others disclose only if asked. Some carve out research and administrative uses. Others draw the line at documents filed with the court. Some focus on affidavits and witness statements. Others sweep in submissions and chronologies.
A lawyer running a matter across two jurisdictions can face two different rule sets for the same document. That is messy. It creates compliance friction, it rewards jurisdictional errors, and it gives practitioners reason to complain.
The friction compounds in practice. A national firm litigating a matter that spans the Federal Court and the NSW Supreme Court might prepare a single chronology using AI assistance to organise dates from raw exhibit material. Under the Federal Court's GPN-AI, that chronology falls within the verification obligations: the responsible lawyer must personally confirm its accuracy. Under NSW's approach, the disclosure calculus differs. The lawyer has to hold both rule sets in mind while producing one document. That is where the "unholy mess" bites hardest, and it is also the strongest argument for harmonisation that Shoebridge could make.
Consolidating those regimes into one national practice note would solve the compliance-friction problem. A single disclosure threshold, a single trigger, a single form of words. Tidier. Easier to train staff on. Easier to audit.
The shared defect across every one of these instruments would remain exactly where it sits. Every version, state or federal, regulates disclosure of AI use. None of them regulates whether the lawyer verified the output before filing it.
Disclosure is a process obligation. It asks: did you tell the court you used AI? Did you describe how? Did you flag it in the body of the document? Those are answerable questions. They produce auditable records. They are also, on their own, weightless.
The substantive obligation underneath all of them, the one the Federal Court's GPN-AI states most plainly, is that the responsible lawyer must personally confirm cited authorities exist, that evidence referenced is in the materials, that facts in pleadings can be proved, that chronologies are accurate. That obligation does not change depending on which jurisdiction's practice note you are reading. It is the same duty whether you are in the Federal Court, the NSW Supreme Court, or the Fair Work Commission. It is the duty that existed before any AI guidance was published, and it is the duty that will exist regardless of how the patchwork is harmonised.
Disclosure regimes tell the court that AI was involved. They do not tell the court, or the lawyer, whether the work product is correct. A lawyer who discloses AI use and files a hallucinated citation has complied with the disclosure obligation and breached the substantive one. A lawyer who discloses nothing and files verified work has breached the disclosure obligation and satisfied the substance. Which of those is the problem?
The disclosure-versus-verification distinction also explains why the patchwork varies so much in the first place. Courts are not disagreeing about whether lawyers should verify their work. They are disagreeing about what to make visible to the court. Some judges want to know which documents touched AI so they can read them with appropriate scrutiny. Others treat AI as a drafting tool indistinguishable from a precedent library and see no need for proactive disclosure unless something goes wrong. Both positions are defensible. Neither addresses whether the lawyer checked the citations.
Framing the patchwork as the disease misses the point. The disease is that disclosure is easy to regulate and verification is hard, so the guidance regulates the easy one. Consolidating ten disclosure regimes into one national disclosure regime produces a tidier version of the same gap.
The VLRC's conclusion that courts have "sufficient powers" is worth sitting with. The Commission was not saying the current patchwork is ideal. It was saying that the mechanism for fixing it, practice notes and directions, already exists and already works. Courts can set rules. They can update them. They can learn from each other, as the Federal Court's GPN-AI clearly drew on prior state-level guidance. The evolutionary process is doing what evolutionary processes do: producing variation, some of which is better than the rest.
What the Commission could not say, because no practice note can, is what the verification standard requires a lawyer to have done. That is a practice gap, not a legislative one. The rule says "confirm the citation exists." The rule does not say how, against what source, with what degree of confidence, or in what time frame. That is left to the lawyer, which is where it belongs, and also where the risk lives.
Consider what "personally confirm" demands in a litigation context. A lawyer drafting submissions late on a Friday needs to verify that a case stands for the proposition being cited. In a pre-AI workflow, that meant pulling the report, reading the relevant passage, and satisfying oneself that the summary was fair. The AI-assisted version of that workflow produces a citation that looks correct, reads fluently, and may still be wrong in ways that only checking the primary source would reveal. The practice note says confirm it exists. The professional obligation says confirm it stands for what you say it stands for. The gap between those two formulations is where a lawyer can be technically compliant and substantively exposed at the same time.
Shoebridge is right that a practitioner should not need a spreadsheet to track which courts require proactive disclosure and which disclose-if-asked. Harmonisation at the level of form would help. But the practitioner who is exposed under any of these regimes is the one who filed an AI-assisted document and could not stand behind every citation in it.
That exposure does not shrink if the disclosure regimes are consolidated. It shrinks if the lawyer's research process produces verifiable, traceable results. The gap between disclosure and verification runs through every practice note, and closing it is a question of tools, not legislation.
Habeas was built to close that specific gap. The Search Engine scans over 300,000 Australian cases and pieces of legislation, with results grounded in a closed dataset of legitimate Australian legal sources, so they are verifiable and traceable, never hallucinated. Every answer carries its citation to the source. A lawyer who has run research through that pipeline can answer the verification question with the source document behind it. The substantive obligation the practice notes point at but cannot enforce becomes answerable, because the tool is what makes the work possible in the time available.
The VLRC said it is too early for legislative reform. We agree, though for a narrower reason than the Commission gave. Reform at the legislative level would harmonise the disclosure layer and leave the verification layer untouched. The thing missing from the current patchwork is a research process that produces results a lawyer can stand behind in any court, under any practice note, in any jurisdiction. Build that, and the patchwork becomes manageable. Leave it unbuilt, and a single national standard reproduces the same exposure in one clean font.
Book a demo at habeas.ai.
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: Héctor Berganza on Pexels
