The matter is a no-adverse-action claim. A protected complaint, a subsequent termination, a client asking how strong the case is. The practitioner is running four other files that week and is not a specialist in employment law. They put the question to a general-purpose AI tool. The answer comes back structured and confident, with what sounds like a coherent line of authority. The reverse onus under the Fair Work Act, the feature that defines how these claims run, the reason defendants carry the burden of disproving the connection, is absent from the answer. The burden is characterised the way US employment law characterises it, because US employment law dominates what the model absorbed during training. The practitioner, consulting the tool precisely because they do not already know this corner of the law in granular detail, reads the output and drafts accordingly. The advice goes out.
This is the failure mode that does not announce itself. A hallucinated citation is easy to catch: you look it up and it is not there. A fluent, jurisdiction-miscalibrated answer looks exactly like research, and keeps looking like it until someone who already knows the area reads it.
Writing in Artificial Lawyer this week, LawVu's CEO made an argument that names the underlying mechanism: the competitive edge in legal AI has already moved away from the model itself. GPT, Claude, Gemini, the choice between them is becoming a procurement footnote. The real differentiation lives in the data a system can reach, the workflows it operates inside, and the institutional knowledge it draws on. The model is infrastructure. The system around it is the product. We think he is right. We also think the argument has a sharper edge for Australian legal practice than the piece draws out.
The model-as-commodity thesis tracks a pattern that has played out before. Processing power commodified. Cloud storage commodified. The underlying capability becomes a utility, and value migrates to whoever built the most useful layer on top. Every serious legal AI product will run on one of the same three or four foundation models within eighteen months, if it does not already. At that point, asking which model it uses is roughly equivalent to asking which electricity grid powers the office.
The ceiling on this argument is real. A model that cannot reason through multi-step legal problems is still a constraint, regardless of what data surrounds it. But once foundation models reach a competence floor for legal reasoning, which most of them have, differentiating on model alone stops predicting practitioner outcomes. That is roughly where the market sits now.
Return to the practitioner whose advice went out wrong. The system was trained predominantly on English-language legal material, and most of that material reflects US and UK legal architecture. Common law, yes, but common law shaped by different court hierarchies, different statutory frameworks, different doctrinal traditions. The reverse onus point is not a minor caveat. In a no-adverse-action matter, it is the architecture of the case. A practitioner who already knew that would have caught the error. One building their understanding from the tool's output had no way to notice what was missing.
This is why the jurisdiction question is the central one, and why the LawVu thesis applies with particular force to Australian law. Training on Australian primary law differs from grounding a system's responses against a closed corpus of Australian cases and legislation at query time. And a system that returns citations traceable to the actual source document is a different tool again from one that synthesises whatever the model absorbed during pre-training. The practical difference shows up in whether a practitioner can verify an answer before it goes anywhere near a decision, or whether verification requires returning to the sources the tool claimed to summarise, which negates most of the efficiency gain.
A GC advising across privacy, employment, commercial contracts, and regulatory compliance in a single week cannot re-research every answer from scratch. If the tool's output requires independent verification to be trusted, the efficiency disappears. What remains is extra work wearing the appearance of automation.
The court guidance material makes the stakes concrete. The Federal Court's GPN-AI, signed by Chief Justice Mortimer in April this year, requires practitioners to personally confirm that cited authorities exist and support the stated proposition, regardless of how the draft was produced. The practitioner's signature carries that weight. NSW, Victoria, Queensland, and the Fair Work Commission have all issued AI guidance across the past eighteen months, landing in the same place: verify before you file, accept professional responsibility for every output, be able to account for AI involvement if asked. The pattern is consistent enough to treat as the default regulatory posture now, not a jurisdiction-specific exception.
For the practitioner whose advice went out wrong, this is where the trap closes. The jurisdiction miscalibration was the problem. The court guidance makes it a professional liability. A fluent answer calibrated to the wrong jurisdiction is no longer only a research error; it is a file where the practitioner signed off on something they could not verify, and now cannot account for.
We have been guilty, as an industry, of presenting the distinction between these systems as marketing copy. It describes a real liability allocation: the difference between an answer that can be trusted to reflect current Australian authority and one that requires an independent check before it goes anywhere near a decision.
What the practitioner from the opening needed was a system grounded in Australian primary law with citations traceable to the source. The Search Engine in Habeas scans over 300,000 Australian cases and pieces of legislation in seconds, drawing from a closed dataset of legitimate Australian legal sources, so results are verifiable and traceable rather than synthesised from ambient training data. Foundational research that used to take a full morning can be completed in minutes, with every authority traceable to the document the practitioner can pull up and read. In a no-adverse-action matter, that means the reverse onus surfaces in the research output because it is present in the Australian sources. The practitioner sees the authority. They can check it. The advice that goes out reflects Australian law.
The LawVu CEO's piece is part of a broader recalibration that is overdue. The legal AI conversation spent much of 2023 and 2024 on model benchmarks. It is shifting toward the harder question: what can the system reach, and how do you know the answer is grounded in the right sources for your jurisdiction? Practitioners who have already asked that question tend to notice, quickly, that the answer changes how they work.
The model is a utility. For Australian law specifically, the corpus it reaches, the citations it returns, and whether those citations trace to real source documents: that is where the practice advantage lives. The difference between getting that right and getting it wrong showed up in a matter that went out wrong several paragraphs ago.
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: Nasser Eledroos on Unsplash
