She is a General Counsel at a Series A startup, advising across privacy, employment, and contracts without a team beneath her. She runs a question about Australian privacy obligations through an AI tool she has been evaluating, one that describes itself, prominently and repeatedly, as trained on Australian law. The answer comes back confident. It covers the consent framework, the notification obligations, the thresholds for what constitutes personal information. It reads like the right answer.
Then she traces the citations. Most of them lead somewhere other than Australia.
The cases exist. The principles are real. The jurisdiction is wrong. The system has processed large volumes of English-language privacy law, GDPR commentary, American frameworks, British regulatory guidance, and assembled an answer that sounds Australian without being grounded in what the Privacy Act 1988 actually requires. The gap is material. A regulator examining advice built on that output would find it wanting.
She goes looking for a tool built differently. The question she takes with her is simpler than it sounds: what does "trained on Australian law" actually mean?
The phrase has become ambient. It appears in pitch decks, on product pages, in procurement conversations, repeated often enough that it functions as a credential rather than a claim. Standing in the gap between confident answer and unverifiable citation, she is the one who has to decide whether that credential holds.
She runs another question, this time about unfair dismissal thresholds under the Fair Work framework. Same pattern. Confident output. Opaque provenance. The system's confidence is uniform whether it is drawing on a binding Australian authority or legal commentary absorbed from somewhere else in the global web. She has no way to tell from the output which is which.
This is a corpus problem before it is anything else. The dominant approach to building AI systems treats scale as a substitute for specificity: train on very large, very broad material and assume that volume compensates for bounded depth. For legal work, that assumption collapses. Australian law is a distinct body of doctrine, structured by specific statutes and specific courts. Privacy obligations under the Privacy Act 1988 are not their European counterparts. The Fair Work framework does not map onto what an employment lawyer in another jurisdiction would expect. A system that has absorbed large volumes of English-language legal text will have absorbed patterns and language that sound Australian without being grounded in Australian authority, and the failure mode is producing something confidently wrong in a way the practitioner cannot detect until a regulator or adversary examines the advice.
Better prompting does not fix this. The drift is not a prompt failure; it is a training-data failure. When the corpus is the global web, the system returns Australian-sounding answers drawn from wherever the language pattern matches. She has now spent more cognitive energy verifying jurisdiction than doing the analysis the question actually required.
A closed dataset of legitimate Australian legal sources, curated and maintained rather than scraped from whatever was available, changes what the system can return. When every source in the training set is an actual Australian primary-law document, the answers stay grounded in the jurisdiction whose law is being applied. Citations trace back to actual decisions. The proposition attributed to a case is the one the court decided.
Habeas runs on a corpus of over 300,000 Australian cases and pieces of legislation, drawn from a closed dataset of legitimate Australian primary-law sources. When it cites an authority, that authority is traceable: a practitioner can follow the citation to the source, read the actual decision, and verify that the proposition it supports is the one attributed to it. The output does not wander into GDPR commentary or American case law and present it with the same confidence it would bring to a decision of the Full Federal Court.
The GC who works through that difference describes it plainly: "The Australian-law focus and the depth of nuance in the answers is a major differentiator. It materially changes my confidence and speed when forming legal views."
Confidence and speed grounded in the jurisdiction whose law is actually being applied — that is what the corpus claim is supposed to mean, and that is the standard against which it should be tested.
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.
