You are across the other side's outline of argument, two days before a contested hearing. There is a Supreme Court decision cited in support of a proposition about the proper measure of damages. The style of the title is right, the year looks plausible, the court is the right one. You go to find it. The case does not exist.
This is no longer a hypothetical. It is a documented pattern in Australian courts, and it is arriving week by week.
Damien Charlotin's AI Hallucination Cases Database, which tracks court decisions where AI-generated fictional authorities were raised in proceedings, had logged 96 Australian entries by early July 2026. Four more decisions landed in the first two days of the month alone. The database is updated continuously. The pace has not slowed.
One detail the profession has largely passed over: the large majority of those 96 Australian entries involve self-represented litigants, people with no professional obligation to verify what they file, and no regulatory body to answer to if they get it wrong.
The profession's working model for AI hallucinations is substantially a compliance model. Do not file fabricated citations. Verify your own work. Understand the tools you are using. Maintain professional responsibility for everything that bears your signature. All of that is right. The Federal Court's GPN-AI, signed by Chief Justice Mortimer in April 2026, made the obligation specific and binding in federal proceedings: personally confirm that cited legal authorities exist and support the stated proposition.
A compliance model, though, is concerned with your output. The Charlotin data shifts the lens. If the large majority of hallucinated citations appearing in Australian proceedings come from self-represented parties, the practitioner's exposure runs in a direction that practice notes have not caught up with. You might verify every authority you file with care. Citations sitting in the other side's submissions were prepared with no such obligation and checked by no one who will face professional consequences for getting it wrong.
Courts in Tasmania, Victoria, and New South Wales have all seen this in the past few months. They do not share subject matter or judicial composition, but they share a mounting volume of decisions having to address invented authority produced by a party who asked an AI a legal question and filed what came back.
Consider the version of this problem courts can manage: a self-represented litigant who appears and relies on a fabricated citation gets corrected, and the decision records the correction. Costs fall where they fall. What that version obscures is the cost absorbed by the practitioner on the other side before anyone reaches the hearing.
Finding a fictional case in the opponent's submissions is not a trivial task if you are doing it methodically. A confident-looking citation, plausible judge, plausible date, plausible court, requires a real search to disprove. The task is harder than it might sound, because AI hallucinations are not always complete fabrications. A real case, with the correct court and a real judge's name, can appear in an AI-generated submission carrying a holding that was never in it, or with a proposition drawn from a different decision entirely. Finding the case in the reports does not end the check. You still need to confirm that the paragraphs the submission points to say what the submission says they do. If the opponent has cited several authorities across a complex outline of argument, verifying each of them is research time. That time is charged to someone. Where the opponent is a self-represented litigant who generated the submission in fifteen minutes using an AI tool they do not understand, the economics of the exercise sit entirely with you and your client.
Courts have issued guidance aimed at practitioners. Almost nothing in the current crop of AI practice notes addresses the volume coming from the other direction. The GPN-AI binds parties to proceedings, including self-represented parties, but its verification obligations are practically enforced against lawyers. A litigant-in-person who files a hallucinated citation faces the court's displeasure and perhaps an adverse costs order in the decision. They do not face a regulatory body. That structural asymmetry is not going away.
Citation-checking is quietly becoming a component of adversarial preparation in a way it has not been before. Checking your own work was always good practice. Checking the other side's work, at the level of verifying that cited authorities exist and say what the submissions claim, is now something a careful practitioner builds time for.
Courts deserve fair treatment here. The Federal Court's GPN-AI is a serious and technically literate document. NSW, Queensland, Victoria, and South Australia have all issued AI guidance of varying depth in the past eighteen months. The Fair Work Commission has done the same. Across Australian jurisdictions the direction is consistent: verify before you file, take professional responsibility for everything in your name, be able to account for AI use if asked.
What practice notes have not done, and perhaps cannot do, is address the inputs arriving from outside the profession. Self-represented litigants are a feature of every level of the court system, and they are increasingly using AI tools to prepare their documents. Most of those tools have no grounding in Australian primary law, no citation verification, and no mechanism for flagging when an output is confabulated. The litigant asks, the tool answers, the submission is filed.
Charlotin's database puts a number on that: 96 decisions and counting, concentrated heavily in a population that has no professional obligation to verify what they file.
Picture the specific moment. You receive the opponent's outline the evening before your preparation session needs to close. Five cases cited across three substantive propositions. The names look right, the courts look right, the years fall within a plausible range. You need to know, before you walk into that hearing, whether any of them are confabulations, and whether the holdings they have been pressed into supporting are actually in the judgments.
That is five searches against Australian primary law, conducted at speed, with the answers needing to be reliable enough to anchor your preparation. Each one requires confirming the case exists, that it comes from the court stated, and that the paragraphs the submissions point to say what the submissions say they do. Finding the case in the reports is only the beginning. Reading it against the claim is the work.
This is where a research engine designed for Australian primary law does something a general-purpose AI tool cannot. Habeas's Search Engine scans over 300,000 Australian cases and pieces of legislation, with results grounded in a closed dataset of legitimate Australian legal sources: verifiable and traceable to the source document, never hallucinated. A citation that was confabulated does not resolve. One that exists returns the judgment, the court, the date, and the relevant paragraphs. The check you might otherwise absorb across a prep evening resolves in seconds, not because the problem is simple, but because the corpus is the right one, and it is the only one that matters for this task.
We think the profession is still working out what this shift actually requires. Verification was, for most practitioners, a discipline applied to outgoing work. Charlotin's data suggests it has become incoming work as well. The structural problem, self-represented litigants with no regulatory exposure and no verification discipline generating plausible-looking submissions at scale, is not going to be resolved by tighter guidance aimed at lawyers. Those 96 decisions, four more in the first days of July 2026, growing continuously, are what a changed adversarial landscape looks like.
Litigation practice has absorbed new categories of preparation before. It will absorb this one. Absorbing it efficiently requires the right tools pointed at the right sources.
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: Pavel Danilyuk on Pexels
