A barrister apologises in a murder case. A solicitor faces the regulator. A KC's unreported authority turns out not to exist. These are the stories Australian legal media has told about AI hallucination, and they are not wrong, the events happened, the professional consequences were real, and the reporting served a purpose.
The numbers, though, tell a different story. One worth sitting with.
Damien Charlotin's AI Hallucination Cases database, updated on 11 July 2026, now records 96 Australian decisions dealing with hallucinated material. Four decisions were handed down in the first two days of July alone, across the Supreme Court of Tasmania, the Victorian Court of Appeal, the Federal Circuit and Family Court, and the NSW Land and Environment Court. The pace has not slowed. Seventy-two of those 96 decisions involve self-represented litigants. Twenty-one involve lawyers.
Read that ratio again. Three-quarters of Australia's recorded hallucination decisions, the ones that consumed judicial time, required registry correspondence, sent judges back to check citations that did not exist, involved people who hold no practising certificate, face no disciplinary body, and in many cases had no idea the chatbot they were using could fabricate a law report with complete confidence.
The profession's response to AI hallucination has been framed, understandably, as a conduct story. Practitioners have duties. Those duties extend to the sources they cite. Court practice notes, from the Federal Court's GPN-AI to the guidance issued across state jurisdictions over the past 18 months, enumerate what verification looks like when AI has been used in drafting. Sign the document, own the citation. That is a reasonable principle and it is right that the profession has articulated it.
What the practice notes cannot do is reach the person who filed a submission last Tuesday using a free chatbot, who has never heard of GPN-AI, and who believed the case they cited existed because the tool told them, fluently and without caveat, that it did.
Seventy-two decisions says that person is the dominant source of the problem.
We should be careful about what this observation does and does not support. It does not suggest lawyers are handling AI well; 21 decisions involving practitioners is 21 too many, and every one of them represents a failure of verification that a professional should have caught. The conduct obligations are correct.
What the data does suggest is that the policy response has been calibrated almost entirely around the constituency that produces roughly a fifth of the recorded cases, while the constituency producing three-quarters of them has received little more than a general exhortation to be careful, if they received anything at all.
Courts have always had to manage self-represented litigants. The administrative burden of a litigant in person submitting documents of variable quality is not new. What AI has changed is the specific character of the error. A handwritten submission that misunderstands a legal principle is annoying. A submission that cites a precise, plausibly formatted case name, with a court, a year, and a paragraph number, all of which are invented, requires someone to chase the ghost. Registries check. Judges check. The other party's solicitor checks. Time disappears.
There is a compounding factor. Courts routinely extend procedural indulgence to self-represented litigants, reading documents charitably and allowing latitude that would not apply to represented parties. That indulgence, sensible in its own terms, means a hallucinated citation can survive further into the pipeline before anyone flags it. The registry may not query a plausible-looking citation at the filing stage. The problem surfaces in chambers, or at the hearing, which is a more expensive moment to discover it.
The four decisions from 1-2 July represent four separate occasions on which a court somewhere in Australia had to stop and verify that a cited authority was real. Multiply that across 96 decisions. The judicial cost accumulates in ways no individual filing makes visible.
The deeper problem is that the tool design makes fabrication invisible. General-purpose language models optimise for fluency. They produce text that sounds like a law report because they have seen enough law reports to reproduce the pattern. They have no mechanism to distinguish a real citation from a plausible one, and they have no architecture, no incentive, to stop when they are uncertain. The result looks exactly like research. For a self-represented litigant with no legal training, there is no tell.
This is a design observation. The moral framing doesn't follow from it. Which is also why the answer to the hallucination problem cannot be reduced to asking people to be more careful. Careful about what, exactly, if you cannot tell that something is wrong?
It is worth drawing a sharper line here between two things the industry often conflates. A model that was "trained on Australian law" has seen Australian cases in its training corpus. That means it can reproduce legal-sounding text in the right idiom. A system that retrieves from a closed, verified dataset of actual Australian legal sources does something structurally different: it can point to the document. The first is a fluency property; the second is an architecture property. The hallucination problem is a retrieval problem, not a fluency one, and training data is not a substitute for a closed corpus with live links to source material.
Any tool touching Australian law, for a practitioner or a litigant in person, a paid product or a free one, should make the underlying authority verifiable by default. One click to the source document. Citations as live links, not formatted decoration. A closed dataset of actual legal material, rather than a web crawl that includes commentary, summaries, and the accumulated errors of the internet alongside the real law.
That is infrastructure. The distinction matters because infrastructure is what you require of a tool; a differentiator is what you reward when it exceeds expectations. The 96-decision record from Charlotin's database makes the case for the former more clearly than any general argument could.
Habeas was built on this premise. The Search Engine runs across more than 300,000 Australian cases and pieces of legislation, in a closed dataset of verified Australian legal sources. Every result is traceable to the source document. The design assumption, from the beginning, has been that a citation without a real, checkable source behind it is a liability. Research that used to take a full morning can now be completed in minutes, and the outputs arrive with the citation intact, the mechanism that makes them usable rather than a formatting convention.
The 21 practitioner decisions are also worth a further observation: they represent only the failures that reached a judgment. The verification overhead, the time spent by lawyers and their staff checking whether AI-generated citations are real before anything reaches the court, never appears in any database. The 96 decisions are the visible cost; the invisible cost is every check performed on every AI output, real or fabricated. A closed corpus with traceable citations reduces that overhead because the practitioner can follow the link rather than run a separate search to confirm the source exists.
We are not claiming Habeas solves the self-represented litigant problem. Habeas is a practitioner product, priced and built for legal professionals. The person who files a hallucinated authority using a free chatbot is not our user, and we should say so plainly. The argument we are making is narrower: that traceable citation as default and a closed corpus as foundation are what any serious legal AI product should meet, and that the 96 decisions now on record in Australia are a concrete argument for why that standard exists.
The profession has been told, repeatedly and correctly, to verify before filing. What the database published this week suggests is that the larger question is whether the tools people use to engage with Australian law are built in a way that makes verification possible at all.
Charlotin's database will pass 100 Australian decisions before the year is out. That is a small number relative to the volume of court filings, and we should not overstate it. The courts are coping. But each decision in that database represents a moment where the administration of justice paused to check whether something was real. The pattern the numbers reveal, overwhelmingly litigants in person, across every jurisdiction, at an accelerating rate, suggests the pause is becoming structural.
The answer the profession has given is the right one for the constituency it can reach. The deeper answer is a design question, and it belongs to everyone who builds tools that let people search Australian law.
See what traceable citations look like in practice 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: Jaiju Jacob on Pexels
