In July 2026, a man named Ba brought an unlawful termination claim against Sterling Parts Australia in the Federal Circuit and Family Court. He represented himself. Along the way, he consulted an AI chatbot as his legal advisor, trusting it to guide him through procedure and tell him what documents he needed. Some of those documents did not exist. His evidence list referred to materials that had never existed at all.
Judge Symons was not impressed. The judgment described Ba's reliance as "blind faith" and found the conduct came close to contempt of court. The claim failed.
The case is already circulating as a cautionary tale about AI hallucination. That framing is understandable and almost entirely wrong.
The fabricated documents made the error visible. The error itself was already complete before a single fake citation appeared on Ba's evidence list. He had delegated legal judgment to a system with no practising certificate, no duty to the court, and no duty to act in his interests. Judge Symons drew all three lines explicitly. The AI was not a lawyer. Treating it as one was the category error, and the hallucinated evidence list was the proof of it.
This distinction carries weight beyond the sympathetic facts of a self-represented litigant in an employment dispute. General-purpose AI tools are not designed to hold that line themselves. Sounding authoritative is, in a meaningful sense, the product. A large language model optimised for natural, confident conversation has every structural incentive to produce an answer that reads as though it comes from someone who knows. The fluency is a feature. It serves the consumer use case well. It serves the litigant before the FCFCOA rather less well, because federal proceedings require something the fluency cannot supply: actual grounding in real sources, and an honest acknowledgment of the limits of the tool producing them.
Ba's chatbot apparently offered neither. When it told him certain documents existed, nothing pointed to a primary source. No citation to trace, no instrument the output pointed back to. The confidence was self-referential, which is precisely the failure mode that hallucination in legal contexts exploits. What made Ba's situation recoverable in principle, at any stage before filing, was also what the design of the tool withheld from him: a link back to a real source he could check. Verification requires something to verify against. The Federal Court's GPN-AI practice note, issued just months before this judgment, anticipated exactly this gap. It requires lawyers appearing before the court to personally confirm that cited authorities exist and support the stated proposition. That obligation assumes a person with professional standing is somewhere in the chain. When there is no such person, the chain itself is missing.
The consumer AI industry has largely left the underlying question unanswered. Disclaimers exist in fine print. The interface presents something different: a conversational partner that answers questions, sounds assured, and rarely says "this is beyond what I should be advising on." For someone without legal training, the gap between that experience and competent legal advice is close to invisible, and nothing in the design of most general-purpose tools works to make it visible. "Ask me anything" is the value proposition. It does not come with a carve-out for matters before the federal courts.
There is a harder version of the same question that Australian courts have not yet fully engaged with: does an AI tool that holds itself out as providing legal advice engage the prohibition on unauthorised legal practice? The definition of "legal practice" varies across state and territory legislation, but the core is consistent. Advising a person on their legal rights and obligations, for a matter in which they have an interest, has historically required a practising certificate. Ba's chatbot was doing something very close to that. The question of whether the prohibition extends to software is unresolved, but Ba v Sterling Parts makes it harder to defer.
Symons's three lines, no practising certificate, no duty to the court, no professional obligation to the person relying on it, are structural features of what these tools are, and what they are not. Better training data will not resolve them. They are not limitations that scale away.
The duty to the court has a specific procedural shape that is easy to abstract away from. When Ba filed an evidence list pointing to documents that did not exist, the court's process was concretely compromised. Other parties had to engage with it. The judge had to address it. Court time was consumed by the consequences of an output no one in the room had any obligation to correct before it was filed. The harm was procedural and immediate. Unlawful dismissal claims under the Fair Work Act attract self-represented litigants more than most areas of federal practice. The jurisdictional threshold is low, the filing process is accessible, and the subject matter is one where a person affected is likely to feel, with some justification, that they understand the facts well enough to run their own case. What they often lack is the procedural framework, and that is precisely the void a conversational AI will fill with confidence whether or not that confidence has any foundation. Practitioners on the other side of these matters have always had to manage the complications of unrepresented opponents. AI-advised self-represented litigants present a version of those complications with a new wrinkle: the person may have filed something built on sources that do not exist, without any awareness of what they have done.
The practical question the judgment presses on is not about verification in the abstract. It is about category selection. General-purpose AI performs many tasks well. Correctly identifying what evidence a court requires, understanding the procedural framework of a specific jurisdiction, and assessing the weight of authority in a given statutory context are tasks that require a different kind of grounding, because the stakes of getting them wrong are not corrected by the next conversation. Ba's evidence list did not get a second chance.
Symons's three lines are also a reasonable test for any AI legal tool worth adopting. Does it acknowledge that it holds no practising certificate, owes no duty to the court, and cannot substitute for the professional judgment of someone who does? We have been direct about this at Habeas, perhaps more than is commercially comfortable. Habeas is a research tool: it searches over 300,000 Australian cases and pieces of legislation, with results grounded in a closed dataset of verified Australian legal sources, every citation traceable to the primary document. Foundational research that once consumed a full morning can now be completed in minutes, because the underlying sources are real and the thread back to them is there to follow. That is what makes speed safe. Speed without traceability is a faster route to filing something wrong, as Ba's claim illustrates well enough.
Habeas does not advise clients, assess litigation risk, or substitute for the judgment of a lawyer who understands the full picture of a matter. This is a description of the product, not a compliance caveat. A tool that assists research and a tool that plays lawyer are different products with different failure modes. The profession is better placed drawing that line at the point of tool selection than at the end of a failed claim. Sign up 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: High Court of Australia
