When lawyers submit AI-fabricated citations.

The Dayal matter, Federal Circuit and Family Court of Australia.

On 19 July 2024 a solicitor identified in later reporting only as Mr Dayal appeared before Justice Amanda Humphreys in the Federal Circuit and Family Court of Australia. The hearing concerned an enforcement application in a family law dispute. The judge had requested a list of prior cases relevant to the application. The solicitor supplied it.

When Justice Humphreys and her associates attempted to verify the citations they were unable to confirm them. The matter returned to court and the solicitor admitted that the list had been generated using AI-assisted legal software and that he had not verified the information before submitting it to the bench. An August 2024 judgment from the Federal Circuit and Family Court recorded that the cases did not exist. The solicitor offered an unconditional apology, and stated that he had not intended to mislead the court, but the conduct was accepted as a breach of professional standards.

In October 2024 the judge referred the solicitor to the Victorian Legal Services Board and Commissioner. On 19 August 2025 the VLSB+C confirmed that the solicitor’s practising certificate had been varied. He is no longer entitled to practise as a principal. He may not handle trust money. He may not operate his own law practice. He may now practise only as an employee solicitor under strict supervision (ACS Information Age, August 2025, Lawyers Weekly, UK Law Gazette).

The VLSB+C statement on the variation was unambiguous: “The board’s regulatory action in this matter demonstrates our commitment to ensuring legal practitioners who choose to use AI in their legal practice do so in a responsible way that is consistent with their obligations.” It is the first professional sanction of its kind in Australia.

A King’s Counsel in a Victorian murder trial.

A few months after the Dayal matter became public, Victorian defence counsel Rishi Nathwani KC filed written submissions in a Supreme Court of Victoria murder trial involving a teenager charged with the killing of another child. The submissions contained fabricated quotes attributed to a speech to the state legislature and citations to Supreme Court decisions that did not exist. The errors were discovered by Justice James Elliott’s associates, who could not find the cases and requested that defence provide copies. The defence team admitted the citations “do not exist” and that the submission contained “fictitious quotes” (NBC News, August 2025, CBS News).

Nathwani, a senior barrister with the formal title of King’s Counsel, took “full responsibility” for the filing. Justice Elliott, for his part, made the point that has quietly become the core rule for AI use in Australian legal practice: “It is not acceptable for artificial intelligence to be used unless the product of that use is independently and thoroughly verified.” That sentence is, in practical terms, the operating standard.

The Law Council of Australia’s president at the time, Juliana Warner, acknowledged that fake citations are a “serious concern” while noting that an outright ban on AI in legal work would be impractical. Several other Australian lawyers have since been referred to state regulators over similar incidents. The broader pattern is now a named professional risk in state law society CPD material.

The original: Mata v. Avianca, S.D.N.Y.

The Australian cases sit inside a longer international pattern that the profession has been watching since mid-2023. In Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023), New York attorneys Steven A. Schwartz and Peter LoDuca filed a brief that relied on cases generated by ChatGPT. Among the fabricated authorities were Varghese v. China Southern Airlines and several similar invented decisions. Judge P. Kevin Castel, after an extended show-cause process, sanctioned the lawyers and their firm for using “fake judicial opinions with fake quotes and fake internal citations.” The sanctions, US$5,000 in total, were modest by legal-fee standards but the reputational damage was durable.

Mata v. Avianca mattered because it was the first well- documented case where a court issued a public, reasoned decision about AI-fabricated authority submitted in good faith. The failure mode was a lawyer who used a large language model to “research” a question without a product architecture that told him when the answer was not supported by real case law.

The UK approach, briefly.

In June 2025 the UK High Court, through Justice Victoria Sharp, warned that providing false AI-generated material as if it were genuine could be treated as contempt of court and in the most egregious cases as perverting the course of justice, which carries a potential maximum sentence of life imprisonment. Whether that warning is ever applied at its upper end is a separate question. The warning itself is a signal about direction of travel.

Australian regulators have so far taken the professional- conduct route rather than the contempt route. The VLSB+C action against Mr Dayal is the first concrete Australian sanction under that approach.

The common failure mode.

Across all three incidents (Dayal, Nathwani, Mata v. Avianca) the mechanism is the same. A capable language model is asked to “find” cases that support a proposition. The model produces output that looks like case citations, in the right style, with plausible party names, plausible case numbers, and plausible reasoning summaries. The output is not retrieved from a real legal corpus. It is generated. The lawyer, under time pressure or without the muscle memory for AI-specific verification, does not independently check the citations before submitting the work.

The failure is not that AI is used. It is that the product used was a general-purpose language assistant rather than a citation-verified legal research tool, and that the workflow did not force a verification step before the output left the lawyer’s desk. Every step of that failure is architecturally avoidable.

What an AI-safe research product looks like.

The architectural answer is not “be careful.” It is a product that makes the careful behaviour the path of least resistance. Four design choices, in combination, reliably block the failure mode above.

The first is retrieval before generation. An AI research tool that searches a real indexed corpus of Australian case law and legislation, and that only generates text grounded in retrieved passages, cannot fabricate a case out of whole cloth. If the corpus does not contain the case, the tool cannot cite it. This is the single most important architectural decision. It is not a feature we treat as optional.

The second is explicit flagging when a claim cannot be supported. When a user asks a question whose answer cannot be grounded in a verifiable authority, the product must say so, in plain language, on the same surface that shows the answer. In Matter Desk the exact phrase is “Authority not verified.” It is an output primitive, not a deprecation pattern. Seeing it is a normal part of a research session.

The third is citation hover cards that show the source text inline. Every citation in an AI research answer in Matter Desk is a hover interaction that reveals the actual passage from the judgment or the statute, with a link to the full source. A lawyer reviewing the answer can verify in seconds that the citation is real, that the paragraph says what the summary claims, and that the authority is applied correctly. Citations that cannot be hovered are, structurally, flagged as not-yet-verified.

The fourth is a human-review gate. Substantive AI output in Matter Desk does not go to a client, a court, or a partner without lawyer review of the underlying authorities and the reasoning. This is a product behaviour, not a disclaimer in the terms of service. The workflow is shaped to make review the default, not an optional extra step.

Any one of those four would have caught Mr Dayal’s list of non-existent cases, Mr Nathwani’s fictional quotes, or Mr Schwartz’s invented appellate decisions. In combination they remove the failure mode.

Why this is a product problem, not a lawyer problem.

It is tempting to frame the Dayal and Nathwani matters as failures of professional diligence. They are, in part. But the more useful frame is that general-purpose language assistants are the wrong tool for the job, and the professional community has not yet developed collective instincts for knowing which tool is which.

Mr Dayal used AI-assisted legal software. The product he used generated plausible citations and did not flag that the output was generated rather than retrieved. That product was, for this task, unsafe. Mr Nathwani is a King’s Counsel with decades of experience. If the failure mode can reach him it can reach anyone. The answer is that the failure mode should not exist in a properly built research product.

This is why the architectural decisions above are load-bearing. Regulatory pressure, ASCR 2015 obligations, and professional-conduct sanctions will continue to shape the space. But the products themselves are where the risk is actually resolved, and where firms can buy confidence rather than buy training modules.

ASCR 2015 and where the profession is going.

Australian Solicitors’ Conduct Rules 2015 place responsibility for the product of legal work squarely with the solicitor, regardless of how the work was produced. State law society AI guidance published during 2025 has reinforced the point: the duty to the court, to the client, and to opposing parties does not diminish because a tool was involved in drafting.

The direction of travel is clear. Firms will be expected to demonstrate that AI output was independently and thoroughly verified before it left the firm. That expectation is easier to meet when the product shows every citation with its source text, logs every AI action with user, matter, timestamp, and retained output, and requires human confirmation on anything that could touch a client. Matter Desk is designed to make that expectation the default behaviour of the workflow, not an extra burden on top of it.

Sources.