What happened
In early 2025, the applicants in Murray on behalf of the Wamba Wemba Native Title Claim Group v State of Victoria sought to amend their native title determination application. The applicants were represented by Massar Briggs Law, a small Melbourne firm. The amended application took the form of an Amended Form 1, the standard Federal Court document by which a claim group sets out who they are, the country they claim, and the evidentiary basis for the claim. Each substantive paragraph in the pleading is typically supported by footnotes citing anthropological and historical reports, ethnographic studies, prior native title determinations and the rules of evidence those determinations rely on. The footnotes are not ornament. The respondents, the State of Victoria and several other parties, have to be able to verify every reference, because the cited material is part of how the Court tests whether the claim group’s connection to country is made out.
The task of preparing the footnotes for the Amended Form 1 (which were also incorporated into a summary document filed in the proceeding) was assigned to a junior solicitor at the firm. She was working from home. The physical copies of many of the documents to be cited, the anthropological reports, the older Federal Court judgments, the secondary sources, were in the firm’s Melbourne office. The electronic copies were on the firm’s internal network. From home, she had access to neither.
To produce the citations she turned to an online tool rather than the firm’s own library. The judgment does not pin down exactly which tool: Justice Bernard Murphy found that the false citations “likely arose from the use of generative AI, possibly through Google Scholar.” Whatever the source, the citations that came back were a mix of references to documents that did not exist and references to documents that existed but had been wrongly cited. Murphy described the failure as one of failing to check and verify the output of a search tool, contributed to by the inexperience of the junior solicitor and the firm’s failure to have systems in place to ensure her work was appropriately supervised and checked.
The amended application was filed. First Nations Legal and Research Services (FNLRS), the Victorian native title service body, was tasked with locating the documents referenced in the footnotes. FNLRS could not locate most of the cited material. Some citations led nowhere. Others led to documents that, on inspection, said something other than what the footnote claimed. The discrepancy was raised with the Court.
Massar Briggs Law did not contest the account of what had gone wrong. The principal solicitor, Jason Briggs, filed an affidavit explaining the circumstances. The Court’s task became the costs question: who should bear the cost of the respondents’ work in unpicking the fabricated and incorrect footnotes?
In a decision delivered on 22 April 2025 as Murray on behalf of the Wamba Wemba Native Title Claim Group v State of Victoria [2025] FCA 731, Justice Murphy ordered that the applicants’ solicitors (Massar Briggs Law) pay the respondents’ costs on an indemnity basis. An indemnity costs order, for non-lawyer readers, is an elevated costs award reserved for conduct the Court considers outside the ordinary run of litigation conduct, where the losing party is required to cover almost all of the winning party’s actual legal costs, rather than the partial recovery that ordinary costs orders provide. Importantly, the order was made against the firm itself, not the client. The judgment was specific that the conduct was not the choice of the Wamba Wemba claim group and the costs should not be borne by them. His Honour did not consider it appropriate to refer the solicitors’ conduct to the Victorian Legal Services Board.
The judgment landed in a period when Australian courts had been visibly tightening their approach to AI use in legal practice. The Federal Court issued a formal Notice to the Profession on AI use on 29 April 2025, a week after Justice Murphy’s decision. The Supreme Court of New South Wales and the Supreme Court of Victoria had each, by mid-2025, issued practice notes addressing generative AI in court materials. The Victorian Legal Services Board had already acted on its first AI-related matter, the Dayal sanction recorded in HD-INC-011. The Nathwani matter (HD-INC-012) had moved through the Supreme Court of Victoria the prior year. The Massar Briggs decision was the first time an Australian superior court had translated the practice-note language into a costs order, and the first time AI use in a native title claim had been considered by the Federal Court at all.
What an auditable version would have shown
Two records that should have existed did not.
The first is a record at the firm end of what was done with which tool, by whom, and when. The junior solicitor was, the judgment records, working without access to the firm’s physical or electronic source library. A signed record at the moment of citation drafting, naming the tool used (Google Scholar) and the absence of an independent verification step against the firm’s source materials, would have surfaced the gap at the moment it opened. The firm’s pre-filing review would have been able to see, before the document went to court, that one of the steps the firm’s own procedure assumes (verification against the source) had been skipped. The supervising solicitor would have had a concrete, structured signal rather than an implicit assumption that the work had been done correctly.
The second is a record at the verification step that should have run between the junior’s draft and the firm’s signature on the filing. A verifier takes each citation, queries an authoritative source (AustLII for cases, Trove or a university library catalogue for older anthropological reports, the National Native Title Tribunal database for prior determinations), and produces a structured outcome per citation: found, not found, found but the cited proposition does not match the source. The verifier’s output is the supervising solicitor’s pre-flight check. The output is itself a signed record, retained on the matter file, available to be produced to the court or the regulator if the question is ever asked. In the Wamba Wemba matter there was no such record because no such verifier was run. The first time the citations were checked against authoritative sources was when the State did it, on the State’s time, at the State’s expense, which is the cost the Court has now made the firm indemnify.
Where the gap was
The gap was supervision, not technology.
Justice Murphy’s language is careful on this point. The conduct is described as a failure to check the output of a search tool, and the failure is attributed in two parts. The junior solicitor was inexperienced. The firm did not have systems in place to ensure her work was supervised and checked. Both are needed to produce the outcome. A more experienced solicitor would, on most accounts, have noticed that something was wrong with the citations and gone back to the office to confirm them against the source library. A firm with a structured pre-filing verification step would have caught the citations regardless of who drafted them. Neither was present.
The remote-work element is worth dwelling on. The judgment does not blame remote work for the failure. The relevant fact is structural: the firm’s source library was not accessible from where the work was being done. The junior solicitor’s choice to use Google Scholar was a workaround for an access problem the firm had not anticipated when it allowed the work to be done from home. Firms that have moved to hybrid work, which is most Australian law firms now, need their pre-filing controls to be the same controls regardless of where the work is being done. A verifier that runs over every filing before it is lodged does not care where the drafter was sitting. A supervising solicitor who only checks for citation accuracy when the drafter is visibly in the office is operating a control that hybrid work has quietly broken.
There is a broader lesson the Federal Court has been signalling for some time. The duty to verify what goes in front of a court does not transfer to a tool. Mata v. Avianca (HD-INC-002) established the principle in the United States. Cohen (HD-INC-003) confirmed it. Dayal (HD-INC-011) was the first Australian regulator action. Nathwani (HD-INC-012) was the senior counsel version of the same conduct in the Supreme Court of Victoria. Massar Briggs is the version where the financial consequence sits squarely on the firm, on an indemnity basis, because the firm’s supervision systems were not in place. The pattern across all five matters is the same. The conduct is older than the technology. AI just makes it cheap to produce a large volume of work that looks correct and is not.
What governance should have looked like
For a firm using AI-assisted research tools, the governance question is what runs before any document leaves the firm, regardless of who drafted it or where they were sitting when they did so.
A verifier that runs against authoritative sources is the first move. AustLII for cases, the National Native Title Tribunal for prior determinations, Trove or the firm’s own library catalogue for the historical and anthropological materials specific to native title work. The verifier is not optional and it does not produce a recommendation. It produces a structured outcome per citation and a pass or fail at the filing level. The firm’s matter-management system refuses to release any document where the verification step has been skipped or has failed without an explicit, recorded override and a documented reason.
A drafter-tool-disclosure field on the verification record captures what tool was used, where the drafter was working, and whether the firm’s source library was accessible from that location at the time. The disclosure is the supervisor’s signal. A draft produced from home, using Google Scholar, against an inaccessible firm library is exactly the configuration that needs the verifier’s strongest check before signature. The record makes the configuration visible rather than implicit.
A retained signed log of every verification outcome is the firm’s evidence if the question is ever asked. The court can produce it on its own initiative. The regulator can call for it. The firm’s professional indemnity insurer can review it during renewal. For lawyers acting in good faith, the record is protective: it proves the verification step happened and the citations were what the drafter and the supervisor believed them to be. For matters where the verification did not happen, the absence of the record is itself the evidence, which is the position Massar Briggs Law found itself in once FNLRS started trying to locate the cited documents.
The Federal Court has, with Massar Briggs, set the price of skipping the verification step. It is the other side’s costs, on an indemnity basis, paid by the firm. Practice notes, conduct rules and CPD modules across the Australian profession have moved in the same direction through 2025 and 2026. The firms that build verification into the filing pipeline now, alongside the AI-assisted tools their juniors are already using, are the firms whose next pre-filing supervision conversation does not have to be conducted in front of a Federal Court judge.
The reference implementation of CitationVerifier, VerificationGate, and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. Anyone can read every line and verify the signatures. The repository is public now.
Sources
- Murray on behalf of the Wamba Wemba Native Title Claim Group v State of Victoria [2025] FCA 731 (AustLII)
- Solicitor to Pay Costs for Submissions Affected by the Use of AI (Hearsay)
- Melbourne law firm caught using fake AI citations (Information Age, ACS)
- Melbourne firm submits AI-fabricated citations (Lawyers Weekly)
- Australian Federal Court orders costs against law firm over use of Generative AI (Mills Oakley)
- Case alert: A warning on the use of Artificial Intelligence in native title claims (MPS Law)