The first Australian lawyer publicly sanctioned for AI misuse handed a family-court judge a list of cases that did not exist

What happened

In October 2024, in the Federal Circuit and Family Court of Australia (Division 2), Judge Amanda Humphreys was hearing a property and parenting matter between a separating couple. The case is reported as Dayal [2024] FedCFamC2F 1166. The husband's solicitor, referred to in the regulator's subsequent statement only as Mr Dayal, taking the pseudonym from the family-law case anonymisation, had been asked by the judge to file a list of authorities supporting the relief his client was seeking.

Mr Dayal filed a list. The list contained case names, paragraph references and summaries of what each authority had decided. The judge sat down to read the cases. She could not find them. None of the cited authorities, on the citations given, existed.

When asked, Mr Dayal acknowledged that he had prepared the list using AI-based legal research software. He had not independently verified that the cases were real. He had assumed that the software's outputs were grounded in real caselaw. They were not. The summaries the software had generated were plausible. The citations were not.

Judge Humphreys' published reasons noted that the conduct had been stressful for the solicitor and that she did not regard it as likely to be repeated by him personally. She nevertheless referred the matter to the Victorian Legal Services Board (VLSB+C). Her stated reason was not punitive towards Mr Dayal individually. It was the broader regulatory issue of AI use in litigation, which the Board had not yet had a clean test case on which to act.

The Board took just under ten months. On 19 August 2025 it published a statement explaining the action it had taken on Mr Dayal's practising certificate. He was prohibited from practising as a principal lawyer. He was prohibited from receiving or handling trust money. He was required to undertake supervised legal practice for two years, with quarterly check-ins to confirm continued compliance. The Board described its action as a commitment to ensuring practitioners who choose to use AI in legal practice do so in a responsible way.

The action was widely reported as the first time an Australian legal regulator had taken public disciplinary action against a solicitor specifically for AI-related conduct. The professional duty applied, the basic obligation to verify what you put in front of a court, was older than the technology. But the action established the precedent. Other Australian regulators have since cited the Dayal matter when issuing AI-related conduct guidance.

What an auditable version would have shown

Two records that should have existed did not.

The first was a record at the software end that distinguished between retrieved authority (here is a case I found in the AustLII database, here is the URL where you can read it) and generated authority (here is a case-shaped block of text I produced from my model, which may or may not correspond to a real decision). The software produced summaries that read as if they had come from a database. They had not. The output did not flag itself as generated rather than retrieved. A signed record at the moment of generation, naming the source as "model output" rather than "database hit," would have given Mr Dayal a structured warning that the citations needed independent verification before submission.

The second was a record at the firm end. A signed log of the citation-verification step that should have run between the AI software's output and the document going to court. The verifier checks each citation against an authoritative source, AustLII, Jade.io, LexisNexis Australia, and produces a structured outcome for each: found, not found, ambiguous. The record is appended to the firm's matter file. The court can request it. The regulator can audit it. Mr Dayal could not produce such a record because the verification step had not been run. Once it was clear the citations did not exist, the absence of the verification record was, in itself, the evidence of the conduct.

Where the gap was

Mr Dayal is not a careless lawyer. The Board's published statement notes that his cooperation throughout the investigation was constructive and that the conduct does not appear to be characteristic of his practice generally. The fact that the Board nevertheless varied his practising certificate is the more important point. The gap Dayal exposed was not a lapse by an outlier. It was a workflow gap that exists in many Australian firms that have integrated AI legal-research tools into their practice without integrating a corresponding verification step.

Australian firms began adopting AI legal-research tools at scale through 2024 and 2025. Most adoption ran ahead of standing court orders mandating disclosure or verification of AI use. The default workflow in those firms was: the lawyer asks the tool, the tool returns text, the lawyer copies the text into the filing. No checkpoint required the citations to be confirmed in an authoritative source before the document was lodged. No signed record captured where the text had come from. The first time many firms had to think about either was when their own version of the Dayal moment arrived. The Board's decision moved the conversation from voluntary best practice to enforceable standard.

The Federal Court of Australia, the Supreme Court of NSW and the Supreme Court of Victoria have each since issued practice notes addressing the use of generative AI in court materials. The notes vary in their detail but converge on a common requirement: disclosure of any use of generative AI in preparing court materials, and an affirmation by the lawyer that the contents have been independently verified.

What governance should have looked like

The fix is the same fix as for Mata v. Avianca (HD-INC-002) and the Cohen matter (HD-INC-003): every citation in a court filing passes through a verifier before the document leaves the firm. The verifier confirms each citation against an authoritative source. The verifier's output is signed and retained. The filing pipeline refuses to release any document where the verification step has been skipped, or where any citation has failed verification without explicit override and a documented reason.

from headlights import CitationVerifier, ConductRecord, sign, chain
from datetime import datetime, timezone

verifier = CitationVerifier(
    authoritative_sources=["austlii", "jade", "lexisnexis-au", "thomson-reuters-au"],
    fail_on_unverified=True,
    require_ai_disclosure=True,    # Federal Court practice-note compliance
)

filing_citations = [
    # The citations Mr Dayal's AI tool had produced. None existed.
    "Re W and the Family Law Act [2019] FedCFamC2F 412",
    "Mason v Mason [2020] FamCAFC 88",
    "In the matter of Patel [2018] FedCFamC2F 705",
]

results = verifier.check_all(filing_citations)
# Each result records: citation text, sources_checked, found (bool),
# database_record_id (if found), timestamp.

record = ConductRecord(
    workflow="court_filing_pre_submission",
    matter_id="dayal-property-and-parenting",
    drafted_by="practitioner_id_MD",
    drafted_with_ai_assistance=True,
    ai_tool="ai_legal_research_software_v_n",
    ai_disclosure_filed_with_court=True,
    citations_submitted=filing_citations,
    verification_results=results,
    verification_passed=False,
    timestamp=datetime.now(timezone.utc),
    previous_record_hash=last_record.hash(),
)

signed = sign(record, key=firm_private_key)
chain.append(signed)

if not record.verification_passed:
    raise FilingBlocked(
        "Three citations could not be verified against any Australian "
        "authoritative source. Review required before submission."
    )

Citation verification at the filing gate. Every Australian firm using AI legal-research tools needs a verifier in the pipeline. AustLII is free, Jade.io is widely licensed, the major commercial databases all have APIs. The verifier runs in seconds. Fake citations are caught before they reach the court.

AI disclosure baked into the matter file. The practice notes now in force in the Federal Court, the Supreme Court of NSW and the Supreme Court of Victoria each require disclosure of generative-AI use in preparing court materials. A conduct record that captures the AI tool used, the prompts and the verification step makes the disclosure mechanical rather than a memo written in a hurry on the morning of the hearing.

An audit trail the regulator can read. The Board's investigation of Mr Dayal turned on a reconstruction from memory and emails. If the firm had kept signed records of every AI-assisted draft, every verification check and every citation submitted, the regulator's inquiry would have been one document long: either the records show verification ran or they show it did not. The regulator's decision becomes evidence-driven rather than admission-driven. For lawyers acting in good faith, the records are protective: they prove the verification step happened.

Dayal is the first Australian sanction on these facts. It will not be the last. The firms that build verification into their workflow now, alongside the AI tools they are already using, are the ones that will not be reading their own version of the Board's statement in twelve months.