Generative AI and the Australian practitioner: an update on court expectations, professional risk and what your firm needs to do
This is an update to the author’s earlier article published in this journal here. That article introduced professional guidance emerging around generative AI in litigation. This article reports significant developments since then.
The regulatory moment has arrived
Twelve months ago, Australian lawyers were asking whether they needed an AI policy. That question has been clearly answered. On 16 April 2026, Chief Justice Mortimer issued the Federal Court’s General Practice Note on Generative AI (GPN-AI), applying to everyone in proceedings: lawyers, litigants, witnesses and third parties. The NSW Supreme Court’s Practice Note SC Gen 23 has been in force since 3 February 2025. The Supreme Court of Victoria has issued its own Guidelines for Responsible Use of Artificial Intelligence in Litigation.
Across the major Australian courts, the message is consistent: generative AI may be used, subject to conditions, but it does not in any way reduce professional responsibility. The question courts and regulators are now asking is not whether AI was used. It is whether that use can be explained and justified if it is ever scrutinised.
What the courts now require
GPN-AI imposes a personal verification obligation. Where a lawyer uses generative AI to prepare submissions, chronologies or lists of authorities, the responsible person must confirm that cited authorities exist and support the stated proposition, that evidence cited is in the materials and reasonably likely to be admissible, and that factual statements are based on what the party reasonably considers can be proved.
SC Gen 23 goes further still. AI must not be used to generate the content of affidavits, witness statements or character references, although this does not prohibit the use of Gen AI for work that is merely preparatory to the drafting of the affidavit or other document setting out the witness’ evidence and/or opinion. Citations in written submissions must be independently verified, and not by AI. Suppressed material, material subject to the Harman undertaking, and subpoena material must not be entered into any AI tool unless strict platform conditions are met. Expert reports attract additional requirements: AI must not be used to draft report content without the Court’s prior leave, and if leave is granted the expert must disclose which parts were AI-assisted, identify the program and version, and retain records of prompts and outputs.
The scale of the problem in Australia
One French legal academic, Damien Charlotin, maintains what appears to be the most systematic publicly available tracker of AI hallucination incidents in courts globally, drawing from published judgments. His Australian dataset, as at May 2026, records approximately 65 reported incidents across every level of the court system, from the Federal Court and state Supreme Courts to the Fair Work Commission, the Administrative Review Tribunal and various state tribunals. Where sanctions are recorded these include bar referrals, personal costs orders against lawyers, adverse costs orders, reprimands and warnings. Those figures should be read with care: the dataset captures only incidents that were detected, raised in proceedings, and recorded in published reasons. There is no reliable way to know how many AI-generated errors have been filed in Australian proceedings without being challenged or noticed. On any reasonable view, however, the published cases likely reflect a fraction of what is actually occurring.
What the data does confirm is the breadth of the problem. Lawyers, not just self-represented parties, appear as the source in a material number of entries. Many jurisdictions are represented. The dominant failure type is consistent: fabricated case citations that look authoritative, are formatted correctly, but do not exist.
The privilege and confidentiality risk
One of the more recent consequential developments came from the UK Upper Tribunal. In Munir v Secretary of State for the Home Department [2026] UKUT 81 (IAC), the Tribunal found that uploading confidential documents to a publicly available AI tool such as ChatGPT places that information in the public domain, breaches client confidentiality and permanently waives legal professional privilege. While the decision is not binding in Australia, it reflects precisely the concern GPN-AI addresses at paragraphs 4.13 to 4.15.
The Tribunal drew a clear line between publicly available open-source tools and closed-source enterprise platforms. For Australian practitioners, the practical implication is clear: if your firm does not know whether the AI tools in use are open or closed source, you don’t have the information you need to protect your clients’ privilege. Of course, privilege, once waived, cannot be recovered.
When policies are not followed: real consequences
In August 2025, the Victorian Legal Services Board took action against a Victorian solicitor following findings in Dayal [2024] FedCFamC2F 1166. The solicitor had filed AI-generated content containing inaccurate citations and summaries. The outcome was significant: loss of the right to practise as a principal, loss of authorisation to handle trust money, prohibition on operating a law practice, a requirement to practise only as an employee solicitor, and two years of supervised practice with quarterly reporting obligations. That is not a costs order; it is effectively the end of a practice.
An international example is instructive for a different reason. In April 2026, Sullivan & Cromwell, one of the world’s most prestigious law firms, apologised to a US federal bankruptcy judge after filing court documents containing numerous AI-generated hallucinations, including fictitious case citations. The firm had comprehensive AI policies, mandatory training modules, and internal guidance requiring lawyers to verify everything. Those policies apparently weren’t followed, and the errors were identified by opposing counsel, not the firm. Within days, the story had appeared in numerous media reports around the world.
The reputational damage in that matter was swift and public. A single filing error, amplified across global legal media, attached itself to a firm’s name in a way that is difficult to undo. For Australian firms, large and small, the risk is the same: a potential disciplinary referral, an adverse costs order, or a news article about a hallucinated citation travels faster and further than any correction.
Culture is the harder problem
The Dayal disciplinary outcome and the Sullivan & Cromwell matter share a common thread. In both cases, the failure was not a lack of guidance, it was a failure in application.
This is where the junior solicitor question becomes important. Many lawyers entering the profession now have used generative AI throughout their studies and in previous roles. For them, prompting an AI tool is intuitive and familiar. That familiarity is an asset, but it carries a specific risk: comfort with a tool can reduce the critical distance needed to verify its output, particularly at a time when maturity of legal knowledge and judgment is yet to be developed. A junior lawyer who has used ChatGPT fluently for three years may be precisely the person most likely to trust a response that looks authoritative but is wrong. Whether or not that is detected comes down to appropriate supervision or technical barriers, both of which are ultimately a principal’s responsibility.
For principals and practice managers, the culture question is this: when a junior lawyer is under time pressure preparing a filing, does the environment they work in make it likely they will apply the verification standard the policy requires? Or does the pressure of the moment, combined with confidence in a familiar tool, make it likely they will not? That is not a policy question. It is a question about supervision and the signals that partners and principals send about what matters.
What to do now
Practitioners and practice managers should treat the following as a minimum baseline.
Confirm that your firm’s AI policy addresses the specific requirements of GPN-AI and SC Gen 23, including disclosure obligations, verification standards and the treatment of confidential and privileged material. The policy must distinguish clearly between public AI tools and approved enterprise platforms.
Review whether training has moved from awareness to application. Can practitioners in your firm demonstrate how verification is done in practice? Knowing that verification is required is not the same as doing it consistently under pressure.
Pay particular attention to junior solicitors and paralegals. Their comfort with AI tools is valuable, but it needs to be paired with a clear understanding of the verification obligations that apply when those tools are used in a professional context.
Consider the reputational dimension. A hallucination that reaches a filed document is not just a professional risk. It is a public one, and the speed at which such incidents become known is now well demonstrated.
Discuss AI governance at principal or partnership level as a matter of serious professional risk, not only as a compliance exercise.
Australian courts have done the work of setting the standard, and regulators have demonstrated they will act on it. What remains is the less dramatic but more demanding task of embedding that standard into how your firm actually operates, day to day, under pressure. That is a leadership question as much as a legal one.