A recent decision of Deputy President Searle in the NSW Personal Injury Commission (PIC) has shone a spotlight on one of the most pressing issues facing courts and tribunals in 2026: the undisclosed use of generative artificial intelligence (AI) in legal proceedings. While the appeal in Sentinel Community Services Pty Ltd v Clark Chesaites was dismissed on jurisdictional grounds, the decision contained a pointed examination of AI-generated submissions that will be of immediate relevance to all practitioners and self-represented parties appearing before the Commission and, more broadly, before any court or tribunal in Australia.
Background
The proceedings arose from a workers compensation claim made by Ms Clark Chesaites against her employer, Sentinel Community Services Pty Ltd (Sentinel). The worker alleged she had suffered a psychological injury as a result of bullying and harassment in the workplace. Sentinel’s insurer disputed liability on 24 July 2025, relying on ss 11A(1), 33, 59 and 60 of the Workers Compensation Act 1987 (NSW). Following two unsuccessful internal reviews, the worker filed an Application to Resolve a Dispute in the Personal Injury Commission of New South Wales (the Commission) in December 2025.
At a directions hearing on 11 March 2026, Member Whiffin permitted Sentinel to participate in the proceedings in its own right, separately from the insurer, and set a timetable for it to file its own Reply and supporting materials. At a further directions hearing on 13 April 2026, the worker and the insurer advised they had resolved their dispute. Consent orders were made and the hearing was vacated.
Shortly afterwards, Sentinel’s director, Ms Joanne Jones, assuming conduct of the matter in a self-represented capacity after parting ways with the company’s solicitors, filed an appeal contesting those orders in Sentinel’s name. Although the insurer had conducted the original defence of the claim on behalf of Sentinel, it did not support the appeal and expressly confirmed that it had not authorised its commencement or the filing of the appeal application.
The insurer accordingly appeared as a respondent to the appeal and opposed the application.
The appeal raised a number of substantive issues, including the employer’s standing to bring an appeal independently of its insurer and whether the Member’s procedural steps constituted an appealable ‘decision’ within the meaning of s 352 of the Workplace Injury Management and Workers Compensation Act 1998 (NSW). Deputy President Searle dismissed the application on those jurisdictional grounds. However, the decision is most significant for its consideration of the undisclosed use of generative artificial intelligence (AI) in the preparation of appeal material and its application of the Commission’s newly introduced AI rules.
The Errors
The insurer identified a series of errors in Sentinel’s submissions that were either plainly wrong or wholly inapplicable. These included reliance on:
- Incorrect citations and multiple incorrect statutory references;
- Reliance on s 117 of the 1998 Act, a provision which dealt with the admissibility of statements by injured workers, which Sentinel incorrectly cited as conferring standing on the employer to maintain its appeal;
- Reliance on s 13 of the Insurance Contracts Act 1984 (Cth), a federal statute which does not apply to workers compensation insurance contracts;
- Misapplication of r 63 of the Personal Injury Commission Rules 2021, which concerns the service of documents on joined parties, which Sentinel argued supported its standing to appeal; and
- An incorrect citation for Barden v Walgett Shire Council which was referred to as a 2022 decision when it was in fact decided in 2023.
Importantly, not every error was substantive. The legal principle Ms Jones was attempting to advance in Barden was “more or less correct”. The problem was that AI had simply given her the incorrect year. This illustrates a common feature of generative AI and demonstrates that the underlying legal concept may be broadly correct while the supporting citation is misdated or unreliable.
When confronted with these issues during the appeal hearing, Ms Jones conceded she had used generative AI in the preparation of the appeal application materials and in the submissions made before the Deputy President Searle. However, she had not disclosed this use at any point. It was, to use the Deputy President’s phrase, “undisclosed.”
The Commission’s AI Governance Framework
Deputy President Searle set out the Commission’s current AI governance regime in detail. Since 1 January 2026, rules 33A, 33B, 133B and 133C of the Personal Injury Commission Rules 2021 and Procedural Direction PIC13 (Use of Generative AI) had been in force. The core obligations were:
- Verification: Rule 133C requires a person who used generative AI to verify that all citations, legal authorities and case law referred to in the submissions exist and are accurate and relevant to the proceedings;
- Human review is mandatory: Procedural Direction PIC13 expressly provided the required verification must not itself be performed using a generative AI tool. The responsibility for checking authorities remains with the person preparing the submissions.
- Continuing obligations: The use of AI does not qualify or absolve a party of their professional or ethical obligations to the Commission.
His Honour also surveyed the broader landscape of AI governance in Australian courts and tribunals. The Chief Justice of the Supreme Court of New South Wales has expressed the need for a cautious and rigorous approach to permitting the use of generative AI in legal proceedings. The NSW Supreme Court has issued Practice Note SC GEN 23 on the use of generative AI. Each imposes duties of verification on responsible persons, practitioners and self-represented parties alike, and each contemplates that non-compliance by a legal practitioner may result in a referral to the Legal Services Commissioner or a personal costs order. The framework is coherent and proportionate. It does not prohibit the use of generative AI. It requires that those who use it take responsibility for what it produces and that they check the work before it goes before a tribunal.
The Obligations Apply to Everyone, Including Self-Represented Parties
A particularly significant aspect of the Sentinel decision is the Deputy President’s confirmation that the obligation to comply with the Commission’s AI rules applies equally to self-represented parties. Ms Jones sought, in effect, to rely on her status as a non-lawyer to mitigate the consequences of her non-compliance. His Honour did not accept that framing.
There was an additional point the Deputy President noted: Ms Jones had claimed she could not afford legal representation, but the evidence suggested otherwise. Two different solicitors had, at different times, acted for Sentinel in the proceedings. The earlier solicitor had apparently withdrawn after being persuaded by the insurer’s legal representative that Sentinel did not have standing to bring the appeal in the first place. As the Deputy President observed, this “suggests that Sentinel did have legal advice but may not have liked it or wished to follow it.” Disregarding legal advice and instead relying on AI is not, as the decision highlights, a sound appellate strategy.
The Commission did not take the matter further on this occasion, noting that “the deployment of undisclosed generative AI has not affected or perverted the result of this application and for this Ms Jones and Sentinel should be grateful” at [112]. However, the decision was described as “a timely warning to anyone appearing before the Commission of the necessity to comply with the rules and Procedural Directions pertaining to AI disclosure” at [112].
Why AI Hallucinations Are a Particular Risk in Legal Proceedings
The errors in Sentinel’s materials are a vivid illustration of AI ‘hallucination’ where generative AI produces information that appears convincing but is factually incorrect. In everyday contexts, this may be a minor inconvenience, however, in legal proceedings, the consequences are potentially serious.
As Deputy President Searle observed, the use of inaccurate AI-generated material can affect the integrity of the proceedings, the results of decisions and the confidence in the institution of the Commission by the public, the legal profession and appellate courts. Given the Commission’s important role and the large quantity of matters it determines each year, these risks are not insignificant.
His Honour identified three principal harms arising from the undisclosed use of incorrect AI-generated material. First, it undermines procedural fairness to other parties, who are entitled to respond to submissions that accurately state the law. A party who cites a non-existent case or misattributes a statutory provision forces the other side and the tribunal to expend resources identifying and correcting the error.
Second, it impairs the integrity of Commission decisions. A tribunal that proceeds based on incorrect legal propositions, however confidently stated, risks making decisions that cannot be sustained on appeal or that do not reflect the actual state of the law.
Third, it is antithetical to the objects of the Personal Injury Commission Act 2020 (NSW), in particular the obligations under ss 42 and 43 to act in good conscience, equity and with attention to the substantial merits of the case. Submissions affected by AI-generated errors do not advance the merits of any party’s case; they obstruct access to them.
Practical Guidance
For practitioners and parties appearing before the Commission (and courts generally), the following steps are essential where generative AI has been used in preparing written materials:
- Be transparent about the use of generative AI. The Commission’s AI framework contemplates disclosure where AI has been used in preparing written submissions. Failure to disclose may become an issue if errors are later identified.
- Verify citations independently. Procedural Direction PIC13 is explicit that the verification of citations, authorities, and legislative references must not be carried out by a generative AI tool. That means a human must open the case, read the section, and confirm what they find.
- AI does not confer standing, and it cannot select the right provision. Generative AI can produce convincing but incorrect explanations of legal principles, statutory provisions and case authorities. The Commission applies the law as enacted and interpreted by the courts, not as described by an AI chatbot.
- Self-representation does not reduce compliance obligations. The rules apply to all parties and all representatives, regardless of whether they hold a practising certificate.
- Treat AI output as a starting point. Generative AI can usefully assist with identifying issues, structuring arguments or summarising background material. However, it is not a substitute for legal research, and responsibility for the accuracy of any material filed with the Commission remains with the person who signs and submits it.
Conclusion
The Sentinel decision is a timely reminder that the adoption of generative AI in legal practice is not without risk and that courts and tribunals are now equipped with, and prepared to enforce, specific rules to manage that risk. The obligations it imposes, disclosure, independent human verification, and continuing professional responsibility, are not onerous, but they require conscious attention each time AI tools are used.
The errors identified in Sentinel’s appeal materials are a vivid illustration of what can happen when AI output is filed without adequate review. For practitioners, the stakes extend to disciplinary consequences and personal costs orders. For self-represented parties, the cost is the credibility and effectiveness of their case.
The Commission has issued its warning. Parties and practitioners appearing before the Commission, and before courts and tribunals across Australia, would be well advised to take it seriously.
This article may provide CPD/CLE/CIP points through your relevant industry organisation.
The material contained in this publication is in the nature of general comment only, and neither purports nor is intended to be advice on any particular matter. No reader should act on the basis of any matter contained in this publication without considering, and if necessary, taking appropriate professional advice upon their own particular circumstances.