1. Introduction
Publicly available generative artificial intelligence (“GenAI”) platforms are now widespread and widely used. With the rapid ascent and democratization of this technology (Socol de la Osa & Remolina, Reference Socol de la Osa and Remolina2024) have come “perilous” uses, notably where GenAI is misused by participants in juridical proceedings, resulting in the submission of false or fake material to courts and tribunals (Legg et al., Reference Legg, McNamara and Alimardani2025). This is of immense concern for the court system, as it likely causes additional costs and delays; and more broadly, for the administration of justice, if confidence in lawyers and the judicial system is undermined. In the United States, this fake material has even found its way into judgments,Footnote 1 illustrating the potential danger to the rule of law.
Many courts have responded to the increasing appearance of GenAI in legal proceedings by issuing specific guidance or rules for court users. In this article, we examine Australia, Hong Kong, New Zealand and Singapore. We focus on these four Australasian jurisdictions as presenting contrasting approaches, yet where the enforcement mechanisms available to the courts are broadly the same. They are drawn from a wider database of case law from ten common law and hybrid jurisdictions (Centre for the Future of the Legal Profession, 2025).
Courts also have the power, as we detail below, to control and regulate their own processes in pursuit of the due administration of justice. This extends to both making general rules and to actions which individual judicial officers may take within specific proceedings. In many ways, the “problem” of GenAI misuse is both novel and yet recognizable; both lawyers and lay people appearing for themselves in court proceedings may tender materials to the court that are imperfect or defective, or even fraudulent or fictitious. Almost never before, however, would completely fake case law, or artificially generated statements, be produced unwittingly, or even negligently, by participants. The court is heavily reliant on lawyers, as officers of the court, to uphold their professional ethical obligations by drawing the court’s attention to relevant precedent, not misstating facts or law and following only lawful and proper instructions from the client. At the same time, the scale of GenAI use is enormous, as the accessibility and no or low cost of many tools has driven huge uptake. Indeed, it has been suggested that using artificial intelligence (AI) may become a facet of a lawyer’s duty of competence and to maximize efficiency for clients (Perlman, Reference Perlman2024). The potential to increase access to justice through the use of GenAI is of interest to many in the legal assistance sector, yet all are struggling with how to do so responsibly and ethically (Cesta, Reference Cesta2024).
Given the rapidity of developments, and the differences between courts and jurisdictions, it is not yet clear what “the best” or ideal response to GenAI misuse in legal proceedings looks like. Our analysis here is grounded in an approach which characterizes courts themselves as regulators, and their varied responses as forming part of a hierarchy of regulatory options. Responsive Regulation by Ayres and Braithwaite (Reference Ayres and Braithwaite1992) has become the model of regulatory enforcement in many jurisdictions and of many regulators (Baldwin et al., Reference Baldwin, Cave and Lodge2012, p. 260; Freiberg, Reference Freiberg2025, pp. 589, 597). Moreover, it has been adopted in relation to lawyers and judges (Boon & Whyte, Reference Boon and Whyte2019; Scott, Reference Scott2022; Webb & Nicolson, Reference Webb and Nicolson1999). Here, drawing on the theory of Responsive Regulation, we argue that it offers a nuanced and contextualized approach to understanding the courts’ response as regulators in a complex and dynamic environment. As far as we are aware, this is the first time that the courts’ own exercise of power to control processes and manage compliance has been considered through the lens of Responsive Regulation. More usually, legal proceedings or enforcement represent a layer, usually the apex, of the “regulatory pyramid.”
The structure of this article is as follows: Part 2 describes the increasing appearance of GenAI misuse in the courts, the courts’ response and the general powers that allow specific address within proceedings. Although increasing use of GenAI in this way is occurring worldwide, here we focus on courts in Australasia as illustrating divergent potential responses, ranging from permissive to prohibition. Part 3 explains the theory of Responsive Regulation and describes two sets of regulatory responses: to lawyers misusing GenAI, and to self-represented parties doing so. Part 4 points to areas where the GenAI problem is confounding to this regulatory theory; and how it applies imperfectly to self-represented individuals, in particular, but also how these limitations themselves provide guidance. Part 5 concludes.
2. The GenAI genie
Since November 2022, GenAI use in court and tribunal proceedings has escalated among legal professionals, self-represented litigants (“SRLs”) and third parties such as expert witnesses. A persistent concern with GenAI use is AI-generated “hallucinations,”Footnote 2 with fake or inaccurate citations appearing in written and oral submissions.Footnote 3 In addition to hallucinations, other potential risks include:
• inaccurate summaries, flawed legal reasoning and fallacious arguments presented to courts and tribunalsFootnote 4;
• incorrect research of legal and other “facts”Footnote 5; and
• prolix, poorly drafted or legally incorrect documents, or flooding courts and tribunals with voluminous documentation.Footnote 6
Much media and public commentary on GenAI misuse in litigation focuses on its confirmed or inferred use by legal practitioners (Browning, Reference Browning2024; Perlman, Reference Perlman2024). However, in Australasia, and reflecting a consistent pattern across many jurisdictions globally, GenAI use resulting in problems within litigation processes has more often been by SRLs. As noted in the Introduction, cases and analysis in this article are drawn from a dataset tracking examples of GenAI use in legal proceedings (Centre for the Future of the Legal Profession, 2025). The dataset has been developed based on searches of reported case law accessible via public legal information websites and legal subscription databases from the four jurisdictions mentioned above (Australia, Hong Kong, New Zealand and Singapore) but also Canada, India, Ireland, South Africa, the United Kingdom and the United States. As of March 16, 2026, the whole dataset comprised 1027 cases where problematic use of GenAI was confirmed or inferred in litigation. Examples involving legal practitioners comprised 27 percent of cases, those involving SRLs 69 percent, with the balance comprising witnesses and other individuals, including a small number of judicial officers. These numbers are almost certainly an underestimate of the actual scale of GenAI misuse, as not all instances will be picked up, occur in cases which come before the court, or be mentioned specifically by the decision-maker. This is, therefore, a growing issue, even in jurisdictions where courts have specific GenAI protocols or other forms of guidance in place. In adversarial, common law legal systems, the court is heavily reliant on the material presented by the parties, making this problem even more pronounced.
2.1. The court response
Misuse of GenAI in litigation is problematic for a number of reasons: it impedes the efficient conduct of legal processes, wastes the time of courts and other parties, risks misleading the court, undermines fairness and perceptions of fairness, increases the chance of fake or inaccurate “authorities” making their way into legal precedent, and ultimately undermines the due administration of justice. Moreover, this is occurring against a backdrop of courts’ ongoing experimentation with whether, and how, they themselves can use AI safely and ethically, without compromising judicial values (Bell et al., Reference Bell, Bennett Moses, Legg and McNamara2026).
Recent Australasian cases illustrate the range of potential problems. In a Victorian case, Director of Public Prosecutions v GR [2025] VSC 490, defense submissions on behalf of a juvenile convicted of homicide contained AI-generated errors that neither side detected. In response, Justice Elliott stressed:
The ability of the court to rely upon the accuracy of submissions made by counsel is fundamental to the due administration of justice … any use of artificial intelligence without careful and attentive oversight of counsel would seriously undermine the court’s processes and its ability to deliver justice in a timely and cost-effective manner (at [79]).
No less significant issues arose in Singapore in Forbes Monaco APAC v Seiji [2026] SGHC 16, where the unrepresented applicant submitted AI-hallucinated materials and engaged in other abuses of court processes. In setting aside the claimant’s case, Justice Tan observed:
…[the SRL] had cited no less than nine fictitious legal authorities, and even his attempt to downplay his mistake was contrived and dishonest. In my view, [his] dishonest conduct went beyond an abuse of process – it was nothing short of an attempt to pervert the course of justice (at [2]).
The serious implications for SRLs deliberately or inadvertently submitting AI-hallucinated materials to courts have also been raised in other Australasian cases, such as Weedbrook v Partlin [2024] QDC 194, where Justice Porter observed (at [40]):
This incident reflects the damaging effect of using such models to generate legal submissions. The errors in Mr Partlin’s submissions took up the time of counsel for Mr Weedbrook. It also took up considerable Court time in checking the authorities. This kind of error impacts on the fairness and efficiency of legal proceedings.
The judge noted at [41] that, despite the absence of professional obligations, SRLs nevertheless could commit contempt of court by interfering in the due administration of justice.
These cases also highlight an important point: as officers of the court, legal practitioners owe a higher duty of care than SRLs, and GenAI misuse can potentially trigger a breach of practitioners’ professional and ethical duties.Footnote 7 Although SRLs are not bound by the same duties, they must still comply with relevant court rules and orders when using GenAI, including not misleading the court or other parties.Footnote 8
Courts also have powers to respond to risks and regulate their own processes. In Australia, the inherent powers of the state Supreme Courts (superior courts of record) and the superior Federal Courts support supervision of the conduct of legal practitioners and the exercise of disciplinary powers (Dal Pont, Reference Dal Pont2024, p. 750, [23.10]). As we explain below, this disciplinary role in relation to practitioners exists in tandem with that of legal professional regulators. Pursuant to the same inherent powers, courts may also issue practice notes or guidelines expanding on court rules or specifying procedures for certain types of proceedings.
Given the uptick in GenAI use in court and tribunal proceedings, some Australasian courts have moved to control its use by issuing guidance, directives or practice notes on the use of AI in legal proceedings, and guidelines addressing how judicial officers or court staff may or may not use GenAI in their work. Supplementary Appendix 1 summarizes these responses.Footnote 9
Several trends are apparent from these emerging frameworks. A spectrum of approaches is discernible, ranging from highly proscriptive regimes, where AI use is restricted, disclosure is mandatory and guidance is highly detailed (NSW); to more permissive guidance allowing responsible AI use (Victoria, Queensland, South Australia, New Zealand and Singapore); principles-based guidance (Hong Kong); or reliance on existing professional obligations. In addition, there is divergence of guidance by audience, as some jurisdictions have published guidance or protocols that are designed to address GenAI use specifically by SRLs; some by all court users; and others specifically by judicial officers and their staff.Footnote 10
Overall, emerging frameworks reflect the regulatory challenges associated with technological change. However, the combination of various courts’ approaches to addressing AI use, divergence of guidance by audience and resulting regulatory fragmentation has the potential to increase uncertainty regarding AI use for court users. This issue is heightened by divergent responses from judicial officers faced with AI use by parties during legal proceedings. For example, as we discuss below, some judges have referred legal practitioners to a regulator; others have imposed their own sanctions, mainly costs; some have named the individuals involved, others have suppressed names.
Additional challenges are those faced by courts trying to keep up with rapid technological change, and ongoing persistent gaps in AI literacy across much of the judiciary and legal profession as well as among SRLs.
2.2. Sources of judicial regulatory power
Lawyers are subject to stringent professional regulation. In the classical formulation, this is one-half of the “regulative bargain” with the State, granting legal practitioners a professional monopoly over legal work, in exchange for submission to intensive requirements for legal education and training, and disciplinary oversight (Suddaby et al., Reference Suddaby, Cooper and Greenwood2007). In many jurisdictions, legal professional regulation is no longer a purely self-regulatory system – oversight may be by an independent regulatory body, or there may be co-regulation between a professional and an independent body. Courts have always been part of the regulation of the profession. When admitted to practice, lawyers become officers of the court through signing the court roll (Committee to Review the Regulatory Framework of the Singapore Legal Services Sector, 2014, p. 7), and as such, a superior court may exercise disciplinary powers over them, up to and including removal from the roll. The primary goal of the disciplinary jurisdiction is to protect the public, enabling people to place their trust in lawyers and have faith in the administration of justice. The court’s inherent power to discipline is exercised in parallel with statutory powers conferred by the legal profession legislation.Footnote 11 The Legal Profession Uniform Law (“LPUL”) and Australian Solicitors Conduct Rules (“ASCR”)Footnote 12 set solicitors’ ethical and professional duties in three Australian states, including their paramount duty to the court and to the administration of justice (Legg, Reference Legg2024).Footnote 13 Analogous regimes apply to legal practitioners elsewhere in Australasia.Footnote 14 Legal practitioners are expected to uphold high standards of conduct in all aspects of legal practice, including in the conduct of litigation.
Such laws fulfill various purposes, including to promote the administration of justice and “regulation of the legal profession that is efficient, effective, targeted and proportionate.”Footnote 15 Thus, as discussed below, one avenue for courts is to refer legal practitioners to the disciplinary body for it to investigate potential wrongdoing and impose sanctions if deemed appropriate.
These disciplinary powers are part of the court’s broader, inherent jurisdiction to regulate its processes and ensure public confidence in the administration of justice (Goh, Reference Goh2011; Jacob, Reference Jacob1970). Lacey (Reference Lacey2003) explained: “It has long been accepted that a court’s inherent powers may be invoked by a court to ensure the integrity, efficiency and fairness of its process and in a manner that protects, among other things, due process and the provision of a fair trial.”
The court’s power to make rules of court and practice directions is attributed to the inherent power to ensure that legal proceedings are conducted fairly and conveniently (Mason, Reference Mason1983, as cited in Lacey, Reference Lacey2003).Footnote 16 Courts may also act to prevent abuse of their processes. Lower-ranking courts are granted these powers by their establishing Acts. However, in EML (a pseudonym) v AFT (a pseudonym) [2026] VCC 734 [60] the County Court of Victoria considered there was a real question as to whether, as a non-superior court, it could exercise the same inherent powers to discipline legal practitioners as the superior courts (at [60]).
Nevertheless, courts have the powers to regulate and sanction the misuse of GenAI in proceedings, and some have continued to rely on these general powers in preference to creating specific guidelines. As we discuss further below, however, while specific guidelines may not be strictly necessary (and may contribute to an “illusion of control” problem), they may also serve an educative function, especially for the legal profession.
3. Regulatory model
In this section, we provide an overview of Ayres and Braithwaite’s theory of Responsive Regulation. We then apply this model first to the courts’ regulatory actions in regulating GenAI use by lawyers, and then to the activities of SRLs. The benefits of this approach are to make explicit the courts’ regulatory powers; but also to consider the actions of judicial officers in light of a broader public policy schema. In Section 4, we turn to consider the limitations of Responsive Regulation in this context and how this may, paradoxically, improve its utility.
3.1. Responsive regulation
We draw on Responsive Regulation as a theoretical and policy framework for organizing judicial responses to GenAI misuse. Responsive Regulation sought to combine compliance and deterrence, or persuasion and punishment, through the enforcement pyramid. The base of the pyramid, where most regulatory action occurs, encompasses the use of persuasion to coax compliance. There are, then, a number of steps in which the regulator’s action is the use of an enforcement tool that gradually becomes more punitive, and the approach is to deter misconduct through various forms of sanctions. At the top of the pyramid are criminal prosecutions that result in imprisonment or the revocation of an entity’s ability to operate (Ayres & Braithwaite, Reference Ayres and Braithwaite1992, p. 20). The Responsive Regulation thesis is that regulatory agencies are best able to secure compliance when they are “benign big guns,” that is, regulators will be more able to “speak softly when they carry big sticks (and [have access to] a hierarchy of lesser sanctions)” (Ayres & Braithwaite, Reference Ayres and Braithwaite1992, p. 19). Note: Responsive Regulation was reworked in later incarnations, but here we draw on its original formulation.
Central to Responsive Regulation are the following assumptions. An actor’s decision as to whether to comply with the law may be driven by whether it is economically rational to do so – a weighing of potential costs and benefits – but also by their concern to do what is right because of their own perception of themselves as law-abiding citizens who are socially responsible. A strategy based totally on persuasion and self-regulation will be exploited by those entities who are making decisions based on economic rationality. As persuasion and self-regulation do not have the same sort of costs as civil litigation or criminal prosecution, it may be rational not to comply. However, a strategy based mostly on punishment will undermine the goodwill of those actors who are motivated to comply because of their self-perception as responsible, law-abiding citizens (Gunningham, Reference Gunningham, Baldwin, Cave and Lodge2010, p. 124). Further, punishment is expensive as contraventions need to be proven and due process followed, usually in a court. Persuasion is cheap as the actor will voluntarily comply. Yet, for compliance to be forthcoming, it needs to be supported by a hierarchy of sanctions so that a more costly and punitive response is available if needed (Ayres & Braithwaite, Reference Ayres and Braithwaite1992, pp. 10, 20, 22, 24; Gunningham, Reference Gunningham, Baldwin, Cave and Lodge2010, pp. 125–26).
The regulator needs to establish a synergy between punishment and persuasion so that the threat of punishment underwrites persuasion as something to be preferred. The aim is to give regulatory agencies the maximum capacity to lever cooperation by allowing them to escalate their response in a way that responds to the degree of uncooperativeness of an entity, while also taking account of the moral and political acceptability of the response (Ayres & Braithwaite, Reference Ayres and Braithwaite1992, p. 36). A regulator needs to have not just a hierarchy of enforcement steps or sanctions it can invoke but also a sufficient number of enforcement steps so that it is not placed in a position where non-cooperation at one level leads to it needing to escalate the response to a level that is considered unacceptable from a political perspective. The advantage of a pyramid of sanctions is the ability to invoke a response to misconduct commensurate with the contravener’s actions – mistaken compared to deliberate conduct, one-off infractions in contrast to multiple or repeated contraventions, replying honestly and demonstrating contrition such as voluntarily addressing the misconduct and ameliorating its impacts, compared to denials and obfuscation which necessitate investigation. The pyramid also creates an incentive for compliance, or at least self-reporting, due to the desire to avoid costly, higher-level outcomes. This has the advantage of reducing the regulatory resources that are required. Judicial resources are typically scarce and ideally should be directed to efficiently resolving disputes rather than dealing with misconduct (Gunningham, Reference Gunningham, Baldwin, Cave and Lodge2010, p. 126). However, as we show in Part 4, it may also be that in certain contexts, some levels of the pyramid – in this case, especially as they relate to SRLs – may be less effective.
3.2. Responsive regulation applied to courts regulating lawyers
Building on the Responsive Regulation framework, we now apply the concept of an enforcement pyramid model to courts’ regulation of lawyers. Figure 1 adapts Ayres and Braithwaite’s enforcement pyramid to the context of confirmed or inferred GenAI misuse by legal professionals in litigation, illustrating the breadth of regulatory levers available to courts.
Responsive regulation pyramid for legal professionals.
At the base of this pyramid is guidance and education (the level where most activity currently occurs). This activity aims to encourage responsible adoption of AI, coupled with appropriate verification and relevance checks. Courts can escalate their response when required to the mid-level of the pyramid, starting with corrective measures and then moving to progressively more interventionist responses, ranging from reputational impact and judicial commentary to costs and other orders. At the apex of the pyramid are the most punitive regulatory responses, being professional discipline and criminal or other legal sanctions (such as referral to the police or contempt proceedings).
This model provides a structured lens for how courts might calibrate their responses to GenAI misuse, depending on the gravity and context, that is, whether the conduct was ignorant or deliberate, accidental or careless, isolated or repeated. The following analysis applies the pyramid framework to reported cases of GenAI use in legal proceedings, beginning with responses at the base.
3.2.1. Guidance and education
Cases of GenAI misuse in legal proceedings were reported in Australasian courts during early 2023,Footnote 17 but formal AI-related guidance followed more slowly.Footnote 18 Early Australasian cases sometimes simply note the confirmed or inferred use of GenAI to generate materials submitted to the court, before moving on to the merits of the case.Footnote 19 As concerns about misuse of GenAI increased, Australasian courts began addressing the issues more directly. Courts generally emphasize that AI use per se is not improper – the issue lies in the potential breach of ethical and legal professional duties when AI is used without exercising “judgment and professional skill in reviewing the final product to be provided to the court.”Footnote 20 Even in GenAI cases where outcomes are more punitive, judgments also perform a “guidance and education” function, aimed at raising awareness of professional obligations and potential risks.
Re Mr Dayal [2024] FedCFamC2F 1166 (a pseudonym) is the first reported Australian case where a legal practitioner was formally disciplined after misusing GenAI. The practitioner used an AI tool embedded in practice software to generate a list of fictitious case authorities and summaries, and filed it without first verifying its accuracy and integrity (Re Mr Dayal, at [8]). In determining whether to refer the practitioner to the legal regulator for his conduct, Judge Humphreys observed at [10]:
The use of technology is an integral part of efficient modern legal practice. At the frontier of technological advances in legal practice and the conduct of litigation is the use of AI. Whilst the use of AI tools offer opportunities for legal practitioners, it also comes with significant risks.
Her Honor referenced AI guidance issued by other courts (at [13]), and the ASCR (at [17]). The case attracted publicity that amplified its educative effect (Taylor, Reference Taylor2024). Another Australian case with similar facts and impact is Valu v Minister for Immigration and Multicultural Affairs (No 2) [2025] FedCFamC2G 95, where a NSW legal practitioner (whose name was also suppressed) inadvertently filed submissions citing nonexistent case law and purported quotes, generated using ChatGPT (at [22]).
The impugned conduct occurred before the issue of the NSW Supreme Court’s practice note on GenAI (NSW Supreme Court, 2024). Although that note did not apply to the proceedings in Valu No 2, which took place in a Federal court, it circulated widely and attracted commentary in NSW legal circles (Barbera, Reference Barbera2024; Ea, Reference Ea2024). The practitioner referenced the note in submissions explaining his conduct, saying that “if he had access to the Practice Note at the time he drafted the submissions, he would not have incorporated the material generated by AI” (at [23]).
In the course of reviewing the practitioner’s conduct and referring the person to the NSW legal regulator, Judge Skaros observed at [18]:
The conduct … falls short of the standard of competence and diligence that the applicant in the substantive proceedings was entitled to expect from his legal representative. The conduct also falls short of a legal practitioner’s duty to the Court, including the duty to ensure that the Court is not deceived or mislead, even if unintentionally …
Several other Australasian GenAI cases illustrate the ongoing need for court-led guidance and education linking responsible AI use to practitioners’ duties. In Wamba Wemba [2025] FCA 731, Murphy J observed at [12]:
Whilst the use of AI in the legal profession is growing, practitioners must be aware of its limitations. It is critical that legal practitioners use proper safeguards to verify the accuracy of the work produced. Any use of AI must be consistent with the overriding duty of legal practitioners as officers of the Court and their fundamental obligation to uphold, promote and facilitate the administration of justice.
This emphasis is also evident in JNE24 v Minister for Immigration and Citizenship [2025] FedCFamC2G 1314, where Gerrard J stated “To be clear, it is not the initial reliance on AI that constitutes the vice in such matters. It is the placing before the Court of false authorities or evidence that constitutes improper conduct and a breach of a legal practitioner’s duty to the Court” (at [21]; see also Tajudin bin Gulam Rasul and Anor v Suriaya bte Haja Mohideen [2025] SGHCR 33).
3.2.2. Corrective measures
Corrective measures sit between guidance and education, and more punitive responses in the enforcement pyramid. Corrective measures might include actions such as giving leave for a lawyer to file corrected documents, mandating forms of continuing legal education and warning of future sanctions (see Legg et al., Reference Legg, McNamara and Alimardani2025; Merken, Reference Merken2025).Footnote 21 To date, Australasian courts have largely preferred guidance and education, although they do have access to these corrective measures; or else escalated to more punitive responses, bypassing this intermediate level.
Corrective measures are still an escalation, and may involve a degree of coercion or further escalation if misuse persists. They can also be calibrated to a practitioner’s intentionality and prior history; actual or potential adverse effects of GenAI use; and promptness and completeness of remediation once the issue has been identified.
3.2.3. Reputational impact and judicial commentary
Given heightened professional and public interest in GenAI cases, courts are sometimes using reputational levers, such as “public naming and shaming” and critical judicial commentary, to encourage responsible adoption, especially for legal practitioners. This is unlikely to be as effective with SRLs (as we discuss below), as they generally do not have the same vested interest in avoiding adverse publicity due to AI misuse.
Australian courts have tended to use this lever with restraint. In Dayal, the court anonymized the name of the practitioner involved. In Valu No 2, Skaros J accepted that the legal practitioner was “deeply embarrassed by his conduct” and noted that referral to the regulator was not intended to embarrass him further, also electing to anonymize him “to protect his identity” (at [36]). The practice of anonymizing the names of legal practitioners was also adopted in Wamba Wemba and JNE24.
However, other Australian GenAI cases demonstrate the potential for deterrence through naming when conduct warrants it. In DPP v GR, the court identified and criticized senior counsel, his instructing solicitors and also the prosecution, despite apologies and senior counsel taking responsibility for the filing of submissions containing nonexistent AI-generated authorities (at [61]–[80]). Judge Elliott emphasized:
Regrettable as it is to single out counsel and their instructing solicitors in this case for what has occurred, in light of the matters set out above it is important to record that counsel must take full and ultimate responsibility for any submissions made to the court. To this end, it is not acceptable for artificial intelligence to be used unless the product of that use is independently and thoroughly verified. The same may be said for solicitors responsible for producing or filing court documents (at [79]–[80]).
This case involved the criminal prosecution of a minor for homicide, and attracted notable publicity (McGuirk, Reference McGuirk2025). In Re Walker [2025] VSC 714, the court noted the importance of deterrence, but also that it must “give way to proportion”; further, that the public naming of the legal practitioner involved “is itself a significant adverse consequence for her professional standing” (at [81]).
Reputational levers and judicial commentary can be calibrated by courts to the context of each case, anonymizing practitioners where appropriate, and using “naming and shaming” only where essential to protect integrity of legal processes, and to deter repetition of GenAI misuse.
3.2.4. Costs and other orders
Costs orders have become a primary court response to GenAI misuse by legal practitioners, operating to compensate for “costs thrown away” and to deter repetition.
In Dayal, the lawyer involved avoided a costs order by proactively compensating the other side’s solicitors in addition to other voluntary mitigation actions. In contrast, an earlier decision on the substantive issues preceding Valu No 2 resulted in a personal costs order of AU \$2980 for the lawyer’s misconduct (at [80]–[85]). In JNE24, a personal costs order of AU \$8371.30 was made against the lawyer involved (at [36]). The court noted the lawyer’s own client “should not be responsible for the legal fees associated with this improper conduct” and the lawyer reimbursed his client, avoiding an additional order (at [36]–[43]). A personal costs order was also made against Singaporean counsel in the sum of SGD \$800, approximately AU \$945.Footnote 22 The party whose lawyer had used AI was ultimately successful, and the costs represented the inconvenience to the other party of the lawyer, including one fictitious case reference in their materials. The court considered it relevant that the fake case was cited for a “trite proposition” of law, against which the defendant ultimately decided not to proceed.Footnote 23
Finally, in Wamba Wemba, indemnity costs were ordered against the applicant’s law firm for “the firm’s use of artificial intelligence in the preparation of documents” (at [16]). No personal order was made against the junior practitioner involved, in part due to her inexperience and the firm’s inadequate supervision systems (at [15]).
Courts may also use procedural orders to protect the integrity of legal proceedings and limit or refuse reliance on AI-generated materials, without resorting to costs orders. In an Australian federal case, the judge declined to admit AI-generated transcripts of audio recordings of phone conversations, citing risk of hallucination.Footnote 24 In another case, an application to grant an extension of time to file an appeal (presumably drafted by the applicant’s solicitor) was refused in part because of defective submissions that appeared to be written by AI.Footnote 25
3.2.5. Professional discipline
At the punitive end of the enforcement pyramid, courts can refer legal practitioners to the relevant professional regulators, a lever generally justified on the grounds of public interest.
In Dayal, despite proactive mitigation by the practitioner involved (including self-reporting to his professional indemnity insurer), Judge Humphreys held it was in the public interest for regulators to be alerted “given the increasing use of AI tools by legal practitioners in litigation more generally,” and referred his conduct to the regulator (at [21]–[22]). This referral resulted in variations and conditions on his practicing certificate (Victorian Legal Services Board + Commissioner, 2025).Footnote 26
In Valu No 2, although Judge Skaros acknowledged that there were mitigating circumstances and the practitioner had taken steps to address his errors, her Honor noted “the misuse of generative AI is likely to be of increasing concern and that there is a public interest in the [regulator] being made aware of such conduct as it arises,” similarly directing that his conduct be referred to the regulator (at [34]–[38]), and JNE24 also followed this approach.
By contrast, in Wamba Wemba no referral was made. Judge Murphy noted distinguishing circumstances, including the practitioner’s lack of experience and a failure of supervision, as well as apologies, observing:
… an inexperienced junior solicitor was given the task of preparing document citations for an amended pleading, and did so while working remotely and without access to the documents to be cited. In attempting to cite the relevant documents she used an (apparently AI-assisted) research tool which she considered had produced accurate citations when she previously used it. And as soon as [the law firm] was told of the false citations the problem was addressed (at [14]).
Where public interest, potential risks to the due administration of justice and repeated misconduct are considerations, courts may respond with a referral to the regulator. However, courts may refrain from referral where the conduct reflects inexperience and is remedied, instead utilizing less punitive responses such as emphasizing practice improvements.
3.2.6. Criminal or legal sanctions
While no Australasian GenAI case has yet reached the apex of the enforcement pyramid, courts have signaled the potential for this to occur.Footnote 27 In JNE24, referring to the recent UK decision of Ayinde v London Borough of Haringey [2025] EWHC 1383, Gerrard J observed at [20]:
AI is increasingly being used as a research tool in litigation. Ultimately, it is likely to prove to be an invaluable tool for both lawyers and self-represented litigants. There is nothing inherently impermissible about using generative AI programs to assist in research. However, not only is it not an appropriate substitute for legal research, it comes with considerable risks which, if not mitigated, have the capacity to lead to actions which could be construed as a contempt of court…
In conclusion, effective application of an enforcement pyramid model to legal practitioners’ use of GenAI requires adjustment of regulatory levers depending on the context of each case, aligning the court’s response with intended outcomes and escalating to more punitive responses only in the more extreme cases of GenAI misuse.
3.3. Responsive regulation applied to courts regulating SRLs
Cases in which SRLs have used GenAI in court proceedings are far more numerous than cases involving lawyers (Legg et al., Reference Legg, McNamara and Alimardani2025). This is unsurprising given that SRLs are likely to lack legal knowledge; may not realize that GenAI results must be verified; and may not have the means of conducting that verification. Further, when compared to lawyers, they may have varied purpose and motivation for using GenAI, ranging from honest attempts to put their case to the court, through to underhand or vexatious motivations. SRLs’ use of GenAI is also related to broader access to justice issues. While some SRLs choose to self-represent, others may not qualify for State-funded legal assistance yet be unable to afford a lawyer, turning to GenAI in the absence of alternatives. However, the use of GenAI by SRLs can still cause substantial problems, and all litigants are under a duty not to mislead the court or their opponent.Footnote 28
In some cases, the court has simply ignored the tainted material.Footnote 29 This has been a fairly common response from Australian courts confronted with GenAI use by SRLs – noting that use has occurred but making no further comment and taking no action concerning the GenAI use.Footnote 30 In other cases, some of which are discussed below, the whole proceeding has been found to be vexatious or an abuse of process, so that the issue of AI use, while noted, has been minor in comparison.Footnote 31
Figure 2 adapts the enforcement pyramid to the context of confirmed or inferred GenAI misuse by SRLs.
Enforcement pyramid for self-represented litigants.
3.3.1. Guidance and education
The base level, and most common, response of the courts to GenAI use by litigants in person is to adopt an educative approach. The creation of court guidelines, and references to them in case law, or to the risks of using GenAI, may act as attempts to educate parties on the dangers and risks of using GenAI in proceedings. Different approaches can be seen in terms of whether a court has produced guidelines specifically directed at SRLs (such as in Queensland, South Australia and the Federal Court of Australia) or has one set of guidelines that are intended to apply to both lawyers and legal practitioners (such as the Victorian Supreme Court Guidelines). Guidelines or practice notes also vary in terms of style and simplicity (Queensland Courts, n.d.). For example, the Queensland Courts’ guidelines for non-lawyers include a web page summary, but the full guidelines must be viewed as a PDF. The Victorian Guidelines are written in more formal language, for instance: “Generative AI and Large Language Models create output that is not the product of reasoning. Nor are they a legal research tool” (Supreme Court of Victoria, 2024).
Courts have then used specific cases to both draw attention to the guidelines, and to make educative statements about the risks of using GenAI in legal proceedings. Where judicial officers or Tribunal or Commission members have decided to publish their decision (which may not occur routinely in all instances) commenting on GenAI use, they are necessarily performing an educative function. This might include drawing attention to the incorrectness of the material generated,Footnote 32 or even how use of GenAI might have contributed to a poorer legal outcome for the person concerned due to their use of GenAI in earlier interactions.Footnote 33 Courts and tribunals have therefore emphasized the general lack of utility in SRLs using GenAI to prepare material.Footnote 34
3.3.2. Corrective measures and warnings
This level of the pyramid encapsulates the idea of the court more actively managing and intervening in a party’s use of GenAI, often coupled with warnings about the inappropriate use. This includes instances where the court has gone beyond educative statements to criticize the party for their GenAI use or warn them about potential negative consequences.Footnote 35
For example, in a South Australian tribunal case, Blackmore & Anor v Smyth & Anor [2025] SACAT 48, the self-represented applicant had used GenAI to prepare submissions which contained misstatements of legislative provisions and fake cases. The applicant was also, initially, not honest about his use of AI. The Tribunal noted that its governing Act allowed the dismissal of an application in certain circumstances,Footnote 36 including “abuse of process and attempting to deceive another party or the Tribunal,” and considered that misrepresenting legislative provisions and citing nonexistent cases could constitute both of these (at [15]). Ultimately, the Tribunal found, however, that the applicant had not “consciously” sought to engage in deception or abuse of process (at [16]). Similarly, in a Federal Court bankruptcy case, the Court was at pains to emphasize the illegitimacy (and therefore seriousness) of using GenAI, suggesting that the terms “fabricated, fictional, false, fake” were more apposite than “hallucinations.”Footnote 37 The Court referred to the potential for the submission of fabricated materials to the court to constitute contempt,Footnote 38 though ultimately no further actions were taken concerning the AI use.
Courts have also taken additional “case management” measures, such as requiring a party to amend and refile documents, give evidence about the content of statements, or disregarding material.Footnote 39 In an NSW Civil and Administrative Tribunal case, a self-represented applicant explained that she had used ChatGPT to “improve” her statements.Footnote 40 The Tribunal considered that the statements could still be admitted into evidence, as the applicant was cross-examined on them, and was unaware of the Practice Direction prohibiting the use of GenAI in relation to witness statements.Footnote 41 Gribble v ESSENTIAL ENERGY [2025] NSWDC 344, an interlocutory decision of the NSW District Court, is an example of the court making use of these types of powers in the face of extensive GenAI use by an SRL. Judge Gibson struck out the plaintiff’s statement of claim with leave to replead, but mandated that if refiled, the material must “exclude all information obtained through Gen AI.”
3.3.3. Costs and other orders
Costs orders are one of the few penalty mechanisms available to the court seeking to regulate GenAI use by SRLs. The court may also refuse to accept documents for filing; and/or dismiss or strike out claims or defenses (including summarily, i.e., without a full hearing). Even jurisdictions which generally permit the use of GenAI in connection with proceedings, such as Singapore, emphasize the availability of these powers for misuse. In Gribble, the plaintiff was ordered to pay the defendant’s costs of the argument. The court concluded at [29] that the plaintiff’s pleading was hopeless, “given the nature and extent of the errors,” which included fake, AI-generated case names and principles. In the Hong Kong case of Licksun Co Ltd v Occupiers of Lot No 552 in DD 85, New Territories, Hong Kong [2025] HKDC 1287, the court ordered the SRL to pay indemnity costs for lodging an appeal that contained “suspicious” nonexistent authorities, due to misleading the court. However, it was not established conclusively that the fake cases were generated with AI (see at [3]).
In other cases, an SRL’s application has been dismissed, and in some, costs orders have been made.Footnote 42 However, this has not always been directly related to the AI use – rather, that has been only a small part of bigger problems related to the proceedings.Footnote 43 Arguably, using fake AI-generated material likely contributes to a court’s findings that proceedings are hopeless, vexatious or constitute an abuse of process but has not yet independently founded such a finding.
3.3.4. Criminal or legal sanction or investigation
SRLs may commit contempt of court, both civil and criminal.Footnote 44 In a civil contempt case, the court must consider “whether the impugned conduct had the capacity to interfere with, or obstruct, the proper administration of justice in this Court.”Footnote 45 It is sufficient that an alleged contemnor engaged in the acts constituting the contempt deliberately and voluntarily – intention to interfere with the administration of justice does not have to be proved.Footnote 46 A civil contempt may become criminal if it is contumacious or if the intention is to inflict punishment for past breaches (as opposed to being remedial).Footnote 47 To find a criminal contempt, the required standard is the criminal standard of proof, i.e., beyond reasonable doubt. In theory, the provision to the court of nonexistent authorities or legislation could amount to contempt of court. In most cases to date, however, the court has not been satisfied that the SRL has understood the submitted material to be fictitious or fake, such that they could be said to have deliberately attempted to mislead the court.Footnote 48
A further avenue by which SRLs may invite sanction due to the use of AI is through uploading documents produced for the purposes of court proceedings into GenAI tools, thus breaching an implied undertaking to the court.Footnote 49 This is referenced in guidelines from the NSW Supreme Court (2024, para. 9); Singapore Courts (2024, p. 4, para. 5(7)) and Courts of New Zealand (2023, para. 2). In an Australian family law decision, the court speculated that uploading documents from the proceedings to a GenAI model might also potentially fall foul of the prohibition on “communicating” an account of proceedings to the public.Footnote 50 This also constitutes a criminal offense.
4. Limitations of responsive regulation
As explained above, Responsive Regulation provides a lens through which responses to GenAI misuse may be ordered and applied in a more systemic manner. However, Responsive Regulation is not without its shortcomings (Baldwin & Black, Reference Baldwin and Black2008; Baldwin, Reference Baldwin, Baldwin, Cave and Lodge2010, pp. 260–265; Gunningham & Grabosky, Reference Gunningham and Grabosky1998). Recognition of the limitations of the theory may, paradoxically, aid in its utility by alerting the decision-maker – here a judge or the judiciary – as to how to better implement responses to GenAI misuse.
4.1. Competence of the regulated entity
The first weakness is an assumption that the regulated entities have the resources or “wherewithal” to be able to learn from educative processes, including warnings. For the lower levels of the enforcement pyramid to be effective, the regulated entity must have sufficient expertise or competence to be able to comply with legal requirements. Robert Kagan and John Scholz developed three types of regulated entity, drawing on an examination of corporations, which require different responses from a regulator. In addition to the good corporate citizen and profit maximizer that Responsive Regulation attempts to deal with, they also focused on the incompetent corporation that does not have the resources or systems to be able to follow the law (Kagan & Scholz, Reference Kagan, Scholz, Hawkins and Thomas1984).
Examples of Gen AI misuse illustrate that both lawyers and SRLs may not understand how GenAI functions or that it may hallucinate and the need to check the information generated. The question for courts is whether, having made lawyers and SRLs aware of the risk of incorrect AI-generated law, they are then capable of acting correctly. Without competence in compliance, the regulated entity will fail to comply with the law. As a result, the regulator may need to devote additional resources to build competence – what is sometimes called capacity building (Kagan & Scholz, Reference Kagan, Scholz, Hawkins and Thomas1984, p. 82). A lack of competence in compliance is likely to be common among SRLs, but less so among lawyers who have access to assistance and resources (though varying by the size and profitability of their practice).
The lack of compliance competence among SRLs suggests that the resourcing for education may be greater than courts have recognized. SRLs may benefit from more targeted communications and training about the appropriate use of GenAI when researching their legal cases and preparing submissions and other evidentiary material, in addition to general AI protocols or those targeting legal professionals and judicial officers. In Australia, guidance documents or even practice notes appear to be insufficient to completely “solve” the issue, given the quickly escalating numbers of SRLs using GenAI in proceedings. This is because it assumes that the SRL will find, read and understand the guidance provided. However, the Federal Court of Australia and the Courts Administration Authority of South Australia, having produced guidance more recently, have opted for plain English explanations accompanied by examples of “good” and “bad” prompts as well as online directions to free (and accurate) legal sources.
As a result, greater efforts may be required, but a question arises as to whether there are sufficient judicial resources for addressing capacity-building among SRLs. Regulator resourcing is a constant issue in relation to whether a Responsive Regulation approach can be adopted (Baldwin et al., Reference Baldwin, Cave and Lodge2012, p. 263). Moreover, a lack of competence among SRLs seems to be an assumption of the judicial system,Footnote 51 resulting in continued non-compliance. As it stands, courts have different methods of case management, and people also have different reasons for being self-represented. In some fora, SRLs are usual or expected, while the procedures of other courts assume that everyone will be legally represented. Where this is not the case, judicial officers must achieve a fair court process (ensuring a fair hearing) without advising or assisting the SRL with their case (ensuring impartiality is maintained) (Richardson et al., Reference Richardson, Grant and Boughey2018, para. 2.3.4). Richardson et al. have explained that case management of SRLs can “reduce the numbers of adjournments, increase timely filing of material, identification of issues, thereby assisting cases to progress more quickly than they would have without case management” (Victorian Government Department of Justice and Regulation, 2016, p. 480 as cited in Richardson et al., Reference Richardson, Grant and Boughey2018, para. 2.4.6, pp. 46–47). Interventionist case management and the provision of support to SRLs may be a way of ensuring that inappropriate GenAI material does not find its way into documents filed with the court while also improving an SRL’s ability to access justice. This might include asking SRLs to certify that documents filed with the court have not made use of GenAI improperly or without checking case names and citations; and issuing warnings about possible penalties for non-compliance.
There may also be a need for assistance from third parties. Responsive Regulation incorporates “tripartism” (Ayres & Braithwaite, Reference Ayres and Braithwaite1992, Ch. 3), a form of regulatory pluralism, but it is generally overlooked in practice. Indeed, regulatory pluralism, which seeks to harness not just governments but a wide range of different actors (such as business and third parties) as regulatory agents or surrogates, arose due to Responsive Regulation being seen as state-centric. One strand of regulatory pluralism, Smart Regulation, proposed a three-sided enforcement pyramid as a way in which to recognize the role of third parties (Gunningham & Grabosky, Reference Gunningham and Grabosky1998). In the current context, professional regulation of lawyers may be seen as part of pluralism, especially when conducted by the profession or pursuant to a co-regulatory model.Footnote 52 Lawyers will be required and assisted by professional regulation to develop capacity for compliance. This is illustrated by the requirements for admission and continuing professional development.
While professional regulatory systems allow more flexibility and creativity when it comes to regulating lawyers, a greater focus on third-party organizations who may act as regulatory surrogates vis-à-vis SRLs is needed. Moreover, assistance is needed prior to the SRL appearing before the court and finding that their documents or submissions are “misconceived, unhelpful and irrelevant.”Footnote 53 There are various website resources about self-representing in proceedings which could be leveraged to direct people to GenAI resources/guidance. However, earlier studies have found that information on court websites is not helpful for SRLs and most will not be able to use this information effectively (Toy-Cronin, Reference Toy-Cronin2015). Third parties such as legal aid and legal assistance sector organizations may be appropriate destinations for SRLs to receive information about GenAI and its uses and risks (Cameron & Kelly, Reference Cameron and Kelly2002, pp. 313, 320). The effectiveness of these third parties as regulatory surrogates will depend on their accessibility and resourcing in the relevant jurisdiction (Singapore Courts, n.d.).
Responsive Regulation accepts that if the entity lacks the competence to comply, the regulator may need to incapacitate the entity or remove it from the industry (Ayres & Braithwaite, Reference Ayres and Braithwaite1992, pp. 35, 53). This leads to the second and third issues.
4.2. Fear of escalation
The second weakness is that regulated entities fear, or at least have a preference against, escalation up the enforcement pyramid. Responsive Regulation operates on the basis that regulated entities that fail to voluntarily comply can usually be coerced into compliance by the threat of more costly sanctions. The regulator shifts from a compliance approach to a deterrence approach. The threat of escalating up the enforcement pyramid, or indeed the actual use of higher-level enforcement tools, is more likely to be effective with lawyers than SRLs.Footnote 54 In the court and lawyer context, it must be remembered that the judiciary operates in an institutional setting where legal practitioners owe duties to the court and a contravention of such a duty, such as misleading the court, can result in reputational harm, or in serious circumstances an inability to practice.Footnote 55 Further, this type of sanction can be communicated to all lawyers and therefore promote general deterrence (Gunningham et al., Reference Thornton, Gunningham and Kagan2005).
In contrast, SRLs are likely to be individuals or small entities who may have different motivations for compliance and be less susceptible to sanctions.Footnote 56 An SRL has, by definition, determined that they are not prepared to incur the cost of a lawyer, in some cases because they simply lack the necessary resources (Whalen-Bridge, Reference Whalen-Bridge2021, p. 483). As a result, cost- or money-based sanctions are unlikely to be effective as they have no ability to pay. Similarly, reputational concerns may be nonexistent (Baldwin & Black, Reference Baldwin and Black2008, p. 9; Wiig et al., Reference Wiig, Calderwood, Høie and Braithwaite2025). From the SRL’s perspective, they may see themselves as doing their best in an unfamiliar environment. Alternatively, the SRL who is, or leans toward, the category of vexatious litigant may have no interest in complying voluntarily. Accordingly, the court may need to escalate its response to actions that restrict the ability to pursue a claim or rely on particular arguments or evidence in the hope that the SRL places value on not undermining their own prospects of success. Case law to date suggests that courts have imposed the most severe sanctions for GenAI misuse on SRLs who are already behaving more “vexatiously.”
4.3. Assumption of rationality
But this leads to a further assumption, and the third weakness: that the regulated entity has a level of knowledge or rationality in relation to the pursuit of the litigation. The SRL needs to be engaging in litigation in a rational manner where their aim is to succeed on their claim in an efficient manner, and they comprehend how the court’s sanctions will impact on the claim. If that assumption does not hold, then striking out claims/defenses or making costs orders may be ineffective and the court may need to rely on criminal law sanctions, such as criminal contempt, to deter misconduct. This then raises the idea of incapacitation as a regulatory response raised above. However, the idea of removing a litigant from the system because of an inability to comply with requirements for, or restrictions on GenAI does not sit well with the role of courts and citizens being able to access the courts to resolve disputes (Brennan, Reference Brennan1998, p. 35; Legg, Reference Legg2020).Footnote 57 Courts may be placed in the difficult situation of trying to calibrate the regulatory response where mid-level sanctions are ineffective but top-of-the-pyramid sanctions are an overreaction.
The response in such situations is typically that the pyramid needs additional levels (Legg, Reference Legg and Legg2011). For example, courts might consider employing the power to stay proceedings (see, e.g., Civil Procedure Act 2005 (NSW) s 67) in a novel manner, such as until the litigant has undertaken some form of education on the use of GenAI in litigation or mandate that the litigant have their case law references certified by a legal practitioner. These types of powers were used by the court in Gribble [2025] NSWDC 344. Importantly, the court in that case considered the plaintiff to have had knowledge of the restrictions on the use of GenAI in NSW, saying at [38]: “Although a litigant in person, the plaintiff is well aware of his obligations in relation to the use of Gen AI, as he swore an affidavit in these proceedings in which he stated that he had not used Gen AI” (at [38]). Later, however, the plaintiff conceded the GenAI use.
4.4. Continuing relationship
The fourth concern arises from the view that Responsive Regulation works best when there is a continuing regulatory relationship. Continuous encounters provide more opportunities to build the relationship, understand the different motivations, backgrounds and objectives of those they regulate, as this will inform the choice of regulatory response and its likely success (Kingsford Smith, Reference Kingsford Smith2011, pp. 708–709, 711). It has been observed that this focus on the interpersonal imbues the theory, and by extension the regulator employing the theory, with “great dynamism and context-sensitivity” because it can respond to the specific circumstances that arise (Ford, Reference Ford2013, p. 16). In short, a more tailored and measured response is possible. In the litigation, context regulator interactions between the judiciary and barristers or advocates may be common. It may be less so for solicitors, but much may turn on the particular jurisdiction. Lawyers and judges are also likely to have a shared background in terms of education and understanding of the judicial system and its rules and requirements. For SRLs who are typically, in the words of Marc Galanter, one-shot litigants, there is likely to be no ongoing relationship (Galanter, Reference Galanter1974, p. 114), and they lack the “advance intelligence” as to how litigation, and indeed particular judges, tend to address litigation before them (p. 98). Consequently, the court is better placed to communicate its expectations to lawyers than SRLs because they have more interactions with the court. There will be exceptions where a lawyer conducts litigation in an unfamiliar court or the lawyer’s practice does not usually include litigation. But the lack of repeated interactions between an SRL and the court system marks another difficulty in effectively regulating SRLs. General deterrence is lessened as SRLs are not a collective and do not share information in the same way that the legal profession does among its members.Footnote 58 Further, SRLs are very diverse, and have widely differing motivations for, and levels of knowledge of, both representing themselves and using GenAI. Where English is not a first language, or even where it is, the use of GenAI as a tool for translation or improvement of written material may be logical.Footnote 59 Further, as GenAI use proliferates, people may be used to using it as an aide in both their work and personal life, such that use becomes normalized for them in other contexts.
4.5. Consistency of response
A final and common criticism of Responsive Regulation is that it is not designed to achieve consistency in enforcement. The enforcement pyramid is a dynamic model which necessarily needs to adapt to the particular circumstances that the regulator faces in relation to an individual regulated entity. As a result, there may be nuances to instances of non-compliance which mean that a different response is needed in relation to different cases that superficially look similar (Braithwaite, Reference Braithwaite2002, pp. 29–30). Moreover, the message from regulator and regulated entity may be unclear or confused due to different approaches being taken in response to particular circumstances (Baldwin & Black, Reference Baldwin and Black2008, p. 63). As a result, it can be difficult for a regulated entity to know how the regulator will decide what level of the enforcement pyramid to utilize. Courts are able to achieve a degree of consistency through rules and practice notes, but ultimately it is an individual judge addressing the specific misconduct that is before them. One judge may strike out materials with fallacious material that was prepared with GenAI and take additional steps such as costs orders and reprimands. But another may identify the false material, disregard it and then simply move on to address the matter before them. Differing responses could be due to relevant differences, but also because of different decision-makers. The goals of education, compliance and deterrence may be hampered by inconsistency. An SRL who is short on time and money may be more likely to take the risk of relying on GenAI if there are no, or minor, repercussions. Thus, consistency may not be achieved, or at least take time to achieve, as a plethora of individual results, perhaps with appellate guidance, accumulate to develop an approach to various forms of misconduct. However, while this is a drawback of Responsive Regulation and the enforcement pyramid, the theory nonetheless provides a framework for judges to consider and utilize in determining what an appropriate response is. Moreover, the highlighting of the potential for inconsistencies and related ramifications also makes judicial officers aware of the importance of how they address GenAI misuse.
5. Conclusion
The GenAI genie is now out of the bottle, and return to an earlier state of affairs is not possible. For regulation to be effective, effective enforcement must occur. As we set out in Section 2, looking at the approach of courts in Australia, Hong Kong, New Zealand and Singapore, we can see considerable divergence of approaches in terms of what controls (if any) the courts have placed on GenAI use in litigation; the targets of regulation; and the response of individual judges, magistrates or tribunal members in the decisions they are making. It is worth noting, too, the differences in the general approach to AI regulation existing in the different jurisdictions we have considered. Arguably, Singapore’s broader governance approach of “responsible optimism” – promulgating consensus-based standards for AI (Allen & Loo, Reference Allen, Loo, Zou, Poncibò, Ebers and Calo2025) – easily melds with the response of its courts in seeking to guide responsible GenAI use by litigants (Singapore Courts, 2024). A whole-of-population approach to education about AI, and harnessing GenAI to support litigants (Singapore Courts, 2025), may also be contributing. In contrast, Australia is moving toward AI regulation at a slower pace, to date marked by the adoption of voluntary standards, some sector-specific requirements but no dedicated AI regulation (Herbert Smith Freehills Kramer, n.d.). Courts across its various federated jurisdictions have implemented variegated guidelines dealing with GenAI use. Meanwhile, New Zealand may have benefited from raising awareness among the legal profession and implementing its court guidance early on (New Zealand Law Society, 2023). Neither New Zealand nor Hong Kong (both small jurisdictions) yet has any reported cases of misuse involving legal practitioners.
These divergent responses, at least where occurring within a single jurisdiction such as Australia, are problematic from a regulatory standpoint. It is our argument, then, that Responsive Regulation provides a useful means of conceptualizing courts as regulators in this space; and of systematizing responses, via the pyramid, based on different indicia. While this does not overcome fragmentation, it illustrates the escalating responses available to the courts and which might be suitable in which circumstances, noting the influence of wider governance cultures. It also enables insights into the strengths and weaknesses of different regulatory approaches. For example, the approach adopted in NSW, which leans toward prohibiting GenAI use, may irritate those actors who feel that GenAI can be used responsibly and that they are being unnecessarily prevented from doing so. Meanwhile, permissive approaches may undermine the utility of regulation if transgressions are not met with an appropriately severe response, as this increases a sense among regulatees that some are “getting away with it” (Gunningham, Reference Gunningham, Baldwin, Cave and Lodge2010 citing Shapiro & Rabinowitz, Reference Shapiro and Rabinowitz1997).
The escalating responses of courts as regulators are described in Section 3 by way of two pyramids of enforcement mechanisms. We showed that methods which may be effective in responding to lawyers may not work for SRLs, with lawyers more likely to behave “rationally,” to have an interest in upholding their obligations to the court and be susceptible to shaming and professional disciplinary mechanisms as impacting on livelihood. Media and professional interest in cases in which lawyers have misused AI promotes general deterrence. In short, there are a range of tools that may be used in relation to legal practitioners, who are not the major problem for the courts when it comes to GenAI misuse. Meanwhile, for SRLs who will struggle to comply, the courts’ only option (unlike the regulators studied in other settings) is to increase education and advice options; and beyond this, tightly controlled case management to try and compel compliance. Mechanisms such as imposition of costs may go some way to achieving redress for the other party or specific deterrence, but are very unlikely to have a general deterrent effect. As we discussed in Section 4, it may be that additional higher layers of the pyramid, or regulatory levers, such as striking out claims or relying on contempt, need to be invoked to address the problem. However, denying access to the courts or imposing sanctions that create insurmountable barriers that result in a lack of access is to be avoided unless the conduct is truly egregious. Calibration of the response is key. Misuse of GenAI by SRLs also points to a much wider issue about support for those representing themselves in legal proceedings and the need for courts, as regulators, to draw on the assistance of third parties to achieve this.
Courts around the world are grappling with, and will continue to grapple with, the appearance of material generated by AI in matters coming before them. Although novel, thinking of courts as regulators of their own proceedings is, we argue, a helpful way of systematically characterizing the possible responses to GenAI misuse. Building on Responsive Regulation, as many others have done, also paradoxically illustrates where additional resourcing and a change of approach are necessary.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/cfl.2026.10056.
Author Contributions
CRediT Taxonomy Felicity Bell Writing – original draft–Equal Michael Legg Writing – original draft–Equal Vicki McNamara Writing – original draft–Equal
Funding statement
No funding to declare.
Competing interests
No competing interests to declare.
Open access
