Chief Justice Sundaresh Menon: Speech at the Australian Judicial Officers Association Colloquium 2026

1. Good afternoon. Let me begin by associating myself with Magistrate Jacqui Payne’s wonderful Acknowledgement of Country and of the traditional owners of the land on which we meet, and by paying my respects to their Elders past, present and emerging.

2. It is a real pleasure to join you in the lovely city of Brisbane for the Australian Judicial Officers Association Colloquium. The ties between our judiciaries are long-standing and warm, and I have benefited greatly from our many exchanges over the years. I speak to you in the year when the Singapore Judiciary celebrates the Bicentennial of its establishment with the issuance of the Second Charter of Justice in 1826. Our celebrations to mark that event will culminate in a number of significant events in November, and as part of that, I have undertaken to deliver a series of five lectures on the rule of law. I was therefore especially grateful to the Association for the gracious invitation to address you today, this being the second in that series of lectures. And it is a contribution to a subject that is moving, with astonishing speed, from the horizon into the daily work of judging.

3. I should begin with the elephant in the room. Some of the most prominent voices in the field have cautioned that advanced AI systems could pose risks of an existential order. Anthropic’s leaked prospectus contemplates that advanced AI could resort to self-preserving conduct, including deception and blackmail.⁽¹⁾ My decision not to focus on such risks today is not to diminish their gravity. They demand an important conversation about AI safety among governments, technologists and society at large. For judges, however, there is a more immediate debate to be had about the impact of AI on law and justice – a transformation that is already well under way, and one that I suggest we have a direct responsibility to shape.

4. That debate is inseparable from our responsibility for the rule of law, because AI bears on it in at least three ways.

5. First, AI is already transforming the legal profession. That is hardly surprising, because law is predominantly a text-based endeavour, and the generation and manipulation of text is precisely where generative AI excels. The extraordinary scale of investment in AI reflects the expectation that it will transform knowledge work, and legal work will plainly be part of that transformation.⁽²⁾ Because the legal profession is integral to the administration of justice, changes of this magnitude will inevitably have consequences for the rule of law.

6. Second, AI bears directly on the legal method. The legal endeavour has traditionally been carried out through a particular body of principles and techniques for interpreting, developing and applying the law. Yet, generative AI can produce seemingly plausible legal answers without employing that method at all. That raises distinct rule of law concerns, because legal authority depends not only on the conclusions reached, but also on the methods by which they are derived and justified.⁽³⁾

7. Third, the implications of AI for the rule of law come into perhaps the sharpest focus in the work of the judiciary. Chief Justice Gageler observed last November that the challenges posed by AI are existential and that the Australian judicature “has no option but to engage”.⁽⁴⁾  The same is true of all our judiciaries. Courts bear a particular responsibility for administering a justice system that gives practical expression to the rule of law. The introduction of AI into judicial work bears directly on how that responsibility is discharged.

8. AI has already entered the justice system through the parties, the profession and the tools used to administer our courts.⁽⁵⁾ As Chief Judge Brian Devereaux mentioned in his introduction, we are well past the question of whether judges should engage with AI. What we should be asking is how far that engagement should extend, for what purposes, and on what terms. How should we harness so formidable a technology without mistaking capability for infallibility or compromising the integrity of the judicial process? This is my specific focus in this address.

9. The answer, I suggest, must turn on the particular judicial function for which AI is deployed. The closer that function comes to the exercise of judgment or the justification of a decision, the greater the wisdom in restraint.

10. I will develop this thesis in three parts. First, I will identify three requirements of sound adjudication, which provide the ends against which the means of technology can be judged. Second, I will consider the adjudicative process and the role that AI might properly play at the four broad stages of that process. Third, I will turn from principle to practice and examine the choices that courts will have to make about questions of design, governance and supervision.

I. What we ask of adjudication

11. Let me begin with what adjudication is for. At its most basic, adjudication is concerned with the authoritative resolution of disputes. But arriving at an outcome that carries the force of law is not necessarily where the judicial task ends. Our decisions will typically disappoint at least one party. What we must expect, nonetheless, is that even the losing party will accept that the decision was reached fairly.⁽⁶⁾ That willingness to accept an adverse outcome ultimately rests on the degree of trust that is reposed in the courts.⁽⁷⁾

12. The link between trust and acceptance is well-established. Professor Tom Tyler of Yale Law School spent years studying why people obey legal authority. In one well-known study, he and his colleagues interviewed some 1,500 Chicago residents about their encounters with the police and the courts. They found that compliance was driven less by fear of sanction than by a belief in the legitimacy of those institutions. That belief, in turn, depended less on the outcome than on their experience of the process: whether they had a voice, whether the decision-maker was believed to be neutral and trustworthy, and whether they were treated with respect.⁽⁸⁾

13. There is, of course, another view that is more attractive to those who view judging as a science. On that view, what matters most is the quality of the output. Professor Eugene Volokh has argued that if a machine can consistently produce opinions as cogent and persuasive as those of a competent judge, it should be regarded as a competent judge, much as a system that produces reliable diagnoses might be regarded as a good diagnostician.⁽⁹⁾

14. I suggest that the latter account might be incomplete. If judging were only about producing the correct outcome, much of what courts do would seem needlessly elaborate. We sit in open court when it might be quicker to decide on the papers. We afford parties a full opportunity to be heard, even when their submissions stray or revisit familiar ground. And we give reasons, often after considerable reflection and revision, when the order could have been stated in a line. These are not mere rituals, inefficiencies or inconveniences to be engineered away. They are an essential part of what clothes a judicial decision with authority. Just as a patient expects more of a physician than an accurate diagnosis, society expects far more of its judges than oracular answers when it entrusts them with power over the liberty and livelihoods of others.

15. That perspective can help us crystallise what sound adjudication requires in at least three key respects. The first is judicial integrity, which encapsulates both the essence of the judicial task and also the character of those who hold judicial office and their abiding commitment to the judicial calling. It is reflected in the generally solemn formality of the process, and as far as the judges are concerned, in their independence, their pursuit of professional excellence and their devotion to achieving justice in every case.⁽¹⁰⁾ The second is procedural fairness, which concerns the quality of the process by which judicial decisions are reached. Those who come before a court should have a fair opportunity to present their case; but, more than that, they should be acknowledged as persons whose concerns deserve attention and respect.⁽¹¹⁾ 

16. But integrity and fairness, indispensable though they are, are not sufficient. Professor Volokh is right that results matter. People come to court to have their disputes resolved – and resolved correctly. Sound adjudication therefore depends on a third requirement, which I will call institutional effectiveness.⁽¹²⁾  A justice system must produce outcomes that are generally correct, and it must do so in a way that is accessible, affordable, reasonably expeditious and generally reliable. If justice exists in principle but remains beyond reach in practice, the courts would be little more than a mirage in the desert. 

17. Much of the case for the judicial application of AI is a case about institutional effectiveness. Its attraction lies in its scale, its speed and its baseline reliability – it seemingly enables courts to serve more users, more quickly, while continuing to produce outcomes that are generally correct. In systems where justice needs greatly outstrip judicial resources, typically manifesting in persistent, intractable backlogs, this capacity to operate at scale may be transformational. I have suggested elsewhere that the real comparison in such cases is not between human adjudication and technologically mediated adjudication, but between a sufficiently workable form of justice and no justice at all.⁽¹³⁾ Such cases raise important questions about access to justice, but they are not the principal setting I want to examine today. My focus is instead on the use of AI in what may be regarded as the mainstream of adjudicative work: where existing judicial processes currently function reasonably well, but where AI could conceivably enable those processes to work better or at least as well even in the light of the growing complexity of legal work.

II. AI across the adjudicative process

18. To answer that question, we need to look inside the composite activity of adjudication and disaggregate its working parts.

19. For this second part of my address, I will adapt a framework developed by Raja Parasuraman and his colleagues in their study of human interaction with automated systems.⁽¹⁴⁾ They identify four functions that I think we can broadly map to the principal stages of adjudication:

(a) The first is information acquisition and management: I map this to gathering, sifting and organising factual and legal material. 

(b) The second is information analysis: or, in our context, interpreting the law and applying it to the facts. 

(c) The third is decision selection: or identifying the range of available outcomes and choosing the one that law and justice require. 

(d) The fourth is decision communication: which would cover explaining and justifying the result to the parties and, typically also to the public.

20. This framework is, of course, a simplification. Chief Justice Niall observed in an address last year that these stages are connected and form part of a single judicial task.⁽¹⁵⁾ That is undoubtedly correct. Judging is not an assembly line, and these functions cannot always be parcelled into neat compartments. But the framework remains useful, because it enables us to distil the values engaged by each function and to expose the trade-offs that attend different uses of AI.⁽¹⁶⁾

21. The broad pattern I suggest is this. As we move from locating and organising material to exercising and explaining judgment, the case for AI progressively weakens. It is most compelling where AI is used to acquire information, more qualified where AI analyses that material or proposes outcomes, and most limited where it formulates and articulates judicial reasons.

A. Information acquisition and management

22. Let me begin with information acquisition and management, where the appeal of technological assistance appears most immediate. Most judges will have encountered the bundle that runs to thousands of pages, or the convoluted affidavit that generates more heat than light. Through a mass of information that may be voluminous, fragmented or heavily contested, the judge must trace each issue through the evidence, reconcile conflicting accounts and piece together a coherent picture of the dispute. This work is indispensable, but much of it is also laborious.

23. To manage this complexity, judges and advocates have long deployed tools such as chronologies, summary tables and lists of issues. One aspect of what generative AI could do is to supercharge these familiar forms of assistance.⁽¹⁷⁾

24. It could do this in at least three ways. First, it could shorten the time needed to digest a case by bringing structure and clarity to a mass of material, leaving more time for more complex aspects of decision-making. Second, it could improve the quality of the adjudicative process by bringing the real issues into sharper focus. Third, in highly complex litigation, it may go beyond enhancing efficiency to addressing a deeper problem of institutional capacity. Let me say more about the second and third points.

25. I can illustrate the second point with Singapore’s Small Claims Tribunals. The tribunals receive about 10,000 claims each year, a substantial number for a jurisdiction of our size. These claims go up to about SGD 20,000, or SGD 30,000 if the parties consent, which is about AUD 34,000. Parties appear without lawyers and typically present their cases as life has generated them: through receipts, screenshots, emails and message threads, often in large quantities and with no obvious structure.⁽¹⁸⁾ 

26. In December 2024, we introduced AI-assisted translations of key court documents into Chinese, Malay and Tamil, the three other official languages of Singapore. In March 2025, we extended this to documents filed by the parties. Then, in September 2025, we introduced a document-summarisation tool for tribunal magistrates and made it available to the parties soon after.⁽¹⁹⁾ In time to come, we hope to extend such assistance to the preparation of claims and defences, the organisation of evidence and the consideration of settlement options.⁽²⁰⁾

27. The significance of these tools goes beyond efficiency. For the self-represented party, they make the material easier to understand and the process easier to navigate. For the magistrate, they provide a clearer starting point for identifying the real dispute and the matters requiring clarification. That is especially important because the magistrate plays a significant role in eliciting the facts and arguments needed to resolve the dispute. Translation and summarisation tools therefore do more than save time: they improve the conditions in which sound decisions may be made and procedural justice achieved.⁽²¹⁾

28. Turning to the third advantage of AI that I have alluded to, this is most evident at the other end of the spectrum, in cases whose scale tests the capacity of the court. Digital evidence, technical proof and dense legal frameworks have greatly enlarged the informational burden of adjudication in recent decades.⁽²²⁾ The Bell Group trial in Western Australia occupied 404 hearing days, involved more than 86,000 documents and culminated in a judgment of more than 2,600 pages.⁽²³⁾ The Roberts-Smith proceedings in the Federal Court ran for about 110 hearing days and produced reasons exceeding 700 pages.⁽²⁴⁾ In litigation of this order, AI may be needed not merely to save time, but to help the judge maintain command of a record that would strain ordinary methods of review. This is not simply a question of easing the burden on the individual judge, but of enabling the judge to manage the material effectively, and beyond this, there are systemic concerns because a case that consumes a disproportionate share of judicial resources inevitably limits how effectively other litigants are served.⁽²⁵⁾

29. My starting position, therefore, is that AI holds its greatest promise for information acquisition and management, and that courts ought to actively explore its use. As Justice Michael Lee of the Federal Court of Australia forcefully argued in the 16th Sir Harry Gibbs Memorial Oration, “there is … little sense in insisting that human beings continue to perform, at enormous expense, tasks involving no [human] judgment which machines may now be able to perform incomparably more efficiently”.⁽²⁶⁾ I agree but would add that the case for embracing these tools does not require us to overlook their present limitations. Let me name three of these.

30. The first is that organisation and summarisation are never neutral. Every summary reflects judgments about what matters and what does not.⁽²⁷⁾ Studies of machine-generated legal summaries have consistently found at least some incidence of material omissions, unsupported statements and/or distorted emphasis.⁽²⁸⁾ More crucially, the process of organising and summarising information is itself an integral part of how judicial thought develops. Each reading generates new questions, which refine the inquiry and return the judge to the material with a renewed understanding of the case.⁽²⁹⁾  What is gained in speed at the outset may be lost in the quality of the analysis that follows.

31. A second limitation is that an AI system can work only with what has been captured and rendered into a digital form it can process.⁽³⁰⁾ Some matters never become data at all. A transcript may document words that were spoken, but it cannot fully convey how the evidence was given or the understanding that a judge accumulates over the course of a long trial.⁽³¹⁾ When complex human events are rendered as data, the individual pieces may remain, but the context and relationships that give them meaning may be flattened or lost.⁽³²⁾

32. A third limitation concerns the hearing. While this may come within the stage of information acquisition and management, a hearing is not a one-way transfer of information, least of all where a party is self-represented. The judge may need to explain an unfamiliar process, test a position that has not been clearly articulated and ensure that the other side has a fair opportunity to respond. Through that exchange, the judge’s own understanding develops, the parties are placed on a fairer footing, and they come to know that they have been heard by the person who will decide their case. AI may help prepare us for that exchange, but it cannot conduct it in our stead.⁽³³⁾

33. These constraints point to three features that should shape the design of AI tools. The first is grounding. A tool confined to an authoritative and ring-fenced body of material is less likely to generate unsupported output.⁽³⁴⁾ The second is traceability. Every synthesis should lead the judge back to the underlying material, so that the output serves as an initial map of the record rather than a substitute for it.⁽³⁵⁾  The third concerns scope. The tool may assist the judge in working with the record, but the hearing and engagement with the parties must remain the judge’s responsibility.

34. The importance of these design choices was brought home to me during a visit last November to the Shenzhen Intermediate People’s Court.⁽³⁶⁾ There, about a dozen technologists working with about a hundred judges identified discrete judicial tasks for which AI might assist. Within a single platform, a judge is able to review the case materials by issue, trace summaries to the underlying sources, detect inconsistencies in a party’s position and consider matters that the system flags for further clarification. At key stages, however, the system will not proceed until the judge has decided, for example, which authorities to rely on or which submissions to accept. Only then does the platform offer more advanced analytical and drafting assistance, drawing on a curated corpus of legal materials.

35. The reported results have been encouraging. In 2025, case disposal rates reportedly rose by 50% while appeal rates were halved. The Supreme People’s Court has developed a system that draws on key features of the Shenzhen model and piloted it across 23 courts and 98,000 cases in the first seven months of 2026 alone.⁽³⁷⁾ This demonstrates, in concrete terms, how a justice system could use AI to enlarge a judge’s capacity to work with the record while preserving responsibility for the steps that call for judgment, and do so in a manner that is acceptable to that justice system in the light of the constraints that it must contend with.

B. Information analysis and decision selection

36. Let me turn to the second and third functions: information analysis and decision selection, which I take together because they often overlap in practice. 

37. Studies suggest that AI systems can perform strongly, if unevenly, on legal-reasoning benchmarks and can improve the speed and quality of human performance on some legal tasks.⁽³⁸⁾ But capability alone does not determine suitability. Using AI for information analysis and decision selection raises two distinct sets of concerns. The first arises from technology’s present limitations, which may diminish, and I think will diminish, as the technology matures. But the second is more fundamental because it inheres in the nature of adjudication itself, and would remain even as AI systems grow more capable.

38. Let me begin with the technical concerns. One difficulty stems from the material on which a model is trained. A judicial precedent is not a comprehensive textbook exposition of the law. It is a response to a particular case presented by the parties, the evidence they adduced and the path that litigation took.⁽³⁹⁾ An argument may be discussed at length because counsel pressed it, and not necessarily because it was on the critical path of analysis. Equally, a judgment may say nothing about a potentially decisive point because it was never pleaded or argued.⁽⁴⁰⁾ A lawyer or judge reads a judgment with an understanding of this path dependency and can be expected to interpret both the statements and the silences accordingly. By contrast, a model that is trained principally on the resulting text may lack such tacit knowledge and fail to reconstruct that context reliably.⁽⁴¹⁾

39. Nor is the performance of these systems entirely stable. AI systems are said to be brittle, because small changes in the wording or structure of a prompt or problem can alter the answer, even if the underlying task remains the same.⁽⁴²⁾ They are also bounded, because performance may fall sharply as complexity increases.⁽⁴³⁾ To be sure, experience is developing and some of these claims are contested. But it is safe to say that AI operates across a jagged frontier: highly capable in some areas, unexpectedly weak in others, and not always capable of telling us where along the frontier a particular task lies.⁽⁴⁴⁾

40. But suppose that all these technical limitations are eventually overcome to a large and acceptable extent. Even then, more fundamental concerns about the nature of adjudication would remain. 

41. The first concerns judicial accountability. A language model produces an answer by identifying and extending patterns learned from data. Its output may be cogent and coincide with a result that a judge might reach. But an answer that looks reasoned is not necessarily the product of the legal method.⁽⁴⁵⁾ Nor is a model’s explanation of its answer necessarily a reliable account of how that answer was generated.⁽⁴⁶⁾

42. The legal method matters because it forms part of the terms on which the parties submit to be judged. A judge is expected to reason consistently with doctrine and precedent, to give effect to the purposes of the law and to operate within the constitutional limits of the judicial office.⁽⁴⁷⁾ Judicial accountability requires that the judge’s reasons can be understood, tested and, if necessary, corrected on appeal. A conclusion is not enough. It must be reached by a process that the law recognises as legitimate.

43. Moreover, many cases require more than the extrapolation of past data. They call for an exercise of judicial creativity: in developing doctrine, in applying open-textured standards, and in fashioning remedies that are responsive to the realities of the case. Let me take these in turn.

44. Under the common law tradition, judges are entrusted with a limited law-making power and must sometimes decide whether, when and how the law should develop.⁽⁴⁸⁾ That calls for judgment about principle, policy and practical consequences; about institutional competence; and about the wisdom of leaving a question open for another occasion.⁽⁴⁹⁾ Professor Richard Susskind captures the difficulty when he asks whether a case like Donoghue v Stevenson might pass unnoticed through an automated court, because the system failed to recognise the deeper question of principle concealed within an apparently ordinary dispute.⁽⁵⁰⁾

45. More frequently, judges must apply open-textured standards of what is reasonable, what is proportionate, what lies in a child’s best interests, or what justice requires in the circumstances.⁽⁵¹⁾ These questions require the judge to bring doctrine into contact with the realities of the world, and to weigh competing interests that cannot always be reduced to a common or quantitative measure.⁽⁵²⁾

46. There is a dimension of judicial integrity that cuts across these points. When judges develop the law or exercise discretion, they do so in the name of the community and are accountable for what they decide. They are subject to the same law, bound by the standards they apply and answerable through their reasons and appellate review. A machine does not stand in the same reciprocal relationship with the community.⁽⁵³⁾

47. I alluded to a third form of judicial creativity, which lies in how remedies are fashioned. A useful illustration may be found in Singapore’s family justice system, which embraces the idea of therapeutic justice as its overarching philosophy. Where the law touches a family, the court’s task is rarely confined to declaring who is right. So far as the law permits, the court should chart a course that enables the family to move forward and to heal.⁽⁵⁴⁾ That calls for practical wisdom, empathy and an appreciation of how the decision will be experienced by those it affects. These inescapably human dimensions of the judicial task might look grossly out of place in a computerised system, no matter how artificially intelligent and capable it might seem to be.

48. None of this means that AI cannot have a legitimate and supportive role in some aspects of decision selection. Many disputes do not require the development of doctrine or the exercise of broad remedial judgment. Take claims for unpaid invoices, tenancy deposits and minor motor accidents. These typically involve the application of settled principles and some arithmetic to largely undisputed facts.⁽⁵⁵⁾ They also tend to be high in volume and relatively modest in value, which makes institutional effectiveness especially important.

49. For such disputes, there may be scope for AI to generate indicative, non-binding assessments that the parties could accept or reject, while retaining the right to proceed to a human decision-maker.⁽⁵⁶⁾

50. Such a form of digital neutral evaluation is less novel than it may seem. The State Courts of Singapore have long offered early neutral evaluation, and online marketplaces resolve millions of disputes each year through technology-assisted processes.⁽⁵⁷⁾ Even where the parties do not accept the indication, an early, objective and inexpensive assessment of the merits can help to temper expectations and provide a common basis for informed negotiation.⁽⁵⁸⁾ Sir Geoffrey Vos, who recently delivered the John Lehane Memorial Lecture in Sydney, predicts that as litigants come to have more confidence in machine-made decisions, they “may well choose to have their cases determined in minutes by a machine at little or no cost rather than waiting months or years to receive the decision of a human judge”.⁽⁵⁹⁾

51. I should emphasise that the case for AI neutral evaluation cannot rest simply on the truism that human adjudication is expensive while a machine-generated indication is cheap.⁽⁶⁰⁾ These disputes matter greatly to the people involved, and a court-supported process must uphold standards of fairness and accountability. But a more principled justification, I suggest, lies in the requirement of institutional effectiveness and the scale of the justice gap. The World Justice Project estimates that more than five billion people have unmet justice needs, while many everyday legal problems fall within recurring categories governed by settled principles.⁽⁶¹⁾  As I observed earlier, for many of the people behind those numbers, the real choice is not between an AI process and a human judge, but between some reliable process and none at all. That need cannot realistically be met simply by appointing more judges, training more lawyers or overworking both groups. In this context, technology holds much promise as a force multiplier that can amplify the reach of existing institutions beyond their finite resources.⁽⁶²⁾

C. Decision communication

52. I turn finally to the fourth stage of decision communication, where the case for caution, I suggest, is at its strongest.

53. Reason-giving speaks to at least three constituencies. First, reasons speak to the parties and honour their dignity as participants in the judicial process. They provide the clearest assurance that the court has listened to their case and taken it seriously, which is a necessary counterpart to the right to be heard.⁽⁶³⁾ For the losing party, reason-giving acknowledges the case advanced and may enable that party to understand, and eventually come to accept, a decision that is bound to disappoint.⁽⁶⁴⁾

54. Second, reason-giving is how the Judiciary accounts to the public for the exercise of judicial power. Reasons expose the court’s decision to scrutiny, and beyond the immediate case, they guide future conduct, articulate the community’s values and contribute to the development of the law.⁽⁶⁵⁾ 

55. Third, reason-giving imposes discipline on the decision-maker. The rigour of judicial writing is an essential safeguard against error – what Sir Frank Kitto described as “an effective stimulant to judicial high performance”.⁽⁶⁶⁾ Every judge knows why. Writing compels us to work through the case systematically, confront weaknesses in our reasoning and make what is implicit explicit. It is not merely the expression of an analysis already concluded, but often an integral part of how the analysis is done.⁽⁶⁷⁾

56. These factors explain why I am wary of using AI to generate the substance of judicial reasons. For one, the reasons may have no authentic connection with the actual process by which the judge reached the conclusion. Professors Grimmelmann, Sobel and Stein describe this as a “Potemkin interpretation”.⁽⁶⁸⁾  Those words recall the villages that Prince Grigory Potemkin was said to have erected along Catherine the Great’s journey through Crimea. In reality, those villages were nothing but elaborate facades intended to create an appearance of prosperity while concealing emptiness behind them. The story may be apocryphal, but the metaphor is apt. An AI-generated judgment may display fluency, structure and apparent coherence, yet it may bear no relation to the judge’s actual reasoning.⁽⁶⁹⁾ The fact that a judge may be prepared to sign off on a draft is not enough. Judicial reasons are not merely the words the judge is prepared to accept. They are the judge’s public account of why judicial power has been exercised on a particular occasion in a particular way. If the reasoning is adopted only after the conclusion has been reached, the proper order of things is reversed. The reasons may be seen as little more than a post hoc justification, even if the judge might have reached the same result independently.

57. Even where an AI-generated explanation happens to accurately reflect the judge’s reasoning, it may still fail to serve the wider purposes of reason-giving that I have outlined earlier. In some cases, the manner in which the court speaks to those affected matters as much as the decision itself. A custody decision may irreversibly alter the course of a child’s life. A substantial custodial sentence may mark one of the most consequential moments in the lives of the offender and the victim. The communication of such decisions calls for care, sensitivity and a deep understanding of the human condition.⁽⁷⁰⁾ Similarly, where a judgment develops the law on a matter of public importance, the public expects the court to confront the difficult choices involved and account for how it resolved them.⁽⁷¹⁾ As far as I am aware, AI cannot presently discharge these distinctly judicial responsibilities. 

58. This does not mean that AI must be excluded from every part of the judicial writing process. It could assist in at least three limited ways. First, it may serve as a critical reader or sparring partner that flags unanswered submissions, logical gaps and adverse authority. Used in this way, it could help to prompt rather than replace further inquiry,  as long as it is used in a manner that comports with the requirements of natural justice. Second, it may assist with drafting certain discrete parts of the judgment such as the procedural history or uncontested background facts, provided the judge has first considered that material independently and proceeds to check the draft carefully against the record. Third, it may assist with routine decisions for which extensive reasons are neither necessary nor expected, such as some interlocutory rulings, routine costs orders and awards whose basis is apparent from another source. 

59. The controlling principle, I suggest, is that AI might assist with tasks surrounding judicial writing, but it should not take over the writing through which the judge reasons, justifies and assumes responsibility for the decision. The closer AI comes to speaking in the court’s authoritative voice, the stronger the case for restraint.

III. From principle to practice

60. Let me turn to the third and final part of my address, on translating these principles into practice. The framework I have outlined could guide us in thinking about which uses of AI may be acceptable in principle when thinking about the judicial task. However, it does not tell us whether a particular system is safe, or how it should be introduced into our courts. A sound use can be rendered unsafe by the design of the tool, the structure of the workflow or the habits of its users. Drawing the boundary is therefore only half the task. The other half is about the governance.

61. When a machine joins the work of judging, it does not simply remove one part of the adjudication task. Automation is not simple subtraction. It changes the work that is left to the judge and how the judge approaches that remaining work.  This is why the assurance that a human is “in the loop” tells us very little by itself, and similarly why the slogan that AI should “support” and not “replace” human judges also in my view tells us very little. Human review may become perfunctory if confidence in the machine’s apparent reliability displaces the judge’s own assessment.  What really matters is whether the judge retains meaningful control. For example, does the AI system grant the judge liberal access to any and all assistance on demand, or is deliberation and friction built into the process - as I suggest it should be? How easily can a judge interrogate the system’s output, depart from it or suspend the system’s use altogether? 

62. I suggest that there are at least three risks of human-machine collaboration that deserve particular attention. The first is over-reliance. A recent MIT report has warned of “cognitive surrender”, or a tendency to turn to AI at the first sign of difficulty instead of working through a problem independently.  A related challenge is vigilance decrement or decrease. The problem is familiar to users of autonomous vehicles: as the system becomes more reliable, the human is expected to watch for increasingly rare failures. Yet this very reliability is precisely what makes sustained attention harder.  The troubling implication is that the more dependable a tool becomes over time, the less effectively it may be supervised.

63. The second risk is skills fade. A study published in The Lancet found that endoscopists who had become accustomed to using AI to detect polyps performed less well when AI assistance was withdrawn.⁽⁸⁰⁾ The evidence in medicine is still developing, but the concern can be readily understood. The skills involved in mastering a case and crafting compelling reasons are sustained only through regular use. If AI becomes a shortcut around intellectual effort, we may erode not only the expertise of experienced judges but also the very pathways by which newer judges will acquire those skills.⁽⁸¹⁾ Indeed, this concern is especially acute for a generation of judges and lawyers who will enter professional life having only known a world with AI. For them, the danger may not be that hard-won skills will fade, but that they will never be fully formed.

64. The third risk is hidden institutional delegation. Attention must be given not only to the relationship between the judge and the AI system, but also to those who design, procure and configure the systems on which judges will come to rely. Developers and vendors make choices about the data used, the objectives pursued, the feedback the model receives, and the manner in which it responds to difficult or edge cases. Those choices may be embedded in ways that the judge cannot easily alter, inspect or even understand.⁽⁸²⁾ Where that occurs, part of the court’s interpretive and institutional power may shift to persons who neither bear judicial responsibility nor possess the training or experience that ordinarily attend its exercise.⁽⁸³⁾

65. The answer to these risks of over-reliance, skills fade and hidden institutional delegation should not be to retreat from AI, but to govern its use in ways that preserve meaningful human judgment and control. Lawyers are accustomed to speaking in terms of prescription and prohibition, but the art of regulation admits of more nuanced choices. We should design AI systems that judges would want to use, because they see that doing so would make them better judges. And we should design these systems in a way that incentivises judges to use them well, because doing so would reflect their own commitments to judicial integrity and procedural fairness. Let me make five points about what this might entail in practice.

66. First, courts should be more intentional about introducing processes for selecting, introducing and overseeing AI tools. New tools should be sandboxed, piloted and socialised with potential users. New and old tools alike should be monitored and audited, to ensure that they remain aligned with the requirements of sound adjudication. These safeguards are especially important because an AI system can turn a single embedded defect into a systemic source of error.⁽⁸⁴⁾

67. Second, cognitive friction should be built into AI systems to preserve active and independent judicial engagement. In some settings, the judge should be required to engage substantively with the case before seeing the machine’s output – for example, by recording a provisional view before the recommendation is revealed.⁽⁸⁵⁾ Without such friction, the ready availability of a plausible answer may create an illusion of understanding, and encourage the judge to defer to the recommendation instead of interrogating it independently.⁽⁸⁶⁾ The same design principle ought to apply across adjudicative stages: summaries should supplement, not supplant source documents; search tools should distinguish extraction from inference;⁽⁸⁷⁾ and tools used to test or review a judgment should receive the judge’s draft before offering a critique.

68. Third, effective design must be matched by timely training. Judges do not need to become engineers. But they do need enough understanding to be effective judges of these systems and their potential value and their potential dangers. In my experience, AI is most powerful in the hands of users who know enough to ask the right questions, to recognise or be sceptical of a doubtful answer and to know where and how to verify it. It is much less safe when it supplies an answer that the user lacks the knowledge or intuitive sense to question or the means to assess in context. Judges therefore need a working understanding of how these systems work and how they fail, and an awareness of the human tendency towards both over-reliance and over-reaction.⁽⁸⁸⁾ That is the discipline that we are trying to build through the Singapore Judicial College. Our programmes range from primers on the mechanics of AI and prompt engineering, to hands-on sessions on AI-assisted workflows, to design thinking workshops that explore how AI might responsibly support justice-system reform.

69. Fourth, courts will need to consider the question of transparency. Not every use of AI will call for disclosure to the parties. But courts will need a principled approach to deciding when disclosure is warranted, including what form the disclosure should take and whether the parties should have an opportunity to comment on or challenge the use of AI or its outputs.⁽⁸⁹⁾ My tentative view is that the case for disclosure strengthens as the use moves closer to the court’s analysis or decision, and whenever the output has materially influenced the reasons or result.⁽⁹⁰⁾ This is a question, I suggest, to which judiciaries will need to return with more mature consideration as our experience with these issues deepens.

70. Fifth, I firmly believe there is much to be gained from greater dialogue and cooperation among judiciaries. The Singapore courts have sought to foster such exchanges through the Meetings of Chief Justices and Judges in Charge of Technology, which is a platform for the discussion of emerging issues and best practices in the use of technology by courts, and which has benefited greatly from the regular attendance of our colleagues from Australia. The Standing International Forum of Commercial Courts has also established a workstream on AI in dispute resolution, which has seen strong interest in knowledge sharing and practical collaboration. These initiatives point to the value of moving beyond the exchange of experience towards more structured forms of collaboration, such as the development of model frameworks and practical toolkits. We have much to benefit from each other's experiences.

71. Taken together, these safeguards of supervision, design, training, transparency and institutional and transnational cooperation will determine whether AI remains an aid to human judgment or becomes a vehicle for unaccountable delegation.

IV. Conclusion

72. Let me conclude by returning to the three requirements of sound adjudication with which I began. They point us away from two errors. On the one hand, Luddite scepticism may cause us to forgo an extraordinary opportunity to make justice more accessible and our institutions more effective. On the other hand, uncritical enthusiasm may weaken judicial independence and integrity, erode procedural fairness and undermine the trust on which our authority depends. The question is not whether we can use these tools, but whether we have the wisdom to use them well – to capture what AI can add to the administration of justice, to preserve what adjudication cannot afford to lose, and to build institutions capable of governing the use of AI with care, clarity and the capacity to adapt.

73. Artificial intelligence may indeed seem to be highly intelligent, but it is also inescapably artificial. It cannot bear judicial responsibility, and it cannot replicate the human encounter that is at the heart of a fair hearing. Judges, for our part, may only be human. But in adjudication, that is not a deficiency to be minimised. In many ways, it is precisely the point.

74. Allow me once again to thank the Association for the opportunity to contribute to this important conversation. I look forward very much to all your discussions. Thank you.