sg-crest A Singapore Government Agency Website
Official website links end with .gov.sg
Secure websites use HTTPS
Look for a lock () or https:// as an added precaution. Share sensitive information only on official, secure websites.

Justice Kwek Mean Luck: Speech at the 12th International Conference on Judicial Training (IOJT)

12th International Conference on Judicial Training (IOJT)

The Artificial Intelligence Revolution: Strategic Challenges for Judicial Training Institutions

France, 9 November 2026

Justice Kwek Mean Luck, Supreme Court of Singapore


Introduction

1. President of the IOJT, Justice Adele Kent, Deputy Director of the ENM, Monsieur Boulakras, distinguished judges from around the world, ladies and gentlemen. Bonjour and good morning. Let me first thank the École nationale de la magistrature and the IOJT for bringing us together, and for the opportunity to speak here.

2. AI is already changing how judges work, whether or not we plan for it. The question for judicial education institutions is therefore not whether AI will reach our judges, but whether our judges will be ready for it.

3. The organisers have proposed two levels of reflection, and I will take them in turn. At the strategic level, I will speak about how far judiciaries should allow AI into judicial work, who should decide this, and what it means for the mission of judicial education institutions. At the operational level, I will speak about what this means for how we develop our judges. I will set out my view that we should focus on building competencies over knowledge, and in particular, twin competencies, namely core judicial competencies and AI judicial competencies. My central point is this: AI does not reduce the value of the core judicial craft. It raises it.

The limits and the promise of AI

4. What are the limits and the promise of generative AI for judges? To answer this, it would be useful to return to how large language models, or LLMs, operate. Essentially, LLMs operate by predicting, based on probabilities, what is most likely to follow from the patterns in the text they were trained on. I will call this semantic matching. Newer models can now set out a problem in ordered steps. But whether this amounts to reasoning, in the way a judge would recognise it, remains open to debate.

5. This limits how far generative AI can assist us in decision-making. This is why the existing generative AI legal tools are strong in certain areas — classification, summarisation, translation — which suit this kind of capability, and less strong in prediction and judgement, which require some level of logical reasoning.

6. The nature of generative AI’s semantic-based capability affects how effective it can be in particular areas of judicial decision making. First, there are limits to its ability to devise creative and context-sensitive solutions to the particular problems before the court. Judges sometimes need to look beyond the arguments and materials presented, appreciate the practical realities of a dispute, and fashion solutions or remedies that respond to its particular circumstances. Second, as judges we are sometimes called to weigh competing interests or values. There are limits to AI’s ability to properly weigh such interests or values, when it lacks the benefit of a lived life. Third, there are limits to AI's ability to engage in the principled legal reasoning required in hard cases, particularly where established precedents or doctrines do not provide a satisfactory answer. Such cases may require judges to return to first principles, reconcile competing lines of authority, or determine whether and how the law should develop in response to changing circumstances.

7. But that may change in time. With the vast amounts of data being fed to the top AI models, and newer models that set out their working step by step, the day may come – and for some tasks, may already be here – when it appears that generative AI engages in logical thinking, even when its underlying model is that of semantic matching.

8. This is particularly so, when you consider that our human thinking process, is to some extent at least, a product of pattern recognition, of which semantic matching is a type. We learnt the alphabet by pattern — A, B, C. We learnt that one plus one equals two in much the same way, long before understanding why it is so. Tversky and Kahneman’s work on the “representativeness” heuristic showed that people often judge probability by similarity to a prototype – and that this can lead us into systematic error.1 Semantic matching drives more of our real world judgment than we may recognise.

9. If our own reasoning rests partly on a semantic foundation, a machine built on semantic matching may one day appear to reason logically, even in complex areas of analysis. Agentic tools, which can encode specific work methodologies, bring that day closer.

10. As an illustration of how fast AI is changing, consider the issue of hallucination. Two or three years ago it was a key objection to the use of AI tools in legal work. Today, in well-designed tools developed specifically for legal research, the risk of hallucinated citations has been reduced, though not eliminated.2 

11. Let me share, as an example, what has been done by LawNet, a legal research platform for Singapore law. LawNet is run by LawNet Technology Services, which I help to chair. By restricting the source materials to a curated set, engaging in an iterative training process of the AI outputs with the help of legal professionals, and requiring the generative AI replies to include citations (and to own up when it cannot find a citation), LawNet has reduced the risk of inaccuracy from sourcing errors to a very low level.

12. There is a lesson in this. If the limits of the technology are receding, we cannot rely on those limits to keep AI in its proper place. That line must be held by us as judges. So where should it be drawn? How far should judges rely on AI tools? At what point do we devolve our judicial responsibility to that of the AI machine?

13. My view is that, fundamentally, judging is our responsibility. That cannot be devolved to AI. Judicial reasoning is not merely a means to an outcome. It is a substantial part of what makes the outcome legitimate.

14. At the same time, we should not lose sight of what AI can offer. Used well, it can help deliver more timely justice, widen access to justice for litigants in person, and support greater consistency across our courts. The task is to secure these benefits while preserving the judicial values on which public trust rests.

15. Four considerations help to frame where that line is drawn.

16. First, there is a spectrum of disputes, for which the assessment of AI use in decision-making may differ. At one end are high-volume cases involving limited variance in fact scenarios, for which members of the public place a premium on speedy resolution. For example, online consumer disputes below a certain quantum. At the other end of the spectrum are cases involving constitutional issues, or issues of life and liberty. There would be greater public expectation that such disputes are adjudicated by seasoned human judges, applying their minds and judicial reasoning, rather than through automated dispute resolution powered by AI. There is a wide range of case typology in between these two ends, which would need to be examined more deeply as individual categories. Even at the high-volume end, I would expect parties would want to retain a right of appeal to a human judge, so that responsibility for the ultimate outcome still rests with a judge rather than a machine.

17. Second, judging involves many tasks: for example, identifying the relevant evidence, summarising the evidence and arguments, assessing witnesses, and analysing cases and statutes using legal methods and principles. AI may be better suited for some of these tasks over others.

18. Third, it is possible to scope out AI use in terms of assistive modalities and not determinative modalities. For example, we can use AI to check if the evidence cited in a judgment is consistent with the evidence in the materials before a judge. This is particularly helpful when there is a deluge of data or documents in a case. AI could also be used to test and critique an evidential or legal argument, before a judge lands on a particular line of analysis.

19. Fourth, the societal context of each judiciary matters to the nature of the trade-offs being considered. A judiciary operating across a vast geography may see a greater need to use AI to ensure some bounds of consistency of decisions for similar cases across the country. A judiciary faced with very high volumes and the risk of backlogs affecting timely justice also makes different trade-offs in its assessment of how justice is best delivered, compared to a judiciary that does not have such issues. 20. These four points will shape the much-needed deeper discussion on where a judiciary draws its line on AI use. They will, I am sure, be further explored in the course of this conference.

21. This brings me to the question of who should decide where that line is drawn. Rules on AI may come from many sources. In Europe, for example, the EU AI Act classifies AI systems used in the administration of justice as high-risk.3 But within any such framework, how AI is used in judging goes to the heart of judicial independence. The judiciary must have the leading voice. These decisions should not be left, by default, to those who build and sell the tools. Judicial education institutions have a role here too. They are well placed to build, within the judiciary, the understanding needed to make these decisions well.

22. Wherever the line is drawn, one risk is clear: the risk that with AI, judicial decision-making becomes a black box, opaque and therefore unaccountable. A judge who cannot explain a decision cannot truly answer for it. But as AI systems become more capable and more prevalent, the line between reasoning devolved to AI and reasoning assisted by AI will become progressively harder to draw.

23. This has led to the introduction of guidelines in many judiciaries, internally to guide judicial use of AI, and externally to guide the use of AI by lawyers and litigants in person. In Singapore, for example, our courts issued a guide in 2024 on the use of generative AI tools by court users.4 

24. This is necessary, but guidelines can only do so much. Ultimately, this comes back down to the values and the integrity of the judges we appoint, and whom we continue to develop.

25. A judge who understands that the reasoning is an integral part of the judgment, and who feels a sense of responsibility in exercising that judgment, will not surrender it, whatever the capabilities of the AI tool on their desk. A judge who does not may find it all too easy to let the tool do the deciding.

26. Hence, the selection and appointment of judges, and their ongoing education, become central to how a judiciary manages AI well.

27. Casting an eye to the future, there are two risks I would particularly watch for.

28. The first is undetected algorithmic bias in the AI models providing legal assistance. Here, I distinguish between two types of biases arising from the algorithm.

29. First, technical bias. This arises, for example, in how a system is designed to weight its search results, such as whether it would prioritise the most cited cases, the highest authority, or the most recent decisions, or some combination. The weighting affects what users see first. Weighting by citation count entrenches dominant precedents and can bury a newer or dissenting line of authority that may be the future of the law. Weighting by hierarchy can miss developments on novel questions that are working their way up through the lower courts. Weighting by recency can obscure the foundational authorities from which the principle actually flows. Research from Stanford found that general-purpose LLMs perform better in cases that are newer, more salient, and from more prominent legal jurisdictions, which the researchers warned suggests a real risk of legal monoculture.5

30. Technical bias is a concern, but it can be addressed. It results from design choices that can be identified, documented, disclosed and adjusted. In turn, we can interrogate how retrieval is weighted, and we can test the answers.

31. What is of greater concern to me is normative bias. This is the tilt in values that a model absorbs from its training data and from the way it is fine-tuned. Every LLM carries some normative bias. It is a feature of building a system from human-generated text and human feedback. One study, by David Rozado, administered 11 political orientation tests to 24 leading conversational models.6 Most showed a measurable political tilt. The point is not that the tilt runs in one particular direction. The point is that a tilt exists at all.

32. Two further findings concern me more than the direction of any slant.

33. First, bias can be steered. The same researcher showed that models can be steered toward specific points on the political spectrum through fine-tuning with only modest amounts of data.7 Whoever controls the fine-tuning can shift the values in any direction. That is why the question of who procures, fine-tunes and audits the AI tools used by courts is not merely technical. It is a question of judicial independence.

34. Second, models may not see their own bias. Research suggests that what a model reports about its own biases, when asked to reflect on them, can differ markedly from the biases that surface in the content it generates. One study found that the biases models acknowledged explicitly were relatively mild, while their implicit biases showed strong stereotypical tendencies.8

35. If an AI system carries a tilt we cannot see, and we come to rely on it heavily over time, it could gradually modify our understanding of legal principles or societal values, without our realising that this has happened. That risk is very modest today, but it can grow as reliance grows. And the unseen incremental nature of such a change over time is precisely what makes it dangerous.

36. The second risk is skills atrophy in the generation of lawyers and judges coming through, who, for example, may never have to construct an argument from first principles because AI has always produced the first draft. Legal and judicial craft is learnt by doing: badly at first, and then better, through repetition, reflection and intention.

37. Both risks point in the same direction. They call for some degree of friction — deliberately designed into the systems we introduce. This can be a very nuanced exercise. For example, a judge might be asked to set out the issues and a provisional view before seeing an AI-generated summary. They also call for us to reshape the way we train lawyers and judges, with a greater emphasis on both core judicial competencies and AI judicial competencies, so that our judges learn to adapt to and use AI, without losing their core judicial skills.

Judicial Education

38. Let me now turn to what all this means for judicial education. I will take it in four parts: what judges need, namely twin competencies; how deep and how broad those competencies must go; how we teach them; and the wider role of our judicial education institutions.

39. Let me begin with what judges need. We should focus on developing competencies, rather than transmitting information. AI will increasingly supply the information. Judges must still supply the judgment. What judges will continue to need are the core judicial competencies to understand, evaluate and apply what AI tools provide, so as to use those tools appropriately and effectively. Indeed, the more capable AI becomes, the more important it is that judges possess the foundational competencies to scrutinise AI’s output independently. This calls for a return to and an emphasis on the basics: for example, the identification and assessment of key and relevant evidence, critical legal analysis including the ability to understand the first principles, and to reason through grey or hard areas of law. This is particularly important, given the concerns I highlighted earlier with skills atrophy, which can set in as lawyers and judges become more reliant on AI tools over time.

40. The concern about skills atrophy is not speculative. In medicine, a multicentre study compared how endoscopists performed in colonoscopies done without AI assistance, before and after AI was introduced into their practice. Their rate of detecting adenomas in those non-AI procedures fell from 28.4 per cent to 22.4 per cent.9 In a survey of 319 knowledge workers, researchers from Microsoft and Carnegie Mellon University found that higher confidence in generative AI was associated with less critical thinking, while higher self-confidence was associated with more.10

41. Two lessons follow. Skills that are not exercised can erode, even among trained professionals. And it is the professional who is confident in his or her own competence who is more likely to engage critically with what AI produces. Building core judicial competencies is therefore not a nostalgic exercise. It is what equips judges to scrutinise AI.

42. There is also a pipeline dimension to this. Many of us learnt our craft through a form of apprenticeship: as young lawyers drafting submissions, or as law clerks and junior judicial officers preparing the first cut of research, which more senior judges then corrected. If that first draft is increasingly produced by AI, the apprenticeship that formed today’s judges may quietly disappear for tomorrow’s. Judicial education institutions may need to recreate some of that formative experience deliberately, for example through structured drafting exercises without AI, and mentorship that focuses on how a judgment is built and not only on how it reads.

43. This emphasis on building core judicial competencies needs to be twinned and complemented by building up AI judicial competencies. Put simply, we need to help our judges to be ambidextrous – capable of exercising independent judgment while also being proficient in using AI to support their work.

44. Two reasons make this pressing.

45. The first is realism. Our judges are already using AI tools, or soon will be, and they must be trained for that reality. Without it, the danger of shadow AI use is very real. This is not a hypothetical risk. In a UNESCO survey of 563 judicial operators across 96 countries, conducted in late 2023, 44 per cent said they had already used AI tools for work-related activities. Yet only 9 per cent said their organisation had issued guidelines on the use of AI chatbots, and only 9 per cent said it had provided AI-related training or information.11 Of the respondents who added comments, 27 per cent called for courses and training on AI tailored for judicial operators. Our judges will use these tools. The question is whether they do so with training, or without it.

46. The second is discernment. Judges need hands-on exposure to AI tools, to the extent their judiciaries allow, not only to use them well but to understand their limits and flaws. 

47. Without complementing the development of core judicial competencies with AI judicial competencies, we run risks that cut across a spectrum: at one end, judges who continue to work less efficiently without AI; at the other, judges who rely completely on AI without an understanding of how that might affect the proper execution of their judicial responsibilities.

48. Research on AI in legal work underlines the need for both sets of competencies. In what its authors described as the first randomised controlled trial of AI assistance in legal analysis, law students using GPT-4 completed realistic legal tasks considerably faster, but the quality of their work improved only slightly and inconsistently. A later trial using newer reasoning models and specialised legal tools found significant improvements in quality as well.12 The tools are improving quickly. But the better they become, the more they offer, and the more it matters that the person using them can tell sound output from output that is plausible but flawed. That is what it means to be ambidextrous.

49. How deep, and how broad, must these competencies go? Let me start with depth. The importance of progressively developing judicial competencies is already recognised in judicial education. For instance, under the Singapore Judicial College’s Judicial Competency Framework, “101”, “201” and “301” programmes are targeted respectively at the foundational, intermediate and advanced phases of a judge’s career, progressively deepening judicial competencies.13 UNESCO’s AI Competency Framework for Teachers offers a complementary perspective. It sets out competencies across five dimensions, from a human-centred mindset and the ethics of AI, to AI foundations and applications, at three progression levels: acquire, deepen and create.14 A similar progression can apply to AI competencies for judges. Every judge needs to acquire the foundational competencies. Many judges, particularly those in courts that deploy AI tools, will need to deepen them through use in their daily work. A smaller group will need to create: to help design, test and govern the tools. That third level is where the legal engineering competencies sit – a point that I will come to shortly.

50. Turning to breadth, there are, in my view, at least three areas of AI judicial competency development.

51. The foundational area is technical awareness, helping our judges to understand the functionality – and, therefore, possibilities and limitations – of AI tools, and training them to use AI tools in the areas of judicial work for which that judiciary has permitted AI to be used. In the Singapore Judicial College, we provide hands-on sessions that allow our judges to play with and acquaint themselves with new technology, including AI tools. We also train our judges on the ethical and responsible use of AI, so that its benefits can be leveraged without compromising judicial values.

52. What does technical awareness mean in practice? I would suggest that every judge should be able to do at least five things. First, understand in broad terms how the AI tools they use work, and therefore where those tools are most likely to fail. Second, verify AI output against the underlying sources before relying on it. The judicial guidance in England and Wales puts it plainly: judicial office holders are personally responsible for material produced in their name, and judges must always read the underlying documents.15 Third, calibrate their trust, by being alert to automation bias, the tendency to defer to an automated output even when other signals suggest it is wrong. Fourth, protect confidentiality, by knowing what information may be entered into which tools. Fifth, recognise the use of AI by others, from fictitious authorities cited in submissions to AI-generated or manipulated evidence.

53. I would differentiate AI judicial competencies from legal engineering competencies, which involve acquiring a much deeper understanding of the technology of AI and how AI can be engineered to assist with judicial work. Legal engineering competency is a useful and needed area for some members of a judiciary to acquire in order to drive effective policy making and implementation regarding AI use in a judiciary, but it is not one that needs to be developed across the board for all judges.

54. For judiciaries in the European Union, these competencies also have a legal dimension. The AI Act provides that deployers of high-risk AI systems must assign human oversight to persons who have “the necessary competence, training and authority”, and requires providers and deployers of AI systems to take measures to support the AI literacy of their staff.16 Whatever the legal framework, the underlying point holds in every jurisdiction. Human oversight of AI is only as meaningful as the competence of the human exercising it.

55. A second area is AI and the law. AI is forcing us to re-examine many areas of law, such as contract, tort, criminal law, property law and intellectual property. In the Singapore Judicial College, besides drawing on a range of leading experts in this field, we also draw on internal horizon-scanning teams to identify emerging issues and the legal approaches taken in different jurisdictions.

56. The third area is AI in a socio-legal context. This canvases broad concerns about the ethics and regulation of AI as well as how AI can be used to enhance our judicial system, improve access to justice, and also how the use of AI by lawyers and litigants in person can affect our judicial system. It is driven by the dual nature of our judicial responsibilities, in our adjudicative role and in our systemic justice role, as custodians of the justice system. In the Singapore Judicial College, we seek to bring about such wider system level perspectives in our judges, by drawing on the perspectives not only of experienced judges from other jurisdictions, but also of those who have been involved in digital transformations in non-legal areas, in both the public and private sectors.

57. How, then, do we teach these competencies? In practical terms, I see two kinds of pedagogy.

58. First, judicial education that is focused squarely on building core judicial competencies. These programmes build foundational adjudicative skills, such as legal analysis, managing cases through the dispute resolution pipeline or the principles of judgment writing, without the involvement of AI tools.

59. Aviation offers a useful parallel. As cockpit automation became pervasive, Europe’s aviation safety regulator observed that the continuous use of automated systems does not help pilots maintain their manual flying skills, and encouraged operators to consider building manual flying into their training programmes.17 Core judicial education without AI serves the same purpose. It keeps judges able to take the controls when it matters most: in the hard case, or when the tool fails.

60. Second, integrated judicial education that develops core judicial and AI judicial competencies in combination. For example, judges may be trained in, or required to use, AI tools in the course of studying a particular aspect of judgecraft or bench skills. They may be asked to critique an AI generated judgment containing deliberate errors such as a mischaracterisation of an authority, a failure to consider material evidence, flawed legal reasoning or simply poor writing. Or they might be given an AI generated dispute involving a range of court users and be required to participate in a role play to develop courtroom communication skills.

61. How AI is introduced into training also matters. In a field experiment with nearly 1,000 high school mathematics students, those given a standard GPT-4 interface improved their grades by 48 per cent while they could use it to solve problems. But once access was removed, their grades were 17 per cent lower than those of students who never had access. A version of the tool designed with safeguards to support learning largely mitigated this harm.18 The students had used the unguarded tool as a crutch. For judicial education, the lesson is to design friction in deliberately. For example, judges can be asked to identify the issues and form a provisional view first, and only then compare their analysis with what the AI produces. There is learning in that comparison. Exercises of this kind also allow us to assess competencies, and not merely teach them. Whether a judge spots the errors planted in an AI-generated draft is itself a measure of the competencies we are seeking to build.

62. Judicial education cannot rest on programmes alone. Given how fast AI is developing, it would be very difficult for any individual judge to keep abreast of developments across all relevant areas. It is also important to build up a cohort of judges who can horizon scan and prepare for an AI future. So, in Singapore, we complement the Singapore Judicial College’s programmes with the formation of a Technology Law Cluster. This draws in interested judges from across our various courts to form a community of learning, where judges share their respective areas of expertise and build each other up. It is also where judges with an interest in legal engineering can build that specialised expertise. This helps build, within the judiciary, the capacity needed to make sound policy decisions on AI.

63. None of this is possible unless those who teach are themselves equipped. Our faculty, many of them serving judges, need AI judicial competencies before they can build them in others. They need to know the tools well enough to design exercises around them, to recognise when AI has been used in a learner’s work, and to keep pace as the tools change. Investing in faculty development is therefore part of the task, not an afterthought. It is also one reason the Singapore Judicial College has brought in specialists in education and pedagogy to work hand in hand with our judges.19 

64. Finally, all of this points to a wider role for judicial education institutions. We should not see ourselves as simply receiving AI policy and then teaching it. The content of AI judicial competencies depends on where each judiciary draws its line on AI use, and as that line moves, the curriculum must move with it. At the same time, education institutions see, through the judges they train, how AI is actually being used and where competencies are falling short. That makes our education institutions a valuable source of insight for those making policy on AI. And because the challenges we face are shared, there is much to be gained from networks such as the IOJT, through which we can exchange curricula, exercises and lessons learned across jurisdictions.

Conclusion

65. As I conclude, let me return to where I began. The question before us was never whether AI would reach our judges. It already has. The question is whether our judges will be ready.

66. Ready, in my view, means ambidextrous: strong in the core judicial craft, and skilled in the use of AI, with each hand strengthening and supporting the other. A judge with only the first will be left behind. A judge with only the second will find AI tools very obliging.

67. Our judiciaries will draw the line on AI in different places. But wherever that line falls, the mission of judicial education institutions is the same. It is to ensure that behind every judgment there is a judge who made the decision, and who can explain it. For judges cannot be accountable for decisions they cannot fully explain, and they cannot fully explain decisions they did not fully make.

68. This conference gives us a rich opportunity to learn from one another in that common mission. I wish you all a very good conference. Merci beaucoup, thank you.

1 Amos Tversky and Daniel Kahneman, “Judgment under Uncertainty: Heuristics and Biases” (1974) 185(4157) Science 1124; see also Daniel Kahneman and Amos Tversky, “Subjective Probability: A Judgment of Representativeness” (1972) 3(3) Cognitive Psychology 430.
2 Matthew Dahl et al, “Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models” (2024) 16(1) Journal of Legal Analysis 64 at 66: “when asked a direct, verifiable question about a randomly selected federal court case, LLMs hallucinate between 58% (ChatGPT 4) and 88% (Llama 2) of the time”. A follow-up study of leading commercial legal research tools found that “hallucinations are reduced relative to general-purpose chatbots (GPT-4)”, but that “the AI research tools made by LexisNexis (Lexis+ AI) and Thomson Reuters (Westlaw AI-Assisted Research and Ask Practical Law AI) each hallucinate between 17% and 33% of the time”: Varun Magesh et al, “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools” (2025) 22(2) Journal of Empirical Legal Studies 216 at 216. The tools were tested between March and May 2024, and a response was counted as hallucinated if it was “either incorrect or misgrounded” (at 221 and 224).
3 Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), art 6(2) (“AI systems referred to in Annex III shall be considered to be high-risk”) and Annex III, point 8(a), which lists AI systems intended to be used by or on behalf of a judicial authority to assist it in researching and interpreting facts and the law and in applying the law to a concrete set of facts, or to be used in a similar way in alternative dispute resolution. See also recital (61): “the final decision-making must remain a human-driven activity”.
4 Supreme Court of Singapore, Registrar’s Circular No 1 of 2024, Guide on the Use of Generative Artificial Intelligence Tools by Court Users (issued 23 September 2024; effective 1 October 2024).
5 Dahl et al, “Large Legal Fictions”, above, at 66: “LLMs perform better on cases that are newer, more salient, and from more prominent legal jurisdictions, suggesting that the risks of legal monoculture are real”; and at 76: “hallucinations are lowest in the highest levels of the judiciary”. The pattern is not uniform: among US Supreme Court cases, “hallucinations are most common among the Supreme Court’s oldest and newest cases” (at 79).
6 David Rozado, “The political preferences of LLMs” (2024) 19 (7) PLOS ONE e0306621. Rozado administered 11 political orientation tests to 24 “state-of-the-art conversational LLMs”, and found that “most conversational LLMs tend to generate responses that are diagnosed by most political test instruments as manifesting preferences for left-of-center viewpoints”.
7 Rozado, above, reporting that LLMs “can be steered towards specific locations in the political spectrum through Supervised Fine-Tuning (SFT) with only modest amounts of politically aligned data”.
8 Yachao Zhao et al, “Explicit vs. Implicit: Investigating Social Bias in Large Language Models through Self-Reflection”, in Findings of the Association for Computational Linguistics: ACL 2025 (2025) 1. The authors report that “LLMs exhibit a substantial inconsistency between explicit and implicit biases”, and that “[w]hile implicit bias shows strong stereotypical tendencies, explicit bias exhibits relatively mild stereotyping” (s 5.1). Explicit bias was measured “by prompting LLMs to analyze their own generated content”.
9 Krzysztof Budzyń et al, “Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study” (2025) 10(10) The Lancet Gastroenterology & Hepatology 896. Across 1,443 patients at four endoscopy centres in Poland, comparing the three months before and after AI was introduced, the “ADR [adenoma detection rate] of standard colonoscopy significantly decreased from 28.4% (226/795) before to 22.4% (145/648) after the exposure to AI”, and the authors conclude that continuous exposure to AI “may reduce” the detection rate of standard colonoscopy. Being observational, the study shows an association rather than proving causation. Quotations are from the authors’ accepted manuscript (Warwick Research Archive Portal, wrap.warwick.ac.uk/191005), whose wording may differ slightly from the published version.
10 Hao-Ping (Hank) Lee et al, “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers”, in Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (ACM, 2025): “higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking”. The first author is at Carnegie Mellon University and the other authors at Microsoft Research. The findings rest on self-reports, and the authors state that “[o]ur analysis does not establish causation” (s 6.1.1).
11 Juan David Gutiérrez, UNESCO Global Judges’ Initiative: Survey on the Use of AI Systems by Judicial Operators (UNESCO, 2024). The survey ran from September to December 2023; “563 responses were recorded from judges, prosecutors, lawyers, civil servants in legal administration, and researchers working in 96 countries”. 44% “stated that they had used AI tools for work-related activities”; “[o]nly 9% ... stated that their organization had issued guidelines or regulations for using AI chatbots”, and the same percentage said their organisation “had provided AI-related training or information”; and “27% of the judicial operators who voluntarily shared additional comments ... mentioned the need for courses and training on AI specifically tailored for judicial operators”.
12 Jonathan H Choi, Amy B Monahan and Daniel Schwarcz, “Lawyering in the Age of Artificial Intelligence” (2024) 109(1) Minnesota Law Review 147: “We conducted the first randomized controlled trial to study the effect of AI assistance on human legal analysis”, finding that “access to GPT-4 only slightly and inconsistently improved the quality of participants’ legal analysis but induced large and consistent increases in speed”. A later randomised trial with upper-level law students, using OpenAI’s o1-preview reasoning model and the retrieval-augmented legal tool Vincent AI, found that “both AI tools significantly enhanced legal work quality” and that AI assistance “significantly boosts productivity in five out of six tested legal tasks”: Daniel Schwarcz et al, “AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice”, University of Michigan Public Law & Legal Theory Working Paper No 772 (2025)
13 Kwek Mean Luck, Natalie Skead, Justin Yeo & Yeo Mui Lin, “Judicial Education in Singapore: The Past, Present and Future” (2025) 37 SAcLJ 409 at 425 (para 45) and 431 (para 58).
14 Fengchun Miao and Mutlu Cukurova, AI Competency Framework for Teachers (UNESCO, 2024). The five dimensions are a human-centred mindset, the ethics of AI, AI foundations and applications, AI pedagogy, and AI for professional learning.
15 Courts and Tribunals Judiciary (England and Wales), Artificial Intelligence (AI) Guidance for Judicial Office Holders (8 October 2025) at 6: “Judicial office holders are personally responsible for material which is produced in their name”; “Judges must always read the underlying documents”; and “AI tools may assist, but they cannot replace direct judicial engagement with evidence.”
16 Artificial Intelligence Act, above, art 26(2) (“Deployers shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support”) and art 4(1), as substituted by Regulation (EU) 2026/1744 (Digital Omnibus on AI) of 8 July 2026 (“Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff ...”). Under Regulation (EU) 2026/1744, recital (40), the obligations for high-risk AI systems listed in Annex III, including art 26, apply from 2 December 2027.
17 European Union Aviation Safety Agency, Safety Information Bulletin No 2013-05R1, Manual Flight Training and Operations (23 June 2025), which revises SIB 2013-05 (23 April 2013): “Nevertheless, continuous use of automated systems does not contribute to maintaining pilot manual flying skills”; and “Operators are encouraged to consider incorporating emphasis of manual flight operations, as a means of maintaining basic flying skills, into their training programme and, when feasible, line operations”.
18 Hamsa Bastani et al, “Generative AI without guardrails can harm learning: Evidence from high school mathematics” (2025) 122(26) Proceedings of the National Academy of Sciences e2422633122: “having GPT-4 access while solving problems significantly improves performance (48% improvement in grades for GPT Base and 127% for GPT Tutor)”; but “when access is subsequently taken away, students actually perform worse than those who never had access (17% reduction in grades for GPT Base)”; “[t]hese negative learning effects are largely mitigated by the safeguards in GPT Tutor”.
19 Kwek Mean Luck et al, “Judicial Education in Singapore”, above, at 425 (para 43), referring to the “need for specialists with expertise in legal education, pedagogy and curriculum, research and leadership to work hand-in-hand with the Judiciary”.
2026/11/19

Share this page:
Facebook
X
Email
Print