21 July 2026
Introduction
E ngā mana, e ngā reo, tēnā koutou katoa.
I have set myself two questions — should judges interact with AI and if so on what terms — which I could answer briefly as ‘yes’ and ‘on our terms’. Although I am confident with an affirmative answer to the first question, the second issue requires a bit of elaboration. To be frank I am not sure that we are comfortably in a position where we can impose the terms. And trying to do so is more complex and requires a lot more thinking about what those terms might be.
Let me deal firstly with the easier question. In short we will interact with AI because, first, AI and its operations will inevitably be litigated. Second, AI will be pervasive in the community and because it will play an enormous role in the documentation and recording of our lives, it will significantly affect the evidentiary record. Third, it will affect the way parties prepare and present cases. We will not be able to ignore it. Can I briefly elaborate on these points.
Judges will inevitably interact with AI because AI is part of the world we live in and increasingly so. The metrics as to the growth of AI are truly incredible.
First, the scale of investment in AI models, computer chips, and data centres is breathtaking.
Consider the scale. Four of the hyperscalers — Meta, Alphabet, Amazon, and Microsoft — have between them set aside some USD725 billion for data centres and physical equipment in 2026 alone.1 That figure is close to half the size of the entire Australian economy.2 Indeed, investment in data centres and information-processing equipment accounted for almost the entirety of United States GDP growth in the first half of 2025.3
Second, AI and in particular data centres require large-scale infrastructure and energy use. In the USA data centres use approximately 4.5 per cent of the entire American energy output. Closer to home, the Australian Energy Market Operator expects their demand to triple — to around 6 per cent, or 12 TWh, of electricity in the main grid — between 2025 and 2030. That is roughly enough to power every home in Victoria.4
Third, investment, impact and capability is driven by adherence to the fundamental scaling law which provides that more compute, more training data and more parameters makes a better AI model.5 In order to feed the Large Language Models (‘LLMs’), their originators seek access to more and more data often apparently regardless of whether it is subject to IP protection. This has seen the capability of AI expand exponentially with its reach extending to every area of society.
The combination of investment, resource intensity, ubiquity and insatiable thirst for data means that AI will inevitably be the subject of litigation. Already we are seeing major litigation in the United States. Apple is suing Open AI over IP breaches.6 And Anthropic, the maker of Claude, is suing the US Government because of a decision to exclude it from all government contracts.7 That decision was made because Anthropic sought an assurance that its programs would not be used for mass surveillance and autonomous weapons. The dispute arose after reports that Claude had been used in the US operation to capture the former Venezuelan President, Nicolás Maduro.8 The US Department of War insisted that it should be able to use the products for all lawful uses. At around the same time, the President issued a directive to the Department of War to rewire the policy on the use of autonomous weapons. He said the US will ‘responsibly accelerate the use of AI across intelligence and warfighting domains in line with American values’.9 The dispute with Anthropic prompted a memorable post from the President who declared: ‘The Leftwing nut jobs at Anthropic have made a DISASTROUS MISTAKE trying to STRONG-ARM the Department of War, and force them to obey their Terms of Service instead of our Constitution’.10
Next, AI will affect the evidentiary record. This will not only arise in a context where AI is used to create images and data, but also because its capacity to synthesise, analyse and interrogate data enables access to information hitherto unavailable. Criminal trials have already seen the widespread adoption of DNA evidence, telephone location data and CCTV images. AI has the potential to turbocharge that process.
AI has vastly amplified the surveillance powers of the state — not least the ability to sift, in moments, through millions of hours of footage. We have all seen the police procedurals where the officer tracks frame by frame and amazingly finds some critical clue or point of connection.
On 1 March, Israeli intelligence revealed that Israel had undertaken large-scale strikes on ‘the headquarters of the Iranian terror regime’ in Tehran.11 It killed the supreme leader.
Israeli intelligence had harvested vast amounts of video footage from Iran’s traffic cameras to help pinpoint the exact location and timing of a February 28 meeting between Khamenei and his closest aides.12 It is well known that CCTV camera feeds can be accessed and now analysed using language-based searches of video images. The operator can identify people by reference to patterns or behaviours and build up detailed predictive models as to where particular targets may be. The data can be cross-referenced with social media, hacked communications, audio picked up from smart phones or cars and travel histories. New generations of AI-enabled cameras and software can interpret scenes, identify patterns of behaviour and retrieve footage in response to written prompts.13
Given that it is in my view inevitable that judges will be driven to interact with AI, I can attend to the more difficult question of the terms on which judges may adopt AI in the performance of their work.
So, given contact is inevitable, how should we deal with the judicial use of AI? The appropriateness or desirability of judicial use of AI might be approached by a number of perspectives or starting points.
- The first might be to ask: what is AI good at?
- The second might identify that which the courts are not doing well and which AI could be employed to improve.
- Finally, we might start with the question of what courts are for: what is their reason for being?
As with many things, the starting point is important because of its potential to narrow the focus of inquiry and drive the outcome so it is convenient to briefly interrogate each of the three perspectives.
What is AI good for?
AI is good at many things. Indeed its capacity for some functions vastly outstrips that of humans.
Take coding. Coding involves creating instructions in a programming language that enables a computer to execute a relevant function. In recent times vast wealth has been earned by companies who specialise in the development and improvement of code: SAP, Oracle, Salesforce and Microsoft. Often this work has been highly skilled, expensive and sophisticated. Increasingly, AI is being used to prepare code at a very high level. In the face of AI growth, one trillion dollars were knocked off the value of software stocks14 and in the USA they are down about 20 per cent year on year.15 At the same time the Philadelphia SE Semiconductor Index is up 120 per cent year on year. The contrast is partially explained by the ability of AI to replace software development.
I have already mentioned surveillance but to that can be added other functions both benign and malevolent. There are fears that AI can develop biological weapons that turn on specialised knowledge by bringing together vast biological and chemical data sets. AI augmented weapons are an increasing feature of modern military endeavour.
More prosaically AI is very good at analysing and synthesising data and particularly documentary material. In many ways the judicial process is very much in the AI wheelhouse. AI can assimilate vast amounts of information very quickly and accurately. The LLMs are based on probability predictions having regard to the vast resource that underpins the model. The judicial task is often associated with probability. The civil standard is based on the balance of probability and probability reasoning is commonly employed in similar fact evidence or as we call it in Victoria tendency and coincidence. Similarly, causation involves the counterfactual: the hypothetical assessment of the probability of various alternative events had the wrongdoing not occurred. All of these functions are grist to the mill for an LLM.
Of course, the claimed advantages of AI in terms of speed and accuracy have to be seen in the context of a febrile market where valuations are based on exponentially increasing capacity. There remains questions about the accuracy and suitability of AI for some functions. And the differing capabilities between specialist and generalist models is still being worked through.
Undoubtedly, Al models can prepare chronologies, summaries of information, identify conflicts in the material and even express opinions on the correct disposition having regard to legal databases to which it is directed.
When one looks at the complexity of function and the significance of the decision including when and where bombs should fall, it is impossible to argue that AI is concerned with routine, mundane or repetitive decisions and on that basis alone is ill-suited to our complex work.
What are we not doing well?
The next perspective is what courts are not doing well.
When looking at the work of the courts there is a recurring refrain that litigation is both too slow and too expensive. Those complaints are difficult to rebut.
It is commonplace to say that the civil courts are no longer accessible to the overwhelming majority of the community.
When one combines the strengths of AI and the weaknesses of the court it would seem to be difficult to withhold AI from the judge’s arsenal. Is that the end of the matter?
Why are we here?
I turn then to the third perspective, which is the rationale that underpins the courts. Our job is the resolution of disputes according to law. It involves the final determination of legal disputes so that the outcome can be enforced. The cardinal attributes of the process include independence, impartiality, fairness and transparency. These attributes provide the environment in which legal rights and liabilities are determined in a way that is final and accepted.
Central to that role is converting the uncertain to the certain: allegations to findings, legal claims to orders. Certainty is critical to the process and it reflects the application of the known law to individual circumstances in accordance with the judicial method which is designed to produce a single legally correct outcome. Certainty requires that the outcome be accepted as final.
Although the outcome is final and certain, the work of the courts is often undertaken in the grey. An often remarked upon aspect of the rule of law is that the rule must be capable of being understood and applied by those who are bound by it.16 Notwithstanding clarity in the law is an important virtue, the complexities of dealing with human interaction and human foibles means that many legal standards and principles are couched in terms that require evaluation and judgment. So for example we are constantly asked to assess conduct against varying standards: has there been a lack of reasonable care, is the impugned conduct unconscionable, what is the moral culpability of an offender or is the conduct arbitrary or misleading? And that determination also requires resolution of factual contests.
In discharging the judicial oath in this grey zone there are three aspects that are profoundly important. The first is that the judicial function is performed by judicial method. This involves an impartial and independent tribunal determining the questions of fact and law that arise in a way that is procedurally fair to the parties in an open court. The second is it involves the exercise of judgment. This entails both an understanding of the law with all its gradations in standard and expectation, and an ability to apply to the facts of the particular case. It involves a human judgment about contestable issues. The third aspect involves recording and explaining the decision and the reasons for it. This last feature is essential to demonstrate that the outcome is the product of exposed legal reasoning on an expressed factual basis that has been determined in accordance with the judicial method. The reasons assist in acceptance of the outcome by demonstrating that it is not the arbitrary or idiosyncratic decision of the judge.
What terms should we accept?
It seems to me that in approaching the judicial use of AI it is the third perspective that is the most important. We should not be seduced by the efficiency of AI, in the face of the problems that presently subsist, without proper regard to the human dimension of the judicial function and the risks that AI poses to that human aspect.
How then do we obtain the benefits that AI can yield but maintain the human centred judge?
Can we trust judges to avoid the blandishments of AI any more than we expect a factory worker to be alert enough not to put their hand in an unguarded machine? We can, provided we do not lose sight of the judicial function — which is as concerned with method as outcome — and of the ethical framework we work in, and ensure that there is responsible procurement and use of AI.
In this respect, the first point I would emphasise is the importance of governance that provides the framework in which courts and individual judges approach the acquisition and use of AI. That framework should take as an organising principle the essential human attributes of the judicial process rather than the skills or benefits of AI.
One of the essential aspects of any framework must be the overt retention of human decision-making. In the judicial context we see that the Australian Guide to Judicial Conduct states that ‘judges must not use Generative AI for judicial decision-making’ with the further explanation that the ‘prohibition on AI use for judicial decision-making does not encompass supportive uses of AI’.17
This problem is not unique to the judicial process.
In his papal encyclical Magnifica Humanitas, Pope Leo XIV makes an important point that, despite the anthropomorphic approach to AI, AI mimics but does not engage in human thought. He says:
so-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships and do not know from within what love, work, friendship or responsibility mean … They may imitate language, behaviour and analytical skills, or even simulate empathy and understanding, but they do not understand what they produce, for they lack the affective, relational and spiritual perspective through which human beings grow in wisdom.18
The Pope’s deepest objection is a moral one, and it speaks directly to autonomous decision-making. A machine, he reminds us, ‘may imitate language, behavior and analytical skills, or even simulate empathy and understanding, but they do not understand what they produce, for they lack the affective, relational and spiritual perspective through which human beings grow in wisdom’.19 That absence is most grave in the context of war. There is, the Pope writes, ‘no algorithm [that] can make war morally acceptable’.20 Artificial intelligence ‘does not remove the intrinsic inhumanity of conflict; indeed, it can only bring about conflict more quickly and render it more impersonal, lowering the threshold for resorting to violence, transforming defence into threat prediction and thus reducing victims to data’.21
From this he draws a principle that any lawyer will recognise as a principle of responsibility: ‘any technology that facilitates attacks without seeing the face of human beings lowers the moral threshold of conflict’.22 The point is not merely that autonomous systems may err. It is that the act of removing a human being from the decision — the person who must look upon the consequence and answer for it — is itself the danger. Responsibility, on this view, cannot be delegated to a machine, because a machine cannot bear it. It is a moral counterpart to the legal instinct that animates so much of what we do: that for every consequential decision, there must be someone who is answerable.
It is important to recognise that the limitations that AI might be used to overcome: the time and cost of individual human decision-making in an open and transparent way with procedural fairness are inherent features of the system. The application of human thought takes time. The Pope makes the point that everything that appears as a ‘limit’ — incapacity, illness, old age, suffering, vulnerability — ‘tends to be seen primarily as a defect to be corrected.’23 He gives the counterpoint that ‘humanity flourishes not despite limitations, but often through them’.24
Human in the loop
Maintaining a human centred judiciary is not satisfied merely by having a human in the loop.
One recurring theme in discussions surrounding the adoption of AI is the importance of ‘a human in the loop’. This is often framed as a necessary corrective to the errors or mistakes of AI. We have tended to see policies that embrace AI but require that the output is verified. Accuracy in documents that are submitted to a court as part of the judicial process is undeniably a sound ideal. But the issue again is not just accuracy, and the challenge that AI presents for judicial use is not merely addressed by requiring verification with or without overt certification.
Verification is important but I am not persuaded that from a structural viewpoint simply asserting that there will be human verification will be a sustainable position for courts to adopt in their use of AI. That is so for a number of reasons.
It is not sustainable.
Requiring a human to take responsibility for AI’s output is plainly valuable. But I fear it may prove harder to sustain than we would like. In a recent working paper on the economics of artificial general intelligence, three economists argue that the ‘human in the loop’ is, on closer inspection, ‘dynamically unstable’.25 Their argument runs as follows. As AI grows ever cheaper and faster, the binding constraint on the economy is no longer machine intelligence; it is human verification — the scarce capacity to check, audit and stand behind what the machine has produced. Execution can be scaled almost without limit; verification cannot, because it remains bounded by human time and hard-won experience. The ‘frontier of what agents can cost-effectively execute expands faster than our biological capacity to oversee it’.26
The consequence is a widening gap between what these systems can do and what any human can realistically check — and a powerful commercial incentive to deploy them faster than we can oversee them. In economic terms there is higher nominal output but reduced overall utility: ‘unverified deployment becomes privately rational’ and this tends to produce an acceptance of the plausible rather than the correct.27
This process is self-propagating because as AI is increasingly adopted it reduces the skills of the junior as the entry level work is replaced by agentic AI; the expert’s verification forms part of the training data that accelerates their own automation and this leads to a widening gap between execution and verification.
As a matter of principle the correct approach is not one of verification but demarcation. And the line of demarcation is that judgment is a human endeavour that can be supported by but not replicated by AI.
Although there is a general consensus that AI should not be used to make decisions or write judgments may I explore further the problem of having AI decide cases.
A variety of courts, including my own,28 have sought to illustrate the legitimate uses of AI by judges and the lists often include things like: preparation of a chronology, summaries of pleadings, evidence or submissions, extracting parts of transcripts, proof-reading or searching for known material in large court files. One rule of thumb that is occasionally mentioned is that if you would give the function to your clerk or associate then you can safely give it to AI.
We do not have the practice of having our associates or clerks write judgments. As I understand it, it is not an uncommon practice in the United States, up to and including the US Supreme Court. In his thoughtful Reflections on Judging, Richard Posner observes that the ‘clerk-written opinion lacks color, depth, and authenticity’.29 He points out that judgment involves knowledge but also experience and that is why brilliant 25 year olds are not judges.30
In explaining why he does not have his clerks write his opinions, Posner draws the distinction between the writer model and the manager model. According to Posner, the manager, who delegates the first draft and then edits the product:
may not realize that the process of writing, which means searching for words, for sentences, in which to express meaning, is a process of discovery rather than just of expressing preformed ideas; that it reveals analytical gaps; that it gives rise to new ideas; and that fluency in writing comes largely from — writing.31
His own approach is to start with a draft and rely on his clerk to perform the role of editor, filling gaps, and undertaking further work to assist in the next draft.32
Posner’s criticisms of clerk-written opinions apply also to the use of AI to write judgments: the loss of human connection, responsibility and accountability for the judgment.
There is an additional risk that having AI write judgments will flatten the range and expression of judgments which would be a serious failing. As a member of an appellate court I have often been inspired and even provoked by the way colleagues have dealt with various issues. Individual expression can expose different perspectives and different ways to come to a particular point. An important part of writing judgments, especially appellate judgments, is knowing what to leave out, and what to leave for another day. These fine-grain aspects of the method are not easily replicated by a probabilistic model drawn from a larger data base. Maybe we would see the end of the useful and powerful dissent.
How then do we reflect these principles in a workable framework: what are legitimate supportive uses that do not undermine the fundamental human role of the judge? I am reluctant to be too prescriptive.
Being too prescriptive is difficult because there is no single way in which judges prepare for, hear and determine cases before them. The various different ways are a function of circumstance: how busy the judge is and the types of cases they hear. But they are also a function of personality. Some judges need to feel they have covered every point; others focus on the critical points that resolve the case or which interest them. And there are many different ways in which judges approach writing their reasons.
For many judges there are likely to be many supportive uses that do not undermine either the decision-making or authorship of the judge. But for others, many of the associated tasks are tightly bound up with how they approach a case and think about and resolve the issues. For this reason guidance should be principle-based: the decision must be that of the judge, and although there ought be clear lines of demarcation — which I think include the reasoning process, fact-finding, the exercise of judicial discretion and producing reasons for judgment — the guidance need not be too prescriptive. The concept of supportive uses is practical enough, especially when elucidated with examples such as summarising the material, preparing chronologies and explaining legal principles.
I think that a relatively permissive environment that allows the supportive use of AI but is accompanied by training as to the imperative of responsible use and the limitations of AI has the potential to achieve gains in both time and cost. But there are undoubtedly risks. There are also risks of inaction through, in effect, the unthoughtful and unregulated reality of judicial use of AI tools.
Fundamentally, our approach to AI should be driven by a clear-eyed and protective view of our role and the importance of human decision-making. We all have a responsibility for this.
Nō reira, tēnā kotou, tēnā kotou, tēnā tatou kātoa.
[1] Thorsten Meyer, ‘The $725 Billion Question: Hyperscaler Capex Q1 2026 and What the Earnings Don’t Answer’ (Web Page, 6 May 2026).
[2] ‘GDP (current US$) - Australia’, World Bank Group (Web Page).
[3] Nick Lichtenberg, ‘Without Data Centers, GDP Growth was 0.1% in the First Half of 2025, Harvard Economist Says’, Fortune (online, 7 October 2025).
[4] Climate Council of Australia, Clouded Future: Managing the Risks of the Data Centre Boom (Report, 3 June 2026) 12.
[5] Kari Briski, ‘How Scaling Laws Drive Smarter, More Powerful AI’, NVIDIA (Blog Post, 12 February 2025).
[6] Jon Markman, ‘Apple Sues OpenAI For Trade Secret Theft Ater Integrating ChatGPT’, Forbes (online, 12 July 2026).
[7] See, eg, Anthropic PBC v US Department of War, USCA No 26-1049 (2026).
[8] Amrith Ramkumar, Keach Hagey and Vera Bergengruen, ‘Pentagon Used Anthropic’s Claude in Maduro Venezuela Raid’, Wall Street Journal (online, 15 February 2026).
[9] The White House, ‘National Security Presidential Memorandum’ (Presidential Memoranda NSPM-11, 5 June 2026) s 1.
[10] @realDonaldTrump (TruthSocial, 28 February 2026) <https://truthsocial.com/@realDonaldTrump/posts/116144552969293195>.
[11] Mehul Srivastava and Christopher Miller, ‘New AI Espionage Powers Trigger Putin Camera Scare’, Financial Times (online, 8 June 2026).
[12] Ibid.
[13] Mehul Srivastava and Christopher Miller, ‘New AI Espionage Powers Trigger Putin Camera Scare’, Financial Times (online, 8 June 2026).
[14] Jack Pitcher and Xavier Martinez, ‘The $1.6 Trillion Meltdown That Swept Through Software Stocks’, Wall Street Journal (online, 26 February 2026).
[15] Medha Singh and Saqib Iqbal Ahmed, ‘US Software Stocks Slammed on Mounting Fears Over AI Disruption, Lose $1 Trillion in Week’, Reuters (online, 5 February 2026).
[16] Lon L Fuller, The Morality of Law (Yale University Press, rev ed, 1969) 39; Joseph Raz, The Authority of Law: Essays on Law and Morality (Oxford University Press, 1979) 213–4.
[17] Australasian Institute of Judicial Administration, Guide to Judicial Conduct (4th ed, March 2026) 13 [3.3] (emphasis added).
[18] His Holiness Leo XIV, ‘Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence’ (Encyclical Letter, 25 May 2026) [99].
[19] His Holiness Leo XIV, ‘Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence’ (Encyclical Letter, 25 May 2026) [99].
[20] His Holiness Leo XIV, ‘Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence’ (Encyclical Letter, 25 May 2026) [198].
[21] His Holiness Leo XIV, ‘Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence’ (Encyclical Letter, 25 May 2026) [198].
[22] His Holiness Leo XIV, ‘Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence’ (Encyclical Letter, 25 May 2026) [199].
[23] His Holiness Leo XIV, ‘Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence’ (Encyclical Letter, 25 May 2026) [118].
[24] His Holiness Leo XIV, ‘Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence’ (Encyclical Letter, 25 May 2026) [118] (emphasis omitted).
[25] Christian Catalini, Xiang Hui and Jane Wu, ‘Some Simple Economics of AGI’ (Working Paper, 26 February 2026) 2.
[26] Christian Catalini, Xiang Hui and Jane Wu, ‘Some Simple Economics of AGI’ (Working Paper, 26 February 2026) 8.
[27] Christian Catalini, Xiang Hui and Jane Wu, ‘Some Simple Economics of AGI’ (Working Paper, 26 February 2026) 9.
[28] Supreme Court of Victoria, ‘The Use of Artificial Intelligence by Judicial Officers’ (Guidelines, 14 May 2026).
[29] Richard A Posner, Reflections on Judging (Harvard University Press, 2013) 46.
[30] Richard A Posner, Reflections on Judging (Harvard University Press, 2013) 46.
[31] Richard A Posner, Reflections on Judging (Harvard University Press, 2013) 240 (citations omitted).
[32] Richard A Posner, Reflections on Judging (Harvard University Press, 2013) 241.