Administrative Justice in the Age of Artificial Intelligence

This is the fifth in a special series of posts reflecting on 25 years of the Gilbert + Tobin Centre of Public Law. To see other posts in this series, click here.

Janina Boughey

16.10.2026

The rapid development of generative Artificial Intelligence (AI) in the last few years has transformed the nature and pace of work in many sectors. Many of these transformations seem overwhelmingly positive. AI can, for instance, apparently improve the accuracy and speed of medical diagnosis and accelerate medical research. It can transcribe, translate, and caption meetings, improving accessibility and participation, and decreasing the workload involved in record-keeping. Other uses of AI have been controversial, for a host of reasons. For example, students’ use of Large Language Models (LLMs) to do their school and university assignments has led to widespread concerns about ‘cognitive surrender’ – that is, students failing to develop required knowledge or skills in research, writing, and critical thinking. LLMs have also drastically increased the workload of many courts, tribunals, and complaints-handling bodies.

The use of AI by Australian public servants

Like virtually everyone else whose main work tool is a computer, public servants are using AI in their work. So far in Australia, little attention has been given to the benefits and risks of public servants using AI to perform, or assist them in performing, their various functions, or to regulation and oversight of public servants’ use of AI. The Australian Government’s Policy for the Responsible Use of AI in Government (Policy) can be described as overwhelmingly ‘green light’ in its approach. It is very much focused on encouraging the use of AI by public servants and speaks of the ‘possibilities’ and ‘opportunities’ of AI. This encouragement is facilitated by the roll out of GovAI – a bespoke LLM for Australian Government agencies (though public servants are also allowed to use publicly available products, subject to security classification).

The Policy refers to ‘safe, ethical and responsible use’ of AI, and to accountability. However, it is left to each agency to develop standards, determine when the use of AI is justified and ethical, decide what information to provide about their use of AI, and to identify risks and determine how to mitigate them. There is an AI Review Committee, comprised of nominated senior public servants with particular expertise in national security, human resources, and privacy, which provides ‘non-binding advice on AI use cases’. But ‘agencies remain responsible for identifying, assessing, managing and monitoring AI use case risks’.

I do not suggest that the Australian Government is oblivious to the risks of AI. While it originally adopted an extremely enthusiastic approach focused on innovation and investment opportunities, in more recent announcements it has begun to place guardrails on certain aspects and impacts of AI. The particular risks on which the Government is currently focused are the environmental impacts of data centres, intellectual property rights of the creators of ‘data’ on which AI is trained, and the use of AI to create harmful or abusive material. To date, at least, little attention seems to have been given to the risks of public servants using AI to perform the work of government – administering laws, making administrative decisions, delivering services, developing policy, evaluating policies and programs, advising ministers, and communicating with the public, amongst other things.

One thing that is clear is that public servants are using AI for a range of tasks. The Australian Government’s Policy has required agencies to issue AI Transparency Statements, explaining their uses of AI (though not all have). These statements have been (fairly) criticised for being variable in quality and difficult to find. They have improved over the course of this year, but most remain very vague in their descriptions. It is nevertheless clear from the AI Transparency Statements that AI is being used on a routine basis in most government agencies, for tasks such as summarising material, drafting advice, recommendations and correspondence, undertaking risk assessments for compliance and fraud detection, image processing, and a range of other tasks. Research and reports from New South Wales (NSW) and Victoria suggest similar uses of AI by state and territory bureaucrats. 

How should governments’ use of AI be regulated?

The Gilbert + Tobin Centre of Public Law has been at the leading edge of research on governments’ use of technology, through our Public Law and Technology Project and Administrative Justice Project. Our collaborative project on The Automated State brought together law and technology academics, judges, public servants, and parliamentary drafters to explore specific practical and theoretical aspects of the challenges and opportunities of governments’ use of technology from a public law perspective. Core findings in this and related research include the fact that government’s reliance on the private sector for technology affects access to information, how a desire to automate government processes shapes the design of legislation, and the public law implications of the digital tools that many government agencies use to provide information about the law.

Our work in both projects is unashamedly normative. We are concerned with how core public law and administrative justice values – including transparency, accountability, rationality, lawfulness, and dignity – can be protected when governments modernise, morph, and transform the ways they operate. At the same time, we recognise the potential benefits that these changes may bring for other administrative law values and virtues, such as efficiency, consistency, and cost-effectiveness in the use of public funds.

When it comes to the use of AI to perform, and assist in performing, government functions, discussion (and regulation) should be focussed on how to protect administrative law values and how to balance these values (and other benefits) when they come into tension. For instance, in decision-making about the level of support a person needs due to age or disability, AI (or automation more broadly) might offer real benefits in terms of efficient decision-making. This could mean that a backlog is cleared, cases are decided promptly, and people receive the support they need quickly rather than it taking several months or years to do so. Efficient administration is a key element of administrative justice. A system can have all the accountability and lawfulness in the world, but a lengthy wait for an outcome on an issue with significant implications for a person’s life and wellbeing, means the system is not fair or just.

In some respects, the use of AI is a continuation of the move to automate government processes. This is often the way it is spoken about by governments and commentators. Indeed, the example I have just mentioned comes from an algorithmic system used in aged care decision-making, not one using ‘AI’. (I note that there are different uses of the term ‘AI’ and some use it to encompass all machine-based systems, but here I take it to mean systems which operate with a level of autonomy and adaptiveness.)

Scholars and accountability bodies have been concerned with governments’ use of non-AI automated systems for some decades. Our concerns include the challenges with (administrative law) explainability of automated systems and the flow on obstacles for accountability, the risk of bias, as well as the ability of flawed automated systems to make harmful errors at scale. Professor John McMillan, as Commonwealth Ombudsman, launched a better practice guide for agencies using automated assistance in decision-making in 2007 (updated last year). If followed, those principles would have prevented the Robodebt scandal. But they were not, which highlights the need for mandatory, independent oversight of agencies’ use of technology to ensure that it complies with the core values of administrative justice. The NSW Ombudsman has also provided outstanding analysis and guidance on the use of machine technology in administrative decision-making. Most of this existing guidance applies equally to the use of AI. My colleague, Dr Scarlet Wilcock and her co-authors have argued that the focus of regulation should be on the risks presented by a system’s use in context, and not on the kind of technology used. I agree.

But AI, and LLM assistants in particular (such as ChatGPT and Claude), present new challenges and exacerbate others. For instance, the explainability challenge is far greater – arguably insurmountable. It is possible to understand how a non-AI algorithmic system reached a particular outcome. There may be many barriers to that understanding. The explanation may be difficult for someone without technical expertise to comprehend. There may be a gulf in communication between the people who understand what a statute means and those who understand how an automated system works. And if government has used the private sector to develop the system one or both parties might refuse to reveal how it works (see, O’Brien v Secretary, Department Communities and Justice [2022] NSWCATAD 100, [40]–[44] (Senior Member McAteer)). But it is technically possible at least to know how an administrative decision made by a algorithmic system was reached and therefore to know whether it was lawful or not. The same is not true of AI. The best that can be achieved is a systemic understanding – how the system was programmed and the data on which it was trained. A person with sufficient expertise could make an educated guess as to how an individual decision was made, but they cannot know for certain.

Perhaps more concerning is that LLM assistants can provide reasons for their decisions. These mimic the reasons a human could provide and give the illusion of an explanation, but the two are not the same. The reasons produced by LLMs are based on probabilistic predictions of word patterns. LLMs have been described as ‘stochastic parrots’: they do not actually ‘understand’ what they produce, nor do they ‘think’ or reflect the actual reasoning process used. Because they are trained on human language, they give the illusion of being reasons for the AI system’s output. This illusion is dangerous because it may increase the human tendency to trust the outputs of machine systems, and forego critical thinking about its correctness. It is also dangerous because, as experiments have shown, the reasons given by LLMs can hide biases.

Another important feature of LLMs, which distinguish them from algorithmic systems, is that they are generic and adaptable, and so are being deployed by individual public servants for a huge array of tasks. Typically, the introduction of an algorithmic system to perform some part of a decision or policy-making task requires the planned and deliberate development of the system, approved at the agency level. In that context the NSW Ombudsman’s guidance that lawyers be involved in the design of automated systems, to ensure the law has been correctly interpreted, makes sense as a risk-management strategy (at 51). By contrast, the widely used, publicly available LLMs, and the bespoke systems developed from them, are generic and adaptable. They are not designed for specific decision-making tasks, but can be used to perform a range of similar types of tasks, such as summarising large volumes of text, searching for information, and drafting documents. Any individual public servant might use these tools to assist them in performing these, and other, aspects of their functions. This makes regulation of their use a different endeavour. Agencies will need to develop principles about when and how it is acceptable for public servants to use LLMs in their work, and approaches for ensuring those principles are followed. As I noted above, but want to reiterate, the Robodebt experience strongly suggests that there ought to be external oversight of the way agencies do this – perhaps through codes of conduct and public service commissions.

It seems that currently, when it comes to the use of AI in government, the focus of the Australian Government has been more on its potential benefits than the risks its uses may pose to administrative justice. Australia is not alone in this respect. Only a few jurisdictions seem to have given serious thought to the regulation of AI. The European Union’s Artificial Intelligence Act is widely regarded as the leading model. It classifies the ways in which AI might be used by the potential harms and risks (at Ch III, art 1), and primarily places obligations on AI developers of high-risk systems (at Ch III, art 3), with lesser obligations imposed on organisations which use high-risk systems. It bans systems which do certain things, such as social scoring and generating abusive material (at Ch II, art 5). It places some minimal obligations on the organisations which use high‑risk systems, but its principal regulatory focus is developers. For the most part, the Act does not distinguish between public sector and private sector organisations which use or ‘deploy’ AI. I think it is important to do so because identical uses of AI in the private and public sectors simply do not raise the same concerns. It is a different thing, for instance, if a private firm uses AI to draft its communication with clients than if a government agency uses AI to draft reasons for its administrative decisions. Serious scholarly and regulatory attention must be given to protecting administrative justice when public servants use AI in their work. These are questions which our Public Law and Technology and Administrative Justice Projects will be exploring over coming months and years.

Janina Boughey is a Professor in the UNSW Faculty of Law and Justice and the Director of the Gilbert + Tobin Centre of Public Law.

Suggested citation: Janina Boughey, ‘Administrative Justice in the Age of Artificial Intelligence’ (16 October 2026) <https://www.auspublaw.org/blog/2026/10/administrative-justice-in-the-age-of-artificial-intelligence>