Inside Monash’s Frontline Thinking on AI and the Law
“AI Will Absolutely Replace Junior Lawyer Work — But Not Junior Lawyers”
Lawtitude recently sat down with Dr Paul Burgess, Senior Lecturer at Monash Law and Deputy Director of the Digital Law Group, to talk about black boxes, hallucinations, disclosure and why he'd like all of us to eventually stop talking about AI altogether.
Author: Kabeer Sahni, Editor: Seth Lahanis
Dr Paul Burgess’ story is truly unique for someone at the forefront of AI and the law. Having begun his career as a detective in the police force, Dr Burgess would occupy various roles outside of law before pursuing a Juris Doctor from the University of Queensland. He then spent half a decade working in litigation before obtaining a PhD at the University of Edinburgh, examining the political thought of 17th-century philosophers Locke and Hobbes. “There’s no update coming for what they’ve said,” he laughs. AI, by contrast, “is literally changing on a weekly basis.”
Today, Dr Burgess teaches the ‘AI for Lawyers’ unit at Monash University, serves as Deputy Director of Monash University’s Digital Law Group and recently co-founded the TAILER (Technology and AI in Legal Education Research) Lab alongside colleagues in the law, information technology and education faculties.
Lawtitude met with Dr Burgess to discuss the tension which AI creates for long-standing legal principles, how students and practitioners should be using such tools and why he believes the profession’s junior ranks are far safer than headlines suggest.
The Black Box Problem is Not New
Ask any lawyer their hesitations about AI in legal decision-making, and you’ll likely hear some version of: ‘we don’t understand how it arrives at its answers.’ “We know it’s doing things, but we don't quite know how it’s doing things,” Dr Burgess cautions regarding large language models (‘LLMs’, such as ChatGPT and Claude).
“Everyone who has ever been in a class with me... will know about the rule of law, and will know it’s about accountability, predictability and certainty.” At face value, this seems to discredit the appropriateness of AI in legal decision-making. If the law demands transparency and AI cannot explain its reasoning, the role of AI in legal decision-making must be fundamentally limited. However, Dr Burgess challenges this reasoning; humans, he argues, are not the transparent gold standard we assume them to be.
“Judges [give] us the best possible version of why they’ve made a decision,” he says. “But if you actually look at what neuroscientists will say, they will come out and say, no, that’s not why you made the decision… They give us a sort of proxy of what the decision is.” Ask an LLM how it arrives at its answer, and you’ll receive a similar response; “an approximation of the reason... the reason that we think we might want to hear.”
All of this is not to imply the ‘black box problem’ is non-existent or unimportant. It means, in Dr Burgess’ view, the profession must determine an acceptable standard of transparency rather than assuming that humans meet a ‘gold standard’ which AI conspicuously fails to satisfy.
Old Workflow, New Workflow and the Risk of Missing Steps
Dr Burgess is also wary about the implications of substituting traditional legal research for the simpler prompt-based research that LLM's can perform on behalf of lawyers.
Traditional research methods involve understanding the question, identifying the area of law, working out where to find the answer, researching it, then drafting and presenting it. In contrast, much of this scaffolding simply disappears when using AI-assisted shortcuts.
“Missing steps can be great,” he says, but only “if you already know what the answer should be.” Otherwise, missing those steps deprives you of the process by which you would have learnt why the answer is correct, leaving you without the means to verify what you’ve been told.
Indeed, Dr Burgess applies this distinction to his own work. “I use Gen AI every day. I use Agentic AI every day. I am a massive Claude Co-Work fan, it runs most of my life,” though strictly as an efficiency tool for answers he already knows to be correct. “If I don't know the answer... Gen AI, even a system that is a law-specific one, isn't the best solution. You actually have to know what the answer is before you can effectively use it.”
‘Which AI is Best?’ is the Wrong Question
Dr Burgess works across the major commercial LLMs and is also familiar with law-specific platforms such as Harvey AI and Legora. His conclusion is that there is no clear “best” choice; rather, the right tool depends entirely on the task.
Generic models perform well at summarisation, he says, but are weaker on the “super soft skills” which law graduates spend years absorbing, like knowing which of five judicial reasons in a case are binding. Law-specific models are instead trained on narrower legal corpora, and often use retrieval-augmented generation to ground their answers, thus reducing the frequency of hallucinations.
In a study which analysed a New South Wales database of cases, containing a mere ~5% of a much larger training corpus based on multiple international jurisdictions, results became noticeably unreliable. “The systems get quite overawed or confused with the amount of data that's in there,” he explains. For a jurisdiction as small as Australia, that's a real problem for any model trained broadly across “all of the legal data” rather than something more targeted.
Dr. Burgess notes that general-purpose AI tools like NotebookLM, when pointed at a small, specific set of documents, can outperform law-specific AI systems on narrow questions. This suggests that prompting methodology may matter more than which AI system you're using.
Hallucinations are Baked into the Cake; Verify Like a Lawyer
Dr Burgess provided guidance on how to safeguard against hallucinations, which he asserts will never be fully solved. “Knowing and accepting is not just the first step, but it's absolutely the most important step,” he says, enumerating a list of (non-exhaustive) failures AI is susceptible to making with the law:
The model fabricates a claim outright.
The model gives a citation for a real claim, but the citation doesn't actually support the claim.
The citation does support the claim but comes from an unreliable source such as an unverified paper, student note or uncited journal article.
Everything checks out but the underlying authority itself has been misrepresented as more settled or significant than it is.
“You've still got to do the work,” he says, “even when [an output] looks really convincing.” That includes clicking through to sources and not immediately accepting that an AI-generated citation truly exists. This is something he flags as an increasingly common failure point among students and practitioners alike.
Who’s Really Asking for Regulation?
On regulation, Dr Burgess describes himself as skeptical that heavier-handed rules are necessarily ‘the right response’. He's more interested in whether existing legal concepts, including the rule of law itself, can achieve the same outcomes without new legislation. He also points to a pattern of established players in the AI space calling vehemently for increased regulation. “Established market leaders (such as Claude and ChatGPT) will always try to call for more regulations, because it's... getting up and then pulling the ladder up behind you”. He notes that whilst such calls for action are not necessarily made in bad faith, neither are they necessarily unbiased.
Should AI Providers be More Cautious in Providing ‘Legal Advice’?
Dr Burgess contends that AI’s tendency to sound confident in its own answers, even when it shouldn't be, is a problem worth taking seriously. Legal AI tools, he argues, should be trained to hedge the way a competent human professional naturally would. Whereas “a lawyer in practice giving legal advice absolutely should” decline to answer outside their expertise, LLMs generally don't do that, and instead seem to always have an answer whether accurate or not.
Disclosure and Privilege Concerns
Among the unresolved questions raised by Dr Burgess is the extent to which lawyers should be disclosing their AI use to clients, and whether an expectation of disclosure should sit alongside better-established, quieter forms of delegation. “Did I disclose the fact that I used a paralegal?” he asks. “Did I disclose I used a legal secretary?... Do you disclose that you used electricity?” His point isn't that disclosure doesn't matter, only that the profession has yet to develop the “innate professional sense” for where the line sits with AI. Some firms, he notes, are already building opt-in/opt-out options regarding the use of AI, allowing for clients to choose whether AI can be used on their matter.
Dr Burgess then raised an interesting perspective regarding clients who submit their own sensitive material to LLMs before engaging a lawyer. “Have they now waived professional privilege by putting that material out in the public, even before they've engaged a lawyer? ... [We] don't know right now.”
For now, it seems the pace of AI adoption has outstripped the development of regulatory frameworks by which to address these concerns.
The TAILER Lab
Dr. Burgess co-founded the TAILER Lab on the premise that lawyers "don't have all of the answers" when it comes to a shift of this scale, bringing together a multidisciplinary approach to AI and the law.
The Lab sits at the intersection of law, education, and computer science, and examines how generative AI should reshape legal education, from the skills students need to develop to the way clinical and experiential learning is structured. It now has 42 members across Monash Law, Monash's Education, IT and Business faculties, and external partners including UCL, the University of Melbourne, and Justice Connect.
The Lab currently runs around 14 projects, two of which caught our attention:
The New Delegation Problem, which examines how generative AI disrupts professional formation by enabling lateral delegation of tasks to machines, with far-reaching implications for legal and professional education.
The Clinic AI Assistant, an internally facing AI tool designed to support Monash Law students in community legal clinics to assist on matters when a human supervisor isn't immediately available and assist in extending clinic capacity to improve overall access to justice.
Beyond its flagship projects, the Lab has published research in Artificial Intelligence and Law and has an ARC Linkage application currently under assessment, one of several funding pathways it's pursuing as it works toward becoming a self-sustaining, internationally recognised research program.
What will Happen to Junior Lawyers?
When asked directly about student anxiety over AI replacing junior lawyer work, Dr Burgess didn't soften the answer: “AI will absolutely replace junior lawyer work.” But “It will absolutely not replace junior lawyers.” His reasoning is grounded in the belief that senior lawyers need a pipeline of junior lawyers coming up behind them, which means the junior role must keep existing in some form, even as its content changes.
He's also candid that some of that change is welcome. Reflecting on his own early career, Dr. Burgess recalled a moment when he was "locked in a room with 12 big boxes filled with documents," conducting manual document discovery, a task AI can now complete in a fraction of the time. That said, Dr. Burgess doesn't believe junior lawyers need to spend months on this kind of work simply because he once did. He does think, however, there's value in doing enough of it firsthand to understand why it matters, before letting AI take over the rest.
Where this All Settles
Dr Burgess' long-term prediction is, in its own way, a vote of confidence in the profession. Within five years, he'd like the sector to largely stop talking about AI as something separate from legal practice at all. “We don't talk about using a pen or a computer or the internet in legal practice. They're just tools in the way that we do law, and I think AI will inevitably go the same way.” “Lawyers are absolutely like cockroaches,” he adds. “We will survive absolutely everything.” The tools will keep changing, likely faster than any of us can comfortably track.