Want to listen to the full episode and all our other episodes?
Hearsay allows you to fulfill your legal CPD requirements every year.
Our yearly subscription is only $299/year.
With a yearly subscription, you can access all of our episodes AND every episode we release over the next year.
Episode 107
Buy Episode
Ethics and Professional Responsibility Practice Management and Business Skills
Examining the Future: What’s Next for Legal Education in the Era of Gen AI?
Law as stated: 1 December 2023
What is this?
This episode was published and is accurate as at this date.
The Chief Examiner of the Victorian Bar Entrance Examinations, Dr Jason Harkess, dials into Curiosity to uncover the adoption and use of gen AI technologies in legal studies and practice. Touching on the implications of the technology for universities, students, and examiners.
Ethics and Professional Responsibility
Practice Management and Business Skills| What area(s) of law does this episode consider? | Generative AI in legal study and practice |
| Why is this topic relevant? | The law is not exempt from the influence of innovation. As lawyers grapple with the increasing presence of generative AI in the profession, understanding the nuances of its use has become a paramount concern. Education is, of course, one of the areas most impacted by AI – consider the ability of ChatGPT to generate essay length responses to prompts in just seconds. Those very same legal hypotheticals beloved by law schools – which would have occupied a law student for far, far longer – can now be fed to a machine for analysis and response. But with reasoning and then articulating that reasoning being such a crucial component of lawyering, what is the right approach to using these tools in legal education, examination, and practice? |
| What are the main points? |
|
| What are the practical takeaways? |
|
DT = David Turner; JH = Jason Harkess; RD = Ross Davis
| 00:00:00 | DT | Hello and welcome to Hearsay The Legal Podcast, a CPD podcast that allows Australian lawyers to earn their CPD points on the go and at a time that suits them. I’m your host David Turner. Hearsay The Legal podcast is proudly supported by Lext Australia. Lext’s mission is to improve user experiences in the law and legal services and Hearsay The Legal Podcast is how we’re improving the experience of CPD.
If there’s one message that the legal professions learned this year, 2023, it’s that the law isn’t exempt from innovation. As lawyers grapple with the increasing presence of generative AI in their firms, in their profession, understanding the nuances of how it works and how it can be used has become a paramount concern.
Today on the podcast, we’re talking about ethics and generative AI in legal study and in legal practice. Education is of course one of the areas already most impacted by artificial intelligence. Just consider the ability of a model like ChatGPT to generate essay length responses to prompts in mere seconds. Just recently in the United States, the University of Michigan Law School banned the use of ChatGPT for essays accompanying applications to the school. In practice, detecting ChatGPT generated responses is substantially trickier. Now those very same legal hypotheticals beloved by law schools, which would have occupied a law student for a long, long time, can now be fed to a machine for analysis and response.
But with reasoning and then articulating that reasoning being such a crucial component of lawyering in the real world, what’s the right approach to using these tools in legal education, in examining our students, and in practice? Joining us today to discuss this question and more is Dr. Jason Harkess, a Foley’s List barrister specialising in criminal and administrative law. Dr. Harkess is also the chief examiner of the Victorian Bar Entrance Examinations and Principal Examiner for the Law Institute of Victoria’s Criminal Law Specialist Accreditation Program. He also lectures in evidence, administrative law, and contracts at Monash University and is a member of the Victorian Legal Admissions Board’s Academic Course Appraisal Committee. So, he is more than equipped to talk about legal education and artificial intelligence here today.
Jason, thank you so much for joining me on Hearsay.
|
| 00:02:25 | JH | Thank you, David, and I’m happy to be here.
|
| 00:02:27 | DT | Now before we dive into our topic today, I thought I’d ask you to share some of your highlights from practice and how you became so involved in legal education.
|
| 00:02:35 | JH | All right, well I’m happy to start from the beginning. I studied law at law school like most, but not all, lawyers straight after high school. That was in New Zealand where I’m originally from. I studied at Auckland University and did the undergraduate LLB with a BA for five years. I always was fascinated with law conceptually and thought at some point I might actually get into academia, but I did what most people did and started out practice at a large commercial law firm in New Zealand. I worked there for a couple of years, then went to a smaller firm and practiced in criminal prosecutions as well as civil litigation. And then ultimately there were opportunities that presented themselves in Australia. And I came to Australia in 2002, worked for about a year as a solicitor in a commercial litigation firm before ultimately wanting to get more involved with court work. So, I came to the Victorian Bar in 2003. Being a sole practitioner, essentially self-employed, sole trader, I could do other things as well. And so, I actually enrolled in a master’s gravitating towards my academic inclinations. That became a doctorate and because of my interest in higher education and law, Monash University, when I was enrolled in the doctorate there, offered me the opportunity to start teaching, which I quickly seized and actually became quite fascinated and engrossed with that for many years and was very much drawn to the idea of being able to tell other empty minds all about the law and what they should know about the law and describing it in a way that I thought it should be described. And so, that was obviously satisfied in terms of that particular desire on my part. But also, as I was doing it, I realised as I was teaching, when you’re teaching, you’re already forced to actually learn it much more thoroughly as a teacher than a student. And so, every time you teach a particular subject – and I’ve repeatedly taught administrative law and evidence and contract law – every time you teach it, the knowledge becomes more and more entrenched and every time you teach it, your understanding of it becomes slightly more nuanced. So, I continued with that while maintaining a practice. Then because of my straddle into both practice and academia, I was identified by the Victorian Bar as somebody who could actually work in introducing an examination for barristers and that was introduced in 2011. And I’ve been involved in designing that ever since, and was appointed the chief examiner in 2013, and have written every single bar exam since then, the most recent of which was on the 8th of October with 265 candidates sitting the exam, which is now in the process of being marked. And I suppose the rest is history.
|
| 00:05:35 | DT | Now, I don’t think I asked you, what was your PhD topic?
|
| 00:05:40 | JH | It was actually an SJD, Doctor of Juridical Science, which is the professional law doctorate. And the thesis, some might say it was unrelated to the subjects I teach, but touching on it, it was on the moral implications arising out of the criminalisation of anti-competitive cartel conduct in Australia.
|
| 00:05:58 | DT | Right. The moral implications. I know it’s not our topic for today’s interview, but my curiosity is piqued. Tell me a little bit about the thesis.
|
| 00:06:05 | JH | So, cartel conduct, for those who don’t know, is an area of competition law. It’s often regarded as the most egregious kind of violation of competition law, where powerful market players get together and agree on setting a price rather than actually independently setting their prices so that consumers are presented with cheaper prices and the competitors are trying to outdo each other. So, it’s essentially the prohibition against collusion. And in Australia, it hadn’t been a criminal offense until 2009. And of course, in the United States, it’s always been a criminal offense. And so, Australia introduced criminal penalty provisions. It would be regarded as a classic white-collar crime. And so, the question is; “Well, given that the criminal law was designed to punish morally reprehensible conduct, is this really a moral offense?”. And so, I suppose the central question for my thesis was, is this really a moral offense, appropriately sanctioned by the criminal law? And so, I spent some 60, 70,000 words talking about that.
|
| 00:07:13 | DT | Now, I’m sure we could spend the next hour talking about that topic, but we’d better get onto the one that we had planned, generative AI and how it will impact on teaching and testing legal knowledge. Before we talk about generative AI in that specific context, though, I wanted to ask you, Jason, how you’ve been using generative AI yourself, whether in your academic work or in your practice work. Are you using it? Do you find it helpful?
|
| 00:07:37 | JH | Well, the gen AI as a concept really only came to my attention probably late last year, like everybody else in the world, when suddenly everybody started talking about this idea of ChatGPT. And I think I fumbled my way on the internet, looking how to access it and couldn’t, because I didn’t know how. And then eventually one of my own students showed me how to log into ChatGPT. They sent me a link and explained how I did it. And they were saying it can actually be very useful. And so, that’s pretty much how I see gen AI at the moment in terms of the main manifestation, which is causing controversy, certainly within the legal education context and potentially now in the legal professional context. ChatGPT, of course, most people will be aware, is an open-source application accessible from anywhere in the world on the internet. You simply log in to the site and you enter your particular question. You can articulate your question in the most complex way possible. And gen AI in the form of ChatGPT will generate an answer to that question and hopefully provide you with a satisfactory answer. And of course, when you direct specific legal questions, such as a simple question of; “do I have a contract in situation ABC?”, then the chat bot, ChatGPT, will try and give you an answer. It might actually ask for more information, but you can actually input all the information in the message box, ask whether or not there’s a contract according to the facts that I have given you, and then it will generate an answer. And more often than not, it will detect the fact that you’ve raised a legal issue because you’re referred to contract, and it will draw upon from all of the data it has in its extensive database, the precise limits of which I do not know, but it seems to be very extensive. And it will explain that you have a contract for these particular reasons. And if you’re a lawyer reading this, for example, a contract lawyer or a teacher, you’ll see that they must have drawn upon some kind of repository of legal knowledge containing legal principles about contract law. And it sounds right. It sounds like you have got some pretty good legal advice, if you didn’t know any better, just by putting in that question. So, those are the types of questions that are posed to you within the context of legal educational assessments. So, here we’re getting into assignments and tests that are set for you. You can theoretically take home your test, your assignment that you’ve been set for contract law or evidence or whatever particular unit you’re studying. And of course, these days, most assignment questions are put in the form of a Word document or a PDF document. You can simply copy and paste the assignment instructions into the ChatGPT message box. And then at the end of that copy and paste, you can say, give me an answer. And then the ChatGPT will proceed to attempt to give you an answer. And sometimes it actually generates what would seem to be possibly a passable answer, or at least direct you towards the critical issues that need to be addressed and perhaps elaborated upon before you actually submit your assignment.
|
| 00:10:58 | DT | Well, and I think there’s a few things to unpack there, but let’s start by saying, I think ChatGPT is a reasonable analog for the range of tools that are now available. It’s certainly the most widely used, certainly the most widely known. Our own sister company, Ask Lexi, obviously provides a generative AI tool that is more legal domain specific. We have plenty of users who are students, and I have no doubt there are some problem questions going into our own application. But ChatGPT is a good shorthand, I suppose, for these range of tools. And we should say we don’t have time to unpack how large language models and generative AI works on this episode. We’ve talked about them on the show before, and we’ll leave some links to our previous episodes in the show notes if you want to explore how large language models work as next token predictors, what the training data that goes into those models provides them with and what it doesn’t, and some of those limitations around accuracy and specificity of answers when they’re not supported by some external source of factual information. But Jason, I suppose you personally, are you personally using these tools as an academic, as a teacher, as a lawyer?
|
| 00:12:06 | JH | I’ve tried and I’ve dabbled in using ChatGPT to assist me in preparing documents. If I can give an example, as an exam writer, I’m frequently writing exams. And when I saw the first time I used ChatGPT, what it could actually do, I just started adding more information to my particular questions and introducing ridiculous propositions to try and confuse the gen AI. And it actually produced a response that was quite impressive in terms of dealing with everything that I actually put. And then I started actually trying to instruct it to generate a hypothetical problem in relation to an administrative law issue. And then I might identify five particular specific principles of law that I wanted to be addressed in the problem. And it needed to be a hypothetical factual problem. And so, they need to be three people. So, feed all of this information, because those are the types of ideas that immediately come to my mind when I first sit down thinking, okay, I need to write an exam question. And so, those are the thoughts that immediately come to my mind. And then of course, when I actually start to do it, it takes hours and hours refining. And I’m thinking; “okay, well, if I can just feed all the brainstorming ideas into the ChatGPT, maybe it can produce something that will leave me very little work to do”, but still being reasonably original because I’ve identified the issues that need to be disclosed in the problem. I just needed to be articulated efficiently, rather than me wasting time, several hours, sometimes days trying to refine the way it’s expressed, things like that when ChatGPT does a pretty good job. So, I did that a couple of times. And I was pretty much not really satisfied with any of the responses, because I suppose probably the primary complaint I had in terms of what was ultimately produced was that when I had specifically identified particular principles that needed to be tested, which of course, were particular to the unit that I was teaching at the time, which has been quite prescriptive with respect to Australian law, sometimes with respect to Victorian law, sometimes with respect to particular cases, case authorities and the nuanced principles that those particular cases stand for. I think my expectations of what ChatGPT would produce were far too high. And I quickly came to the understanding that either the capabilities of the gen AI being able to identify and extract those particular principles from those particular cases, it just hadn’t reached that level yet of being able to engage in those kind of complicated algorithmic calculations to draw upon that information, or the data from which it was drawn simply didn’t have those particular principles. What often came up were statements of general principle that largely reflected general principles developed in overseas jurisdictions, such as the United States. So, I quickly realised that the answers, the assistance that I’d get from gen AI, certainly at this stage in 2023, would generate assistance that created more problems than it was really worth getting that assistance for. So, I’ve reverted back to my traditional method.
|
| 00:15:26 | DT | Yeah. I think a lot of legal professionals have definitely had that experience of experiencing hallucinations, as we call them, in the output of large language models that have been given a task that requires the answer to be accurate. We’ve, of course, all heard that famous example from New York State, the unfortunate attorney who submitted a number of fictitious case references in his submissions. We don’t need to rehash that. But there’s certainly that risk with using a large language model that doesn’t have that external source of truth. And again, without talking about our own product too much, it’s that ability to retrieve reliable information from an external source that can set some other tools apart that are specific for the legal use case. But despite those challenges, there’s certainly a tool that’s becoming popular with students, both undergraduate students and students who are completing postgraduate coursework. Because whether or not the artificial intelligence is the ultimate source of truth or whether it’s just helping you to solve that blank page problem of getting started, there is an enormous efficiency gain to be made, I suppose, if we leave the ethical or moral considerations to one side, in using a tool like this. So, I expect that there will be a market impact from this technology on the way that we study and the way that we test knowledge of the law in formal education. Are you seeing the impact of artificial intelligence in your courses at the moment? Do you know if students are using it?
|
| 00:16:58 | JH | That’s a very good question and one that’s being discussed amongst my colleagues who teach law quite regularly, but it’s sort of up in the air at the moment. I’m sure everybody appreciates that there’s a real concern with respect to law schools who are charged with the responsibility of educating people who wish to enter the legal profession, that the people they are saying have actually met the requirements of their course, meet the minimum eligibility requirements to actually enter the legal profession. And the way that has traditionally been ensured is that we know that this particular individual, when we set them this particular assessment task, produced this particular piece of work. And we know that came from their body of knowledge stored in their particular mind, because we’ve got these systems in place, primarily in relation to the most important legal units which are required to be passed. Those have to be undertaken in an invigilated examination. So, as a student, you have to attend a big room. This is traditionally, of course, with paper-based exams. And there is somebody from the university literally watching you and standing over your shoulder to make sure that it is indeed you who is answering the question and you’re not getting any external assistance from anyone. But now what we’re presented with is something a bit more sophisticated. TIP: Those core units that Jason’s just described are sometimes known as the Priestley 11. These units are required to be completed by all students seeking admission to legal practice, and they’re named after John Priestley, chair of the Law Admissions Consultative Committee in 1992. They can be found today in the Uniform Admission Rules 2015 at Schedule 1. They include criminal law and procedure, torts, contracts, property, equity, company law, administrative law, constitutional law, civil dispute resolution, evidence and ethics. Are these 11 subjects the essential topics that young lawyers today really need to know? Have they kept up with the times? We’ll leave you to reflect on that.
|
| 00:19:08 | JH | Then, of course, there’s the issue of assignments, which of course are take-home assignments, which make up a lesser component but are still part of the assessment. How do we stop cheating occurring there? Because, of course, what’s stopping you from copying and pasting something you’ve found on the internet? And so, over the last 20 years or so, there have been software applications that have developed where students have to submit their assignments. Perhaps the most well-known one is Turnitin, which then analyses a particular assignment and generates a probability of plagiarism or how much appears to be derived from other sources. And so, that’s a significant deterrent in making sure that students actually produce their own work.
TIP: Those of our listeners who studied on or after the mid-2000s will probably be familiar with Turnitin. It was founded in 1998 in the United States, but it’s now used extremely broadly all over the world by universities as the leading form of plagiarism detection software. It’s basically a SaaS or software as a service tool. Students upload their exams, their assignments, and those exams and assignments are compared with an enormous corpus of other original student works. It’s an effective tool for detecting similarity between two answers provided by students, but it’s not without its complaints. Some complain that it captures and keeps the original works of students for a commercial purpose, and others say that it’s sometimes prone to false positives where there’s a large number of in-line quotes or only so many ways of expressing a particular concept or idea.
|
| 00:21:12 | JH | But now what we’re presented with is something a bit more sophisticated and more concerning. Which is; how do we know that, when a student is given an assignment task, they haven’t just copied and pasted the question and shoved it into a gen AI application and then taken that, product of the gen AI passed it off as their own in terms of the assignment. And because the wonders of gen AI is that every time you click generate with that particular question, it produces a completely original answer. And so, existing applications like Turnitin are not going to pick up on the possibility that this has been plagiarised. There are of course products coming out into the market which purport to be able to detect whether or not that particular assignment that’s been submitted is a product of gen AI or whether it’s actually an original product of a human being. But I’m not sure whether anything on the market is satisfactory in that respect at this stage.
|
| 00:21:46 | DT | Well, yeah, I mean, so far as Turnitin, that technology is ultimately predicated on detecting similarity between two submitted responses. And as you say, provided your parameters are set with a sufficiently high temperature or equivalent parameters, ChatGPT or any other model will produce a largely original response. It might be similar in structure. It might be relatively predictable, but in terms of the language used entirely original. I had heard of one law school in Australia taking this attempted approach to detecting ChatGPT generated responses – and I have to say it’s entirely misconceived – was to generate their own answer to the problem question or to the assignment using ChatGPT. Do that a couple of times, do three or four of them, submit all of those to Turnitin and then wait for Turnitin to detect the similarity of student answers to those answers. But of course, it’s a pointless endeavour because there’s no guarantee that a fifth answer generated by ChatGPT is going to resemble any of the earlier four. So, that’s an ineffective technique. Of course, there is the hope that we’ll come across a reliable tool for detecting generative AI output. Unfortunately, so far as I’ve seen, most of those tools are not performing reliably in production. Of course, the most famous one, earlier this year, OpenAI had its own free classifier that could provide you with a percentage likelihood that certain text that was inputted into the classifier was generated by one of their large language models. They’ve since withdrawn the availability of that classifier because it was producing too many false negatives and too many false positives. There’s just really no hallmark in the way a large language model constructs its texts. That means you can reliably say this is generated by a person or not. There’s of course watermarking technology that is largely theoretical at the moment. I know some businesses are working on watermarking technology to secretly mask output as generated by AI, but that’s largely in the realm of image generating AI, diffusion models rather than text. So, it’s largely a problem without a solution, isn’t it? One solution, and we mentioned it in the intro to this episode, was the one adopted by the University of Michigan Law School. They’ve just said ChatGPT is banned. It’s a breach of the student’s code of conduct to use it for your university coursework. Is that a viable approach? Is that likely to prevent students who want to use it from using it?
|
| 00:24:14 | JH | My own personal view on that idea of banning ChatGPT is I think that’s just pointless, completely pointless. When I discovered ChatGPT and tried different ways of using it just for personal fun but also within a legal academic context, I was very excited about the possibilities that are going to arise from this in the future and it’s only going to get better and better. But obviously I’ve already indicated that I’m not really satisfied in terms of using it as a legal professional at this stage because of its limitations. But I foresee it will only become more and more sophisticated, drawing upon more reliable data pools and identifying which data pools to actually draw upon so, that specific questions can be inputted by anyone, really, law teachers and law students and professionals, and getting out a reasonably comprehensive answer and explanation which they can then take and work with and adapt for their own particular purposes. I don’t think there’s any point in banning it because people will use it because it’s so, useful and it will become ever increasingly more useful. I think the better approach would be to accept the fact that we’re now moving into this next stage of the information technology age and gen AI is going to have a significant impact on the way that students at law school and everywhere else and legal professionals go about their daily business of reading texts, being presented with particular legal problems which they have to answer and then researching and then coming up with a solution. It will provide a very useful tool. In terms of dealing with it at a law school, where of course the primary focus of the law school is to make sure that the person who they’re going to award the law degree to does actually meet the minimum regulatory requirements in relation to standards, we’re producing this law student and claiming that they have got that minimum knowledge set to actually enter into the legal profession as required by the regulations. How can we be satisfied that they do given that there’s that risk which we can’t really stop of them getting assistance from gen AI. One thing to do is to say; “well, look, you can use gen AI, you can use gen AI in completing this particular assessment task but you have to disclose that fact and you have to declare that fact when you submit your assignment”. In fact, some assignments can be set on the basis that you are instructed to use gen AI and you can create a question as a legal educator and part of the instruction is obtain an answer or various answers from a gen AI application and then proceed to critique it and assess to what extent is it correct and identify the deficiencies and so, that’s a way of I suppose implicitly educating up and coming lawyers as to you can actually use gen AI but you have to constantly appreciate its limitations and that you have to know when you actually enter into the legal profession. You are going to be personally responsible for the advice that you ultimately produce in your name and you need to know that gen AI certainly now has made your deficiencies and you need to fix those deficiencies up whenever you purport to rely upon it and so, I think it can be dealt with in that way and as gen AI technology becomes more and more adept at actually producing more precise and reliable legal advice it will just become habitual that lawyers and indeed law students will draw upon it and there’ll be standard disclosures and perhaps a requirement that you identify all parts of the particular submitted assignment that were generated from gen AI and that you’ve verified the authenticity or the reliability of these assertions by your own research. And so, holding law students to account by making that declaration. I think it obviates the need to create a general prohibition against it and then checking to see whether or not they’ve used it which of course is very difficult.
|
| 00:28:21 | DT | Just as you were describing that method of assessment, of saying, “well, use the output as an assessment as a method of critiquing another answer” – effectively marking the AI on its own output – I think what I really like about that is it reinforces the ethical dimension of this both from an academic standpoint but also from a professional standpoint which is that the buck stops with you as the human being. There’s a human in the loop – to use the machine learning terminology there – and it’s that human who’s ultimately responsible for the use of that output. And I suppose we’re teaching not just the practicalities of using a new research tool which may be unreliable or it may be very useful, but we’re also teaching the ethical dimensions and we’re teaching where the responsibility lies for what we’re ultimately presenting to our clients.
|
| 00:29:11 | JH | Yes, there’s an alternative way to address it certainly at law school. And, of course, most law schools are still following the traditional ways of sitting assignments. The biggest problem is in relation to essay assignments where the question is just in the space of three or four lines. If we’re dealing with the subject of criminal law for example there might be an assignment which simply states explain and discuss the concept of the mens rea element of recklessness within the criminal law which of course is gen AI’s dream question because and indeed the students dream question because they can just copy and paste that and they will generate probably reasonably good answer and they can add a bit more and so, flesh it out such that they could probably generate an answer that is actually passable. As we saw in that analysis conducted in the United States of gen AI answers, a lot of these would pass. I suppose the problem is that you probably won’t get an HD or high distinction answer from it but you could get something that’s quite passable because there’s enough information there that if the examiner who’s assessing it takes it on its face; “it seems to disclose that you know exactly what you’re talking about, you’ve done enough research, you seem to understand all the concepts. It doesn’t appear to be plagiarism because the plagiarism detector hasn’t picked it up so, but could have done a few more things, referred to a couple of Australian cases rather than just refer to general principles, but no I’ll give you 60 out of 100”. I mean that’s a pass and there will be some students who will identify that as an easy way of getting through law school as opposed to the high students who want to get more than 60, they’ll be the ones who actually you probably don’t need to worry about in terms of using gen AI. But these particular students who will just see this as an easy way to get through, that needs to be addressed and confronted. And one way of ensuring that this particular student does indeed have the knowledge that appears to be displayed in this gen AI answer they’ve submitted, and we don’t know whether it’s a gen AI answer, we suspect but we don’t know – because there’s no way of really detecting it – why don’t we force every student to supplement that assessment task with an actual in-person oral examination? And that way the examiner will get a pretty good idea as to the extent of the knowledge of this particular individual in relation to the concepts and propositions that they have articulated in their submitted assignment. Obviously, there’s a resource issue there and how many examiners in a particular faculty can be deployed to conduct say 300 oral examinations for the 300 students that are enrolled in that particular unit that semester. But the oral examination is probably a pretty satisfactory way of satisfying yourself as a legal educator that this person does indeed know what they’re talking about insofar as it’s reflective of what they’ve articulated in their written assignment.
|
| 00:32:08 | DT | Yeah, I suppose it certainly is. And we’ve got a kind of a proud tradition of oral assessment, I suppose, in legal education certainly those who’ve finished their practical legal training, I recall a lot of oral assessment being part of that process.
TIP: Now, if you’ve had the experience of being lectured by a practitioner trained in the United States, or who’s studied in the US, or even if you’ve any kind of exposure to American legal media you might have been exposed to a practice called cold calling or to give it a nicer name the socratic method. The idea is to test random students in a class unprepared by calling on them to answer a question about the topic under discussion. There is a few different ways to conduct the process. For example, if a student’s unable to answer a question, requiring them to find the answer in a textbook or reading, and come back to orally answer the question later in the class. Alternatively, and maybe a little bit more mean, in the case of a correct answer deliberately probing that student as far as possible until they eventually get an answer wrong.
|
| DT | The challenge there is that the trend of asynchronous teaching and remote teaching is pulling in the other direction, isn’t it? Since COVID-19 when, by necessity, we had to teach increasingly online and increasingly make adjustments to the way we assessed and taught to cater for our inability to be together. That mode of teaching has persisted and there’s more online teaching there’s more online only courses, and I suppose more online only assessments. Are we being pulled in two directions here by two different technological influences? On the one hand a preference for remote work and remote learning and on the other a need to safeguard the of our learning from generative AI which really requires that level of in-person either invigilated exam or oral examination.
| |
| 00:34:02 | JH | So that’s a complex question that raises a couple of issues. Well, two main issues of course. The first one is as you’ve said law schools and universities generally are now gravitating towards remote learning because it’s been found to be efficient in the sense that obviously we’re forced to do this as a result of COVID-19 and discovered the technology available to create these efficiencies and delivery of teaching. A significant part of that is that lectures can be recorded and saved and uploaded and downloaded or live streamed or streamed at a later time at the convenience of the particular student who’s enrolled. Assignments can be uploaded and you’ve got so many different applications through which assessments can be undertaken online. So, there’s that desire to keep going that way. But, of course, the other concern is with respect to the ease with which people can pass off others work as their own when they do this. And so, there is a need again to ensure that there’s compliance with the regulations that the law school is presenting to the community. A particular person who they say has the minimum level of knowledge that would qualify them to be admitted into the legal profession is in fact that particular person who undertook and completed those assessments. And so, then we have to revert back to the need to actually do live in-person assessments for the law school and the people involved in the administration of the law school need to ensure that this particular candidate is indeed the person who has that knowledge that we claim that they have by awarding them with those particular marks for those particular units. There is, of course, the capability of assessing students online in real-time fashion. Of course, there are what some people refer to as invigilated remote exams. That’s becoming a common practice now, where certainly in some of the major universities where students no longer sit their exam as they did 20 years ago in a big hall with hundreds of other students and have invigilators employed by the university actually standing over them to make sure they’re not cheating. Now many law schools have all of their students sit their exams in the comfort of their own home or workplace and that’s on their own laptop or their own computer and there are remote invigilators. And there are various different software applications that universities have deployed that essentially involve somebody employed by the university sitting at a remote location logging in and literally watching them through the student’s own computer camera to make sure that they’re actually doing the exam. And there are other software applications that ensure that the computer device of the particular candidate is completely shut down from the internet and then once they finish their particular exam answers the application will then upload the answer file and that gets deposited into the central repository at the university and then it’s marked. So, that’s one way of ensuring that it’s that particular student. And of course, there’s always the way of personally interacting with the candidate online, like I am now with you – you’re the student. I’m confident, David, that it’s you there who’s going to be answering my questions and if I throw at you a particular question concerning the law subject that you’re studying I know that you’ll have no idea what question I’m going to ask because nobody else has seen the question you know I’ve prepared.
|
| 00:37:50 | DT | We’ve been talking about how some of our traditional approaches to assessing knowledge better or worse suited to the advent of generative AI but I wonder whether there’s a silver lining to what we’ve largely been describing as a cloud in that generative AI and artificial intelligence more generally can offer opportunities for new forms of assessment. It’s always occurred to me that there’s a level of unreality about the way we sometimes assess legal knowledge especially for say a take-home exam or an essay where as you said we’re responsible for establishing that this student has the requisite knowledge to join the profession. We establish that knowledge largely by recitation of the course material in long form in an answer that’s prepared over a relatively short period of time in the case of a take-home exam. But that doesn’t necessarily reflect the future reality of legal practice where it might be necessary to demonstrate that knowledge in a dynamic environment where, as you’ve just described, you’re not sure what question you’re going to be asked. And oral examinations are a good example of that, of an examination that reflects a reality of legal practice that a prepared take-home exam or essay perhaps doesn’t. What do you think of this idea of using a generative AI tool to create a simulated scenario for students to participate in to have that kind of experience of speaking with a client or a colleague or another practitioner to demonstrate their knowledge in an environment that’s not invigilated or managed by a human academic but which is designed to provide that dynamic or emergent situation where the student can demonstrate their knowledge in a less predictable way?
|
| 00:39:41 | JH | Well, I’ve seen them in operation. I haven’t used one myself but it sounds to me like an excellent educational tool certainly for the student where essentially the gen AI bot can substitute for a real-life human being and I’ve seen some of those generated conversations and they do come across as the questions that a real-life human being client would ask and then the student can give the answer simulating the role of being a real lawyer and of course that would be an excellent tool which of course could then be used by a teacher in the law and they can then assess the conversation based on the student’s responses to the gen AI bots questions.
|
| 00:40:36 | DT | I suppose we’ve been assessing students for so long with a handful of tools. This would be a fairly novel one by comparison to some of the assessments or forms of assessment we’ve used in the past. Do you see having used generative AI yourself in both professional work and in academia any possible pitfalls with that kind of approach?
|
| 00:41:02 | JH | The main pitfalls are I think that the current limitations on the precision and reliability of the gen AI bots when they’re generating responses to the questions that are coming either from the legal educator or indeed the student. From what I’ve seen I’m not entirely confident that the gen AI technology today will produce nearly enough satisfactory answers to say; “yes, this is a great substitution. Let’s go with this”. I think the algorithms, databases, the particular language learning models within the context of law need to be refined significantly and I’m sure they are being worked on as we speak by various IT engineers in the industry. And maybe 10 years from now we’ll get to that position where we can say; “well, look, it’s an effective substitute for this particular context that now can be relied upon by both students, legal educators, and indeed members of the legal profession for this particular task subject”. Of course, to whatever’s produced ultimately being reviewed and accepted by the particular person who’s going to say “this is my document that I rely on and I submit”.
|
| 00:42:13 | DT | I suppose one of the things you need to keep in mind if you’re thinking about using generative AI for any kind of assessment is what you described earlier that it’s the human student who needs to be the source of knowledge and insight and having that method of assessment where the student is critiquing the AI response or where the student is providing answers imparting their knowledge to a AI client is an example of putting the student and their knowledge at the centre rather than relying upon a large language model to be a source of knowledge or truth.
|
| 00:42:47 | JH | Indeed. So, much has to be said in relation to the responsibility of the individual who is presenting this particular work as their own or that they’re relying on to support the position that they’re taking within the legal profession. It’s so important. And that, of course, is coming from the perspective of somebody who’s a member of the legal profession, and that when you enter the legal profession which is highly regulated you have to obtain a practicing certificate you have ethical duties that are prescribed in all jurisdictions in Australia. You have an obligation not to mislead the court. So, those are very important. And so, when you do present a particular document in the legal profession which you say, for example; “these are my submissions that I have written”, if any of those submissions – and of course there’s no prohibition against using gen AI to assist you in drafting those submissions but there is a responsibility on you to read them check them and make sure that you’re satisfied that there’s a proper legal basis for actually presenting those to a court of law, to a judicial officer, who is going to read them and take them at face value. Because they accept that you’re a member of the legal profession who’s entitled to actually present these kinds of legal submissions to them. You need to make sure that the particular authorities and your submissions the case authorities you’ve cited do in fact exist. If you’ve just shoved into the chatbot on ChatGPT saying; “please write submissions related to this legal issue. Hurry up because I’m in court in half an hour” and then copy and paste them into your submissions, and you see that it’s cited these 10 cases and footnotes. Anybody who’s used ChatGPT will know that all of those case citations – they don’t actually exist. They’re just made up. You need to check them and find your own case citations and if you don’t do that and present something that has obviously been presented in a reckless manner then that impacts on your professional reputation and could even be perceived as unsatisfactory professional conduct.
|
| 00:44:55 | DT | You’re right there is no prohibition on its use in the profession. And I suppose that could be a model for our approach in the educational domain as well. It is ultimately the responsibility of the practitioner that the submissions that they make or the advice that they give all the documents they prepare are accurate, are fit for purpose, reflect a correct view and statement of the law. I do wonder whether we’ll start to see some specific either prohibitions or requirements of disclosure around the use of generative AI. We’ve already seen it in a practice note in a court in Texas – not in Australia just yet – but there are courts in the United States requiring practitioners to disclose whether court documents filed there have been prepared with the assistance of artificial intelligence.
TIP: We’ve actually spoken about this Texas practice note on the podcast before. But just to give you a reminder, Judge Brantley Star of the United States District Court for the Northern District of Texas requires a mandatory certification on the use of generative artificial intelligence for all filed documents. The certification presents two alternatives. The first, a declaration that the practitioner hasn’t used any generative AI to prepare any part of the filing. The second alternative, a certification that any output prepared by generative AI in the filing has been checked for accuracy by a human lawyer.
I do wonder whether the purpose of a rule like that is to discover that fact or merely to cool the intention of practitioners to use them. I wonder if the desire there is to just dissuade practitioners from using AI tools as part of their pleadings or submission drafting process because otherwise they will have to disclose it.
|
| 00:46:37 | JH | I think it’s the latter. That is, it’s a rule that Texas courts have imposed to deter practitioners from doing it at this stage because it seems to have been identified as a recurring problem that is entirely unsatisfactory from the point of view of the judiciary in Texas. And it hasn’t happened here yet because it hasn’t been identified as a recurring problem. Certainly, I’m not aware of any particular instances in Victoria or Australia more broadly where a practitioner has attempted to pass off submissions generated by AI as his or her own. But if it did become a problem then I’m sure that the courts would react and perhaps even the legislature may react as well. And look I say that because I see gen AI as developing more and more. I don’t think it’s going to go away. I think it’s only going to get better. Why it’s not currently a problem and we don’t have any rules relating to its use by legal practitioners in court is because nobody’s really picked up on it because of the points they raised before about the extent to which it can actually be relied upon as being precise and generating thorough expositions relating to principles that are actually applicable to the case that I’m actually dealing with. I’ve tried many times to do that and I’ve just grown impatient. Basically, there’s nothing at this stage that I’m prepared to invest the time and to actually use. But say 10 years from now there may very well be products on the market where you’ve got gen AI software applications that allow a barrister such as myself to sit down and upload particular documents in a court proceeding and then I can highlight certain particular issues which I can then ask the gen AI application to find solutions to referring to all relevant authorities. I can direct them to confine itself to finding principles discussed in only Victoria or perhaps draw upon other overseas jurisdictions and because of the databases that have since been evolved they will be able to draw upon all of the case authorities that of course we see today in the likes of Westlaw and Casebase and spit out the answer and it will be a barrister’s dream document that is presented and they need only read it and they will just fly into court and say; “here are all the authorities on the point, your Honour”. And it will be at that stage where we can actually rely on it to produce something so comprehensive that would ordinarily take days and days to prepare in chambers sitting down researching writing finding that obscure proposition on page 253 of the 1000-page judgment when gen AI can do it in a flash.
|
| 00:49:26 | DT | Yeah, I think that use case that you’ve described – finding the needle in a haystack as it were – you know where to look, but it’s a time-consuming tedious task to find. It is a fantastic use of artificial intelligence. And again, our sister company Ask Lexi does just that with its document review and matter review features, because that’s really about using a large language model for what it’s good for – understanding patterns in language. Often when we’re posing a research question for ChatGPT we’re asking a large language model to do something it’s not very good for which is try to find a relationship between a language pattern and a fact in the real world a relationship between language and reality. It’s not well equipped for that but it’s very well equipped to find patterns and connections between two different language objects and so, giving it that sort of reading comprehension task; “here’s the thousand-page judgment, tell me where in it this issue is dealt with”. It’s eminently well equipped to deal with provided that you’re able to provide those sources of knowledge. Jason, one thing that we’ve been talking about a lot in this interview is the impact of generative AI on our undergraduate or postgraduate teaching at university the way we assess students the way we teach students. But you are the Chief Examiner for the Victorian Bar Entrance Examination. Do you think generative AI will have an impact on bar examinees? My initial impression recalling my time studying for and sitting the bar exam it was a largely unguided process. I knew that I had an exam on a particular date or three exams. I knew the topics that were going to be covered in them. I knew which texts and which legislation I should probably acquaint myself with and I had a few months to do that. There were no classes, there’s no interim assessments, and the assessment is done or at least was done live in an invigilated examination. So, in that situation it seems to me that there’s really two situations in which a potential bar examiner could use generative AI one to assist them with their study of the content and I suppose they’re taking a risk on its reliability if they do that and the other is in pre-preparing answers to possible questions, they could be asked in an invigilated exam. And again, I suppose that’s a limited utility until you get into the room. So, do you see generative AI as having had an impact already or potentially having an impact on the bar examination?
|
| 00:51:52 | JH | I don’t at this stage. And perhaps that’s because of the way the Victorian Bar Exam at least is designed. Although there is potential use for gen AI in the event that a candidate gained access to it while they were actually undertaking the invigilated examination. But certainly, there’s nothing wrong certainly from my perspective in candidates using gen AI to generate model answers from previous exams or hypothetical questions they’ve generated themselves because previous exams are generally accessible in some form or another online. There’s lots of discussion groups online about them and people share previous exam questions and model answers and they’re circulated each year and of course generally speaking the same subject matter is examinable and there are similar questions that are posed in one exam that are similar to previous years exams although they’re not identical. Obviously, we don’t want candidates sitting the exam copying and pasting an answer from a similar question that was presented in a previous exam and shoving it in their answer to this exam. And so, the way we try to minimise that is to each year create a completely novel hypothetical factual scenario and then create very specific questions on particular points and principle of law that relate to quite specific facts in this novel factual scenario which would create a headache really for current gen AI applications to be able to deal with. Because the questions will be slightly reframed in terms of the particular principle and they’ll need to relate the principles to the particular facts that are original now and presented to them for the first time. Certainly, we see a lot of candidates answers that appear to be similar and we can infer that there’s a particular model answer part of which candidates have considered seems to be relevant to answering this particular question this time around but that’s not going to carry them through because there’s always original questions and of course the pass mark in the bar exam is 75 which is actually quite hard to achieve. And so, relying on gen AI is not going to carry you through. The concern I have which I don’t think is a concern right now but is really in the future when technology becomes even more adaptive. The current state of affairs with respect to the bar exam and I don’t think that’s going to change is that when COVID hit in 2020, the bar chose to abandon paper based in the big hall type invigilation where all of the candidates had to turn up to a particular location and there were human beings – members of the bar office – who actually supervised them and they submitted their paper handwritten exam at the end of the exam, and we knew it was their exam. But now when COVID hit that we converted to electronic exam, which essentially permits candidates to sit the exam in the comfort of their own home. And we’ve got particular software that we’ve used to ensure that the computer shuts down and minimises the risk of a particular candidate getting external assistance. You know their camera’s on, it’s recorded, and we review the footage to make sure that they haven’t spoken to anybody. And so, we’re reasonably comfortable that they can’t really use gen AI during that because suspicions will be raised if they divert their eyes away from the camera and look down at a device they’re not supposed to have and it looks like they’re inputting. And they have to undertake they didn’t do anything of that nature. And if there are any suspicions that will call in the candidate and question them about it and may even oblige them to undertake a supplementary oral examination if we’re not satisfied. I can’t say that’s happened to date but certainly the mechanisms we have in place would seem to allay any concerns about candidates sitting the exam in that way are not going to be using gen AI. However, technology, of course, is advancing very quickly. And at some point – and I don’t mean to sound extreme here – but we’ve all watched Mission Impossible from the 1990s, and I don’t think it’s too far away where you can place contact lenses in your eyes and that will transmit the computer screen to somebody else in the next room. And you’ve got a hearing device in your ears which can’t be seen and so, theoretically somebody else could essentially sit the exam for you by looking at all the questions using gen AI in a different room and then conveying to you and your hearing device that nobody can see the answers. That sounds a bit extreme but from what I’ve seen of technology I don’t think that’s far away. Maybe 10 years from now that might be possible and that would then become a concern that would have to address.
|
| 00:56:34 | DT | Well even AI generated deep fakes replacing the face of a person with another’s. That technology is increasingly persuasive. It’s perhaps not so persuasive when done live at the moment but you can imagine that being done in the future to such a sufficiently convincing degree that one invigilator tasked with making sure that 20-30 people are all who they say they are might not pick up on the uncanniness of some of those expressions or movements of the deep fake. So, I don’t think it’s too science fiction to consider the possibility that remote examinations might one day have to grapple with that challenge of us not really being able to trust our own eyes. I think that that is a concern held pretty broadly about generative AI in general terms both diffusion models for image and video and large language models for text that we might no longer be able to trust our eyes that there will be a huge amount of AI generated content out there in the information marketplace and it will be very difficult for us to distinguish which content is written by a human being and can be relied upon and which content’s been written by an AI and can’t and we’ll find it difficult to trust what we read and what we see. I think some of the existential concerns about artificial intelligence are a little bit science fictiony and overblown but that explosion of misinformation or an unreliable content is one that I worry about.
|
| 00:58:00 | JH | Yes, I agree, I don’t think the legal profession is being presented with those extreme situations right now. But I certainly see the potential and that said of course the discussion that we’ve had for this past hour or so, David, I just want to say is premised on the obvious concern that we all have that we as legal educators or members of the legal profession may be being misled by somebody who’s trying to pass off something as their own when it’s not. Of course, it needs to be said that those types of people are rare who choose to do that. So, I don’t want anyone to think that this is a prevailing problem that everyone’s doing it. It’s not the case at all. But we need to recognise that there will always be that sort of inclination to try and get ahead in a way that nobody else might know that I’m actually doing the wrong thing by getting ahead, and some people will succumb to that temptation and do it and we need to put in place the mechanisms to try and stop that from happening.
|
| 00:59:02 | DT | Yeah, and I suppose putting it in a charitable way, this is a new kind of ethical challenge for our profession, isn’t it? It’s not one where the rules have been well established for us and I suppose that there are situations where reasonable minds or otherwise professionals who regard themselves as highly ethical will have differing views on what they consider to be ethical conduct and not ethical conduct with regards to the use of AI.
|
| 00:59:28 | JH | I agree completely on that. Because my own view is there’s nothing wrong as a legal professional or legal educator than using AI to assist you in actually producing something that you can claim as your own. Provided that you make sure that you’ve done all the checks in relation to the propositions. That you are saying; “I adopt as my own”, and that they are accurate and that they are ultimately a product of your care to make sure that what you’re presenting is something that you have substantiated yourself with reference to legal authorities that you’ve found by traditional legal research methods. And I don’t think there’s anything wrong with using AI in that respect. So, it’s subject to that qualification that you need like getting a junior or something to do a preliminary draft and if you’re going to sign off on it and pass it off as your own you need to read every single line and check that it is actually correct and relevant and proper to actually present the context of your case.
|
| 01:00:24 | DT | I suppose if you take attribution for it then you also take responsibility for it.
|
| 01:00:29 | JH | That’s exactly right.
|
| 01:00:30 | DT | Well, we’re nearly out of time, Jason. Before we go you have an opportunity, I suppose, to impart some wisdom to some of our law student listeners. If there’s one thing you wanted our listeners who are law students or who are young lawyers newly entering the profession to take away from our conversation today what would that be?
|
| 01:00:47 | JH | It would be don’t be afraid of using gen AI technology to assist you if you think it would assist you in getting your way through law school in terms of brainstorming ways to deal with a particular problem. But, of course, take great care in what use you actually make of what is produced by any gen AI application that you choose to use because the current state of the technology is that it is likely to produce significant unreliable inaccurate statements which you might think; “oh, that looks quite helpful” but in actual fact, it might not be. And may even alert the person who is assessing your work as to the possibility that you have used gen AI and are attempting to pass it off as their own. And, of course, you have to be conscious of any particular rules that your particular law school has in place that either might allow it or might prohibit the use of it absolutely.
|
| 01:01:45 | DT | I think we’ve said on the show before that what generative tools are very good at doing without some reliable architecture sitting around them is producing plausible responses. And those plausible responses are often correct and they certainly often appear correct but they do need that independent verification. Dr Jason Harkess thank you so much for joining me today on Hearsay.
|
| 01:02:05 | JH | Thank you for having me.
|
| 01:02:17 | RD | As always, you’ve been listening to Hearsay the Legal Podcast. I’d like to thank today’s guest Jason Harkess for being a part of it.
As you well know if you’re an Australian legal practitioner you can claim one continuing professional development point for listening to this episode. Whether an activity entitles you to claim a cpd unit is self-assessed but we suggest this episode entitles you to claim an ethics or practice management and business skills unit. More information on claiming and tracking your points on Hearsay can be found on our website.
Hearsay the Legal Podcast is as always brought to you by Lext Australia a legal innovation company that makes the law easier to access and easier to practice and that includes your CPD. Hearsay is recorded on the lands of the Gadigal people of the Eora nation and we’d like to pay our respects to elders past and present.
Thanks for listening and see you all on the next episode of Hearsay. |
You must be a subscriber to access this content.