Do Human Lawyers Dream of Electric Briefs? Using AI Wisely in the Practice of Law

As with any powerful tool, the potential for misuse or misunderstanding is significant. An effective use of AI in your legal practice requires an awareness of its limitations and an understanding of how it all actually works.

• Large language models are trained on vast amounts of text, which can be domain-specific or general material, impacting how they respond and encode knowledge.
• A large language model focuses on next token prediction, where tokens represent text and can be character-based or word-based. Tokens can be individual letters or words like “cat” and “dog,” depending on the model design.
• Modern legal AI companies combine large language models with other technologies like search and retrieval, distinguishing them from simpler models that provide one-shot responses.
• The architecture of retrieval-augmented generation over a defined document or data set is beneficial in legal use cases to help  mitigate the risk of using naive models without legal knowledge.
• Companies vary in their approaches to incorporating retrieval-augmented generation, with different methods of processing data and representing it to users, despite using similar architecture.
• Legal technology engineers are exploring ways to improve AI architecture to address these challenges.

• The adoption of legal AI tools is becoming increasingly important for firms looking to improve their productivity and remain competitive in the legal field.
• A common misunderstanding around AI tools can arise when they are used for a purpose they are not intended or designed for. For example, a user mistakenly uses a legal AI tool built for contract drafting to perform research tasks, resulting in the user becoming frustrated and forming the misconception that legal AI tools more broadly are not suitable for research.
• Consider using specialised, purpose-built AI tools and acknowledge your own personal responsibility and duties when utilising AI in the course of legal practice.
• To make informed decisions around AI usage and avoid overrelying on AI tools, it is important to understand the underlying technology of AI tools and how they work.
• Another common misunderstanding stems from the perception that AI tools can do all the work for you, with little user input. Users expecting perfection from an AI tool after only one query might then dismiss it and give up on further exploring its potential benefits. Whereas users who embrace AI tools as an aid and work with it to leverage its capabilities tend to see more successful outputs.
• Traditional legal approaches can also be adapted by implementing a combination of AI tools to streamline different tasks without requiring entirely new workflows.
• Of course, it is always important that lawyers understand when to leverage AI tools for tasks and when to exercise solely their own skills.
• With the advancement of AI, there is a shift towards broader skill utilisation across various sectors. Upskilling in specialised tech tools, development, coding, or marketing can help lawyers both enhance their legal practice and open up opportunities for a multifaceted and flexible career path beyond traditional law firm roles.

Michael Legg, More people are using AI in court, not a lawyer. It could cost you money – and your case, The Conversation, published online September 29, 2025

Benedict Evans, Are better models better?, published online January 22, 2025

The law is sometimes criticised for being slow to change, but it feels like legal practice is evolving rapidly these days. And there’s few rapidly evolving developments, more disruptive or promising – or for that matter, divisive – as artificial intelligence. As generative AI tools integrate into the daily workflows of lawyers to varying degrees of depth and rapidity, they’re transforming how legal professionals research, draft and think about the law, perhaps. These systems can provide up-to-date conversational answers to natural language questions complete with citations and source links, which change how information is accessed and used, but the reliability of that information is sometimes called into question and raises important questions about a lawyer’s responsibility and independence when using these tools.

As with any powerful tool, the potential for misuse or misunderstanding is significant. An effective use of AI in your legal practice requires an awareness of its limitations and an understanding of how it all actually works.

To unpack these challenges and opportunities, we’re joined today by Will McCartney, founder of Habeas AI, an Australian company dedicated to revolutionising the way Australian lawyers engage with legal research. Habeas provides easily readable responses backed by verifiable legal sources. I’ve known Will for a while in the legal tech community. Love Habeas, love what you’re doing.

Will, thanks so much for joining us on Hearsay.

But at the same time, I will say that definitely throughout my uni years, particularly 2021, 2022, I was getting really into technology at the same time. So I was building projects on the side. I was studying projects, not really just in legal initially, but in other spaces that were deeply fascinating to me. And I liked the fact that with tech you have this enablement, this sense of autonomy that you might not always get in the structure of say, a university degree and what you are studying there, which can be a bit more rigid or constrained to the parameters of what you are working on, right? So I think that’s a bit of my background in the sense of, yes, I’ve got the humanities interest, but increasingly got interested in tech and capacities we could do there. And I think eventually where that led me to was the fact that whilst in law school I saw there was quite a lot of problems that not only students, but also professors at law school that you’re talking with, face around the retrieval of legal information and search engines that we were using at law school even three to four years ago. Of course, the dominant ones that you might be exposed to in law school, like AustLII or bigger providers like Lexus or [Thomson Reuters] that the university has a subscription to, right? And there’d be these common phrases from teachers being like, “oh, you this is good to get you started,” or, “you know, this is how you kind do a keyword or a Boolean search,” et cetera.

TIP: The episode that David just mentioned that gives a realistic look at AI in legal practice is our episode with Dr Armin Alimardani. That one is episode 131 and is called ‘The Lawyer’s Guide to Generative AI: Where It Fits (and Doesn’t) in Legal Practice’. Check that one out if you get a chance. It includes a quick crash course on how generative AI actually works.

And I don’t know if you would agree with this, but at least anecdotally, I feel like there’s less and less magical thinking about AI in 2025. I think in 2023 there was a lot of misconception about how AI worked and a lot of inflated expectations of what it might be capable of. I think people are starting to understand that LLMs are next token predictors. They generate plausible responses, but not necessarily correct ones. But there’s still some of that misconception, right? And I guess there’s still a lot of people who haven’t used them regularly or at all. So, just as a bit of a refresher, the elevator summary when we’re talking about AI in this episode, we’re really talking about large language models and derivatives thereof like, embedding models and things like that. What are we describing when we’re describing an LLM?

The other thing this brings me to is – I think you mentioned this point – around LMS as well and potentially that in certain instances they’re good when they already have the retrieved knowledge at their fingertips. That’s not to say that there’s not a realm of questions in the law and tasks that people are doing that sort of range from subjective to objective. So there’s certain tasks where hundred percent factual accuracy is required to validate a source or to understand where information came from. There is a right answer.

But there’s also a lot of tasks, new areas, fundamentally constructing a new legal argument where creativity, subjectivity, these sorts of factors come into play. And I still think that’s an awesome thing about large language models. They can be a brilliant creative aid in that capacity. They’re still obviously rooted in a model of the world a certain way of, I wouldn’t necessarily describe it as thinking, but reasoning, that we might call it colloquially. And so we can remember that. And then for those situations where that objectivity is required, we can build stacks so that an LLM is always being able to reference real data in a more context informed way, but we should also embrace those aspects where the creative and the ideation capabilities of LLMs come out and are really strong as a tool in the lawyers toolbox.

One thing I do like about Habeas is that you do provide that subparagraph, right? Because what I wanna ask you about next is verifiability. I think you put it exactly right, which is that retrieval, and retrieval that’s specifically adapted for the type of work you’re doing, so, specific to law, substantially mitigates hallucinations, right? But they kind of can’t be solved in the sense that Meta’s chief AI scientist says it’s a fundamental feature of large language models. It will just happen no matter how large they get, no matter how sophisticated. And so there is always a need to verify, right? And there are some lawyers who complain, “well, the verification takes as long as it would take to do the task. So why bother? Like, I should just do the task myself.” And we do see in other domains, the Metr study – that’s METR, not META – in software engineering that found that some software engineers thought they were saving 20% of their time because it took them less time to write the code, that it takes them 20% more time to review the code. So tell me a bit about verification and how you solve for what I’ve heard called “verification fatigue,” which is that challenge of verification being a really laborious task that can perhaps rival the original task in terms of its time, because I know Habeas has some useful features there.

And even there is some interesting research which we will include in the show notes, prompts sometimes behave in the opposite way to the way we think. So you can’t just tell ChatGPT to behave like a PhD level stock analyst and expect it to do it. There was some interesting research that if you told chat GPT to avoid positional bias. which is that if given the options one and two or A and B, it might tend to prefer the first option or the second option. If you tell it to avoid positional bias, it’s more positionally biased than if you didn’t.

TIP: So David just mentioned a study on trying to negate positional bias in LLMs. That study is called ‘Can We Instruct LLMs to Compensate for Position Bias?’ and it was by Zhang, Meng and Collier in 2024. We’ll include a link to that one in the show notes if you’d like to take a further read.

One of the most interesting takeaways from that study is that prompting large language models doesn’t always work the way our intuition suggests. You might think that telling a model to “avoid bias” or “be more objective” would improve its behaviour, but in reality, sometimes it can actually do the opposite.

So the study is looking at something called position bias, which is the tendency for language models to prefer information that appears at the beginning or end of a prompt, while underweighting content in the middle.

The researchers tested whether you could “talk the model out of” this behaviour using instructions. They tried two approaches. The first was what you might naturally attempt: relative positioning. For example, telling the model to focus on “the middle section” or “the later part” of the input. That didn’t work. Across multiple models, including closed-source ones, the models showed little understanding of relative position concepts. In other words, “the middle” doesn’t mean much to a system that doesn’t actually perceive the prompt spatially the way humans do.

The second approach was more concrete: explicitly labelling documents with numeric IDs and telling the model to focus on “Document 2” or “Document 5.” That did work. When the correct document was given a clear identifier and referenced directly in the instruction, the model shifted attention and performance then improved. When the instruction pointed at the wrong document, performance dropped sharply, which is actually good evidence that the model was genuinely following the instruction rather than guessing.

The lesson is that vague behavioural instructions like “avoid bias,” “be objective,” or “act like an expert” often don’t map cleanly onto how models operate internally. In some cases, telling a model to avoid positional bias actually made it more biased, because the instruction itself introduced confusion. So rather than abstract, human-style advice, what does seem to work more is being explicit and giving the model clear anchors, so numbered references, explicit constraints, or concrete targets. 

The broader takeaway is that prompting is about control surfaces. If you want reliable behaviour, you need to point, label, and constrain as a user. That’s a very different mental model from how most people approach AI today.

All of these things tell me that prompt engineering is a really important skill for people like yourself who are testing prompts at scale and designing them as part of a cohesive framework, and who can observe how a prompt works when you’ve run it a hundred times, a thousand times in many different scenarios, but not so obviously useful to the lawyer who maybe prompts once a week.

So I have already seen a lessening of that to a certain extent over the past year. But nevertheless, I’ll even get emails like, “oh, I created my particular agent with this persona to do X, et cetera.” But maybe it had a slightly different outcome on the way it actually went about this research. There’s always an incredible complexity because what people ultimately want is they almost want this J.A.R.V.I.S. human level thing that they can just talk to, and that is the perceived standard in a lot of lawyer’s minds. And then occasionally, I do think sometimes AI hits the mark there and it’s incredible, right? But sometimes it falls short. And that’s important there’s something that maybe goes wrong in the backend or the way the assistant interprets that prompt as well. But yeah, I think in particular, just even to provide more context there, our agents feature, which we call assistants, that’s how we frame it to people, it is an agentic architecture on the backend. When we say agents, it’s just really this concept of rather than exposing it to a one shot answer, the model can actually go and reason and call upon different tools as part of its response. And it can take that initial query, maybe break it down into three to four subqueries, right? Because maybe that initial query covers so much ground and it can call all those tools in tandem, get all that context and then, take it into the inference step into the step where it actually generates an answer as well. So it becomes a much more potentially informed step. However, the interesting thing about that is you are also delegating some decision making to a model, just as you would if you asked a paralegal. Five different people would probably have slightly different methodology and way that they would submit those queries to search engines, right? So this is this interesting step as well is there’s some element of almost choice or autonomy that we’re now giving models, right, to create better answers. So handing off some of that, but then also realising that you can still point things in a very specific direction and that will lead to better search theories or a stronger understanding for the model of what you want and need.

If you’re an Australian legal practitioner, you can claim one continuing professional development point for listening to this episode. Now, as you know, whether an activity entitles you to claim a CPD unit is self-assessed but we suggest that this episode entitles you to claim one professional skills point or a professional skills point – take your pick. For more information on claiming and tracking your points, head on over to the Hearsay website.

Hearsay the Legal Podcast is 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.

Now, before you go, we’d like to ask you a favour. If you’re enjoying Hearsay the Legal Podcast, please consider leaving us a Google review. It helps other listeners to find us and ultimately that keeps us in business.

Thanks for listening and we’ll see you on the next episode of Hearsay.