Legal AI Adoption In Australia: Habeas' Will McCartney on Mitigating Hallucination Risk

Let's talk a bit broad compass, 'cause you're in this space.

This is your day-to-day work, and I look on as an observer and a, and at times a commentator.

But talk us through the trust issues and the hallucination.

This is what I understand as a non-tech guy.

Hallucinations happen because we tend to think it's a person at the other end communicating with it, but it's, um, it's a predictive model.

And so they're really predicting what you should know based on the dataset they're drawing from.

So...

And of course, an AI model doesn't know, uh, won't be able to say, "I don't know," so it-- because of its predictive nature, it's made stuff up at times.

Am I on the money? Is that, is that what's played out, or is it more complex than that, no doubt?

Broadly, you're on the money in the sense that the-- it would be the right framing to say that oftentimes generative AI solutions, just raw large language models, um, that companies are producing, um, always have a risk of hallucination, and it's something that companies have already, like frontier labs that are building these models, um, or have already done a good job of reducing over the years.

Just even in the, in the span of the past three years, the hallucination rate, et cetera, has lowered pretty dramatically.

But yes, it's something intrinsic more to the nature that... of the way that the text works, and so, uh, of the, the way the tech works.

And so one of the, I guess, confusions in this space is a, is a tendency to just kind of conflate the term AI, and people just think of everything in this one broad brush category, when, of course, you have so many different iterations there under the hood.

So you have, you know, machine learning pipelines.

You have generative AI systems.

You have, um, custom platforms that leverage external AI models, but they do so in a way where they've created custom code and architecture around that.

You have platforms like ours that are more premised on the idea of being search first, inference second, which is prioritizing an actual real-world retrieval search or analysis step.

Where we go out, look at real data and information, and that's always cited to the user, and then you kind of get a, a, an analysis or inference layer on top.

Then you've also kind of got the emergence of a lot of people who have engaged in fine-tuning models for particular do- domains, which can have its pros and cons, but that's using maybe a particular data set to focus a model.

And then, um, there's also the rise of other types of models that are of interest for forms of legal work, like encoder models.

And without getting too deep into that, there's basically other types of AI models that we can look at that don't hallucinate, but which also kind of have their pros and cons in real-world legal workflows, right? You have this, you know, core issue, and then you've got a whole range of different providers, apps that have, um, different ways of trying to either solve that problem or build architectures around it.

So is the trust issue, do you think, well-founded historically, but from what I'm hearing from you, it's im- it's, it's significantly improving? I mean, we've all read, uh, cases around AI producing submissions citing cases that don't exist, and some practitioners finding themselves in very hot water about that.

In terms of, say, the Habeas approach, because it's very much research first and draws on existing data, uh, and established case law and, and, and precedent, that it sort of negates some of those issues.

But what I'm hearing more broadly is you think that r- rightly there's been these trust issues, but over time, it, it, it's will significantly improve.

Is that a fair summary?