How MiAI Law built an LLM for real legal reasoning

For the past 15 years, Laina Chan, an Australian barrister and authority on construction, property and insurance law, has published widely across the legal landscape.

Like many others in the legal sector, Chan experimented with generative AI tools to mine and gain insights from her own data, but these solutions didn’t deliver.

“I’ve read so many cases, but I don’t have instant recall. I started trying other [vendor’s] AI products, but they were basically just ‘wraps around OpenAI’ that weren’t giving me the precision I wanted. Most importantly, I just wanted to be able to tap my own IP,” Chan tells ADAPT.

Earlier this year, Chan teamed up with India-based AI specialist, Arjun Jain, and others to build and take to market what is believed to be a world-first legal reasoning platform, MiAI Law.

This large language model (LLM) encodes legal reasoning, which Chan says is missing in the general models used in the sector. She trialled LexisNexis and other legal research products that only retrieve and provide the proposition of a legal case.

“But I wanted to know if that’s the most authoritative case on the subject…we built it so that our model can actually do that. It will retrieve the most authoritative cases; it will check whether it’s still good law and it will give you a reasoned report with everything – pinpoint references and hyperlinks to source materials.”

Without these references, Chan says, a solution is essentially a ‘black box’ forcing users to search underlying case material to verify propositions. This is not the case with this new model, which Chan describes as “a little robot lawyer” operating at a PhD level.

Pure, compliant and encrypted data

MiAI Law, which some lawyers have been using since September, analyses primary legal cases and legislation using its own algorithms. At no point does it rely on information that exists online.

“We are as pure as you can get in terms of our ‘data in’ because every single case analysis is mapped to an original case and we never, ever look at what the internet says. We don’t look at what an author might say a case means, we only ever rely upon the case.”

Chan completed an honours degree in pure mathematics in the early 1990s, only recently realising that her final thesis on algebraic equations, the very calculations that underpin encryption systems, would be so useful when building this tool.

“We make sure that our data is encrypted both at rest and in transit and at the moment, we’re working on two-factor authentication for all users.”

The solution is hosted at Microsoft Azure’s Australian data centres “so we know we are complying with the law”, adds Chan.

“Our backend is completely sorted; everything is governed and packaged. We were really conscious from day one that we knew who our audience was and to make sure our systems would comply so we [wouldn’t be] trying to plug holes.”

Hallucinations showed tools were not fit-for-purpose

Legal professionals in Australia have come under fire for misusing AI in cases. So far, there have been 43 legal cases where generative AI produced hallucinated content.

In August, a lawyer was referred to Western Australia’s legal regulator after preparing court documents that contained AI-generated citations for non-existent cases.

Two Victorian barristers and one South Australian solicitor were recently referred to regulators after AI was used to prepare a court document containing ‘erroneous citations.’

A judgement handed down by the Federal Circuit and Family Court of Australia found that the document, filed as part of a court appeal, was drafted using AI and contained “inaccurate and misleading references to case law.”

In this case, the paralegal had used AI.

“I think it’s outrageous that the entire case theory was built upon the research of a paralegal that had not been checked in any way up the chain. That paralegal obviously wasn’t trained properly on how to use AI…people don’t seem to realise that you can’t take the AI as gospel truth.”

Language models used in these cases are not ‘fit for purpose’, says Chan.

“A language model is just one element in our input; you can’t use it as the ‘be all the end all’ – to do all the legal research for you.”

Click on the video above to watch our full interview with Laina Chan.