Lawyers’ AI use must be verifiable, not vulnerable

“One of the most important things to do is to just really evaluate the tool and make sure that when you are choosing a tool that you are testing it on the kind of work you do,” McLay said.

That evaluation should not be treated as a one-off exercise, she added, arguing that firms need a methodology for testing AI and should make verification part of their organisational culture. This must include, she said, ensuring that outputs are being checked and verified accordingly.

Lawyers should also continue testing tools after they have been adopted, she said, given the pace at which AI systems are developing.

McLay cautioned against assuming that a general-purpose AI system is adequate simply because a lawyer can identify whether an answer “sounds good”. Users of general-purpose AI face additional burdens, including checking the currency and context of information, refining prompts and undertaking more extensive verification, she said.

“It feels like it’s a fast answer, but actually, it’s increasing an ongoing burden for them,” she said.

Jarvis similarly warned lawyers against confusing the sophistication or fluency of an AI-generated response with its reliability. As AI models become more advanced, he said, there is a growing risk that authoritative-sounding answers would be accepted without sufficient scrutiny.

VIEW ALL

The starting point, he said, should be the legal problem itself: lawyers need to ask what they are trying to solve and whether the technology has been designed for that particular problem.

“When we develop solutions like Westlaw, we focus in Australia on the Australian legal content,” Jarvis said, noting that legal technology developed for another jurisdiction would instead be grounded in that jurisdiction’s content.

That distinction is critical because Australian law can be complex and nuanced, Jarvis said. Lawyers should establish whether an AI tool has access to the appropriate information, whether it is designed to address the relevant legal question and whether its output can be traced and understood.

“If it’s not transparent in showing you the response, you need to be concerned [about] how authentic and verifiable that output is,” he said.

McLay also warned that the distinction between general-purpose AI and profession-grade legal technology was becoming increasingly important as client expectations changed.

Practitioners are finding that clients and in-house lawyers are themselves using AI to interrogate legal advice, creating expectations for faster, deeper and more comprehensive work, she said, and that changing benchmarks means firms cannot simply adopt whichever AI tool is most readily available.

“If you start building those around a general-purpose tool that isn’t really fit for purpose, you’re wasting even more time potentially,” she said.

Ultimately, Jarvis said, responsibility for legal advice must remain with the human lawyer, rather than being displaced onto the technology.

“The humans [are] accountable for the legal advice,” he said.

With AI increasingly capable of producing convincing answers, the central safeguard is therefore not whether an output sounds right, but whether a lawyer can establish why it is right.

“Australian law is complex and nuanced,” he said. “If the grounding is missing, what sounds good might just be wrong. If you can’t trace it, you shouldn’t trust it.”

To watch a playback of Thomson Reuters’ session, as well as the other sessions at Tech Innovate, click here.