Back when navigation apps first emerged, skeptical drivers still kept a paper road atlas in the glove box just in case. It didn’t take long, though, before the driving masses stopped questioning and began to follow along. The apps had earned our trust.
Greg Dickason, chief technology officer at LexisNexis Legal & Professional, a legal technology provider, worries something similar may be happening with artificial intelligence (AI), and that it could be to the detriment of professionals in highly consequential fields like law. Despite ongoing wariness of hallucinations and bad information, attorneys have still seen AI get better at researching issues and drafting briefs.
“We're now at the point where you can trust it more, but you can't trust it 100%,” he says. “And that's the big gap. Partially correct is probably more dangerous than wrong.”
The question for legal organisations is how they can harness the massive opportunity of AI without risking sanctions, damaging their reputations, compromising client relationships, or, perhaps worst of all, losing trust.
Why legal AI needs a higher standard
Legal professionals in particular cannot afford to allow inaccuracies, faulty citations, or flawed analyses into their work. It doesn’t help that AI models are generally built for fluency, meaning that even when they’re wrong, they sound confident. “Every day, you read a new report in a legal periodical where an attorney is either in trouble with a court or has an ethical issue as a result of improper use of AI or using an untrustworthy system,” says John Ursin, managing partner at Schenck Price. “We simply cannot take any risks with our attorneys’ professional reputation and our client outcomes.”
Getting AI right in the legal world starts with grounding a model’s responses in reliable, up-to-date sources, and ensuring that users can trace output back to them.
“Grounding the data within our generative AI-based tools, or really any of our tools, is of the utmost importance to the firm,” says Todd Friedlich, director of knowledge management and innovation at New York-based Willkie Farr & Gallagher. Firms need attorneys working from current law, internal documents and other materials it has designated as authoritative—not unverified information pulled from unknown sources.
Building AI into legal work responsibly
Authoritative information is just one part of the equation. Legal organisations also need to think about how AI fits into their larger workflows, and where human judgment enters the process.
Mr Dickason compares the relationship to the way a partner supervises an associate. The associate might do much of the underlying research and drafting, but the partner still needs to review the work before putting their name behind it. And, critically, that review should not start from scratch. The associate needs to show their reasoning, provide supporting evidence and flag any areas where they’re uncertain. “That checking has to be robust,” says Mr Dickason, “but it also has to be easy”.
As legal organisations get inundated with pitches from AI products and services, Mr Dickason encourages them to look for essential capabilities. They should ask whether products can easily trace underlying sources and keep humans in the loop to approve workflow steps and output. They should also ensure the AI can handle complexity across, for instance, multiple jurisdictions. “It might be a Delaware contract for an employee that’s based in California, but the company headquarters are in Texas,” Mr Dickason says. “How are those kinds of failure modes handled by your AI?”
The litigation teams at Wilkie are incredibly excited about the use of workflows to automate the analysis of depositions, to find opportunities during cross-examination.”
Todd Friedlich, director of knowledge management and innovation at Willkie Farr
Simply granting AI access to a vast library of cases and legislations is not enough. The information must be organised and contextualised so the system can retrieve the right material. The LexisNexis process begins by keeping its constantly growing collection of sources—which includes archived statutes, unpublished decisions and manually retrieved offline sources—up to date. To do this, its editorial teams modify and correct about 550,000 cases, statutes and pieces of legislation each year. Its agentic AI assistant, Protégé™, then draws on this editorially enriched content alongside other legal context to help its AI tools identify the appropriate authority.
The company’s Shepard’s Citation Service adds a further layer of scrutiny by helping users validate the legal authority and historical treatment of a wide range of sources. This includes cases, statutes, court rules, legislation and administrative materials. Shepard’s analysis spans more than 60 million points of law and is continually updated. Its graphical analysis, meanwhile, maps citing trends across jurisdictions while its case law analyses undergo rigorous attorney editor review. For support interpreting primary law, LexisNexis offers users extensive secondary content, from news sources and commentary to practical guidance, drawing on more than 5,500 expert contributors.
Still, legal organisations can and should reinforce those safeguards through their own workflows. Willkie Farr & Gallagher, for instance, establishes approved sources, prompts and output formats for specific tasks. “So, the attorney can spend their time adding the real value,” says Mr Friedlich.
Foundations of trustworthy legal AI
Grounded answers
Traceable sources
Verified sources
Uncertainties flagged
Humans in the loop
Trust as an enabler of adoption
Trust can’t be treated as a barrier to AI adoption in the legal field. It’s a precondition, and to get there, organisations must deeply understand how the agentic tools they deploy actually work.
Even after customising the model on the right context and priority materials, they must resist the urge to blindly trust outputs simply because they’ve become reliable most of the time. With the consequences of failures so pronounced, firms must layer in reliable checks from human attorneys.
After all, the goal is not to remove lawyers from the process. It is to build systems that make clear when their attention is needed most.
