Every lawyer knows the nightmare version of AI. It doesn’t just get something wrong: it invents cases, quotes, and citations with total confidence, and hands them to you like they’re real.
That isn’t hypothetical. In Mata v. Avianca, a federal judge in the Southern District of New York sanctioned two lawyers and fined their firm $5,000 in 2023 after they filed a brief built on cases ChatGPT had fabricated, then stood behind those fake cases when the court asked them to produce the opinions. The cases didn’t exist. The lawyers hadn’t checked.
That story hit a nerve, because it touched every fear a firm owner has about this technology:
Will I be embarrassed in front of a judge?
Will I increase my malpractice exposure?
Will I get a bar complaint?
Will I make a mistake that lives online forever?
The lesson from Mata isn’t “never use AI.” That’s the easy read, but it’s wrong. The real lesson is narrower and more useful: never use a plain chatbot as if it were a researcher.
The old workflow was seductive: ask ChatGPT a question, get a fluent, confident answer, and let the polish stand in for reliability. It doesn’t.
A hallucination is what happens when the model fills a gap with something plausible. It doesn’t stop and say “I don’t know.” It gives you an answer that reads well enough to survive a quick skim, especially when you’re busy or asking about something you don’t already know cold. That’s the dangerous part: the answer doesn’t sound unsure. It sounds like a confident associate who never checked the source.
I’ve watched lawyers distrust AI for exactly the right reason and then use it the wrong way anyway. They’ll ask a chatbot “what are the best marketing channels for a family law firm?” or “what are my competitors doing in estate planning?” and then treat the reply as research. It isn’t research. It’s a guess in a nice suit.
The shift worth paying attention to isn’t from bad AI to perfect AI. It’s from summarizing chatbots to research agents.
A summarizing chatbot answers from its own memory plus whatever you paste in. That’s genuinely useful for brainstorming, rewriting, and getting unstuck. It’s not enough when the whole point is to know whether something is true.
A research agent does a different job: it searches, reads real sources, compares them, cites them, and hands you a trail you can inspect. That doesn’t make it perfect. It can still misread a source or overstate a point. But it moves the work from “trust me” to “check me,” and for a firm owner that’s the difference that matters. You don’t need an oracle. You need a faster way to reach sources and patterns you can verify yourself.
Here’s the distinction most people miss: there are two completely different research jobs inside a law firm, and the AI tools split along that exact line.
The first job is the actual practice of law. That means case law research, drafting, discovery, contracts, matter analysis, and that’s really the home of the legal vertical AI tools:
Harvey: enterprise and big-firm legal AI for research, drafting, and due diligence.
CoCounsel (Thomson Reuters, formerly Casetext): legal research and drafting wired into Westlaw and Practical Law, with linked citations.
Lexis+ with Protégé: LexisNexis legal AI built around its own case law and Shepard’s citation checks.
Clio Work: Clio’s legal AI workspace for matter analysis, strategy, and drafting, with a built-in citator.
Vincent AI (vLex): legal research, fifty-state surveys, and litigation intelligence, cited to primary sources.
Paxton AI: an all-in-one legal assistant for research, drafting, and document review, priced for small and midsize firms.
There are new tools that could be on this list every day, and the reason they get on this list is because they’re trustworthy enough for casework, and that’s precisely because they’re tied to a real legal authority: those are databases, citators, primary law. If you’re doing actual legal research for a legal matter, that’s where you go. You are still going to need to verify the authority yourself.
But that’s not the research I’m talking about here.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.