Better AI will not solve the hallucination problem. Today’s AI models are less prone to hallucinate than earlier models, and today’s well-known legal providers’ “grounding” of their AI responses has definitely cut down on obvious hallucinations. Don’t make the mistake of thinking this has eliminated the problem; it’s only made it more subtle. The hallucination that should worry a litigator most isn’t the obviously fake case. It’s the real one cited incorrectly: the fabricated quote or the misstated holding attached to a case that actually exists.
This is not a hypothetical. In United States v. Farris, No. 25-5623 (6th Cir. Apr. 3, 2026), the Sixth Circuit sanctioned a court-appointed defense lawyer over two appellate briefs drafted with the help of a “trusted legal technology provider.” The fact that the hallucination was an incorrect quote rather than a fabricated case didn’t impress the court of appeals and was cold comfort to the sanctioned lawyer.
The lawyer had practiced forty years without discipline and admitted the error immediately; the court credited his candor, but denied him all compensation, removed him from the case, and referred him to the Chief Judge and the Kentucky bar.
The answer to this problem is simpler than you think. It isn’t waiting for the next version of an LLM or the newest offering from your cleverly-named legal AI provider. The answer is structural, and it’s a simple structure that you are already familiar with because it’s how you approach legal research with brand new lawyers.
Don’t ask your AI to “tell” you the answers to your legal question; ask it to “show” you the cases (with the relevant parts highlighted). As a former trial lawyer who did my own legal research and supervised scores of younger lawyers doing it for me, this is the approach I advocate.
It’s also the approach we’ve baked into our own AI-based legal research tool: Align Research.
Here’s how it works:
1. No hallucinations because nothing is generated
For legal research, your AI’s ultimate job should be to retrieve a set of relevant cases and present them to you in full, with key portions highlighted. This structurally eliminates the hallucination risk because the AI never writes a word that you see. There is no invented case to verify and no fabricated quote to catch, because nothing was authored. It forecloses both failure modes not by guarding against them, but by removing the step that produces them. Critically, because it’s only ever “showing” you what is there in pre-existing cases (and not “telling” you the answer or what to put in your brief), it can’t invent.
2. It works like a junior associate
There are lawyers who are learning to be AI experts. It’s amazing to see. It’s also not most lawyers. Most attorneys don’t have the time or inclination to learn the nuances of prompt engineering, context management, and agentic workflows. There’s no reason they should have to.
An AI legal research tool can work exactly like asking an associate: You ask a question; it gives you the cases. As litigators know, the need to “find me the cases” comes up all the time:
- Client question. The need to get smart on the relevant case law before a meeting with a new, or old, client.
- Strategy decisions. Could we argue this? Has anyone ever argued this before?
- The other side’s brief. Are they right? Have any courts gone the other way?
- The letter cite. A demand letter, a discovery letter, a settlement proposal. It’s often nice to have a court citation for the point you want to make.
Then of course, there are the times you need to boil the ocean for case law. In that situation you aren’t going to rely on a single tool. Every legal database has gaps. The question, though, is where you start. An AI case-retrieval tool can get you that start, saving you hours of preliminary research and unbillable associate work.
3. It’s billable to the matter
If your clients are still paying you for your initial case searches, you’re lucky. That won’t last. Clients can do the same math you can. An AI-based legal retrieval tool can do for $100 what an associate would bill hundreds or thousands of dollars to do.
That doesn’t mean clients won’t pay for any research. Reading and analyzing are still heartland lawyer work. But finding cases? Nope. Outside counsel guidelines routinely state that in-house teams won’t pay for basic research and require pre-approval to bill it; some bar any research over five hours without sign-off. Bloomberg Law’s 2025 guidance lists prohibiting hourly billing for first-year-associate research — and shifting to flat-fee, per-task billing — as a best practice for general counsel.
In this context, you want a tool that lets you quantify your AI expenses on a per-matter, per-task basis. The value to the client is then obvious. They pay $100 for a case retrieval task. That’s an easy cost to understand, and one that is way less than what the equivalent retrieval would cost in billable hours. It’s also easy to explain when your client asks you “how are you using AI to lower my bills?”
4. The judgment stays yours
None of this hands judgment to a machine. The cases come back; you read them, you decide what they mean. It’s both a discipline (there are no AI summaries to lean on) and an insurance policy (you are always reading the actual text of the opinions).
My own view is that legal research is one of the places where you shouldn’t delegate what you ought to learn. Reading cases is how lawyers develop a feel for a body of law, and AI isn’t good at the hard analytical questions anyway. Let it do the annoying part — sorting the hundreds of opinions you’d otherwise skim — and you keep the interesting part.
In a market racing to make AI do more — draft, summarize, argue — the tool that does less, on purpose, is the one you can bill with a straight face and put your name on in a filing.
5. No babysitting
Align Research isn’t a chatbot you sit with and steer. A research job runs for a few hours — the agent reads thousands of opinions, sorts them, marks the relevant passages — then emails the results. Meanwhile you do other billable work.
I want to be direct about the tradeoff, because it cuts against every demo you’ve seen: you do not get an instant answer. That is the design. Reading thousands of cases properly takes time, and a system that returns something in four seconds is doing something else. Our tool is closer to a permanent researcher on staff than a search box.
Like a good researcher, it goes off and does its work. In the meantime, you do something else. That’s key. Too many times AI has made our brains less productive because we are forced to sit there babysitting the AI while it works – checking it and course-correcting it every five minutes. That’s not the right workflow. AI should work in parallel to you. That’s how a good AI agent is set up.
See how it works at Align Research. Try a case search for free here.
Sam Davidoff is the founder of litigation technology company Align, makers of Align Research. Sam is a former litigation partner at Williams & Connolly, where he practiced for twenty years.
The post How To Avoid Hallucinations: A Legal Research AI That Works Like A Junior Lawyer appeared first on Above the Law.
Better AI will not solve the hallucination problem. Today’s AI models are less prone to hallucinate than earlier models, and today’s well-known legal providers’ “grounding” of their AI responses has definitely cut down on obvious hallucinations. Don’t make the mistake of thinking this has eliminated the problem; it’s only made it more subtle. The hallucination that should worry a litigator most isn’t the obviously fake case. It’s the real one cited incorrectly: the fabricated quote or the misstated holding attached to a case that actually exists.
This is not a hypothetical. In United States v. Farris, No. 25-5623 (6th Cir. Apr. 3, 2026), the Sixth Circuit sanctioned a court-appointed defense lawyer over two appellate briefs drafted with the help of a “trusted legal technology provider.” The fact that the hallucination was an incorrect quote rather than a fabricated case didn’t impress the court of appeals and was cold comfort to the sanctioned lawyer.
The lawyer had practiced forty years without discipline and admitted the error immediately; the court credited his candor, but denied him all compensation, removed him from the case, and referred him to the Chief Judge and the Kentucky bar.
The answer to this problem is simpler than you think. It isn’t waiting for the next version of an LLM or the newest offering from your cleverly-named legal AI provider. The answer is structural, and it’s a simple structure that you are already familiar with because it’s how you approach legal research with brand new lawyers.
Don’t ask your AI to “tell” you the answers to your legal question; ask it to “show” you the cases (with the relevant parts highlighted). As a former trial lawyer who did my own legal research and supervised scores of younger lawyers doing it for me, this is the approach I advocate.
It’s also the approach we’ve baked into our own AI-based legal research tool: Align Research.
Here’s how it works:
1. No hallucinations because nothing is generated
For legal research, your AI’s ultimate job should be to retrieve a set of relevant cases and present them to you in full, with key portions highlighted. This structurally eliminates the hallucination risk because the AI never writes a word that you see. There is no invented case to verify and no fabricated quote to catch, because nothing was authored. It forecloses both failure modes not by guarding against them, but by removing the step that produces them. Critically, because it’s only ever “showing” you what is there in pre-existing cases (and not “telling” you the answer or what to put in your brief), it can’t invent.
2. It works like a junior associate
There are lawyers who are learning to be AI experts. It’s amazing to see. It’s also not most lawyers. Most attorneys don’t have the time or inclination to learn the nuances of prompt engineering, context management, and agentic workflows. There’s no reason they should have to.
An AI legal research tool can work exactly like asking an associate: You ask a question; it gives you the cases. As litigators know, the need to “find me the cases” comes up all the time:
- Client question. The need to get smart on the relevant case law before a meeting with a new, or old, client.
- Strategy decisions. Could we argue this? Has anyone ever argued this before?
- The other side’s brief. Are they right? Have any courts gone the other way?
- The letter cite. A demand letter, a discovery letter, a settlement proposal. It’s often nice to have a court citation for the point you want to make.
Then of course, there are the times you need to boil the ocean for case law. In that situation you aren’t going to rely on a single tool. Every legal database has gaps. The question, though, is where you start. An AI case-retrieval tool can get you that start, saving you hours of preliminary research and unbillable associate work.
3. It’s billable to the matter
If your clients are still paying you for your initial case searches, you’re lucky. That won’t last. Clients can do the same math you can. An AI-based legal retrieval tool can do for $100 what an associate would bill hundreds or thousands of dollars to do.
That doesn’t mean clients won’t pay for any research. Reading and analyzing are still heartland lawyer work. But finding cases? Nope. Outside counsel guidelines routinely state that in-house teams won’t pay for basic research and require pre-approval to bill it; some bar any research over five hours without sign-off. Bloomberg Law’s 2025 guidance lists prohibiting hourly billing for first-year-associate research — and shifting to flat-fee, per-task billing — as a best practice for general counsel.
In this context, you want a tool that lets you quantify your AI expenses on a per-matter, per-task basis. The value to the client is then obvious. They pay $100 for a case retrieval task. That’s an easy cost to understand, and one that is way less than what the equivalent retrieval would cost in billable hours. It’s also easy to explain when your client asks you “how are you using AI to lower my bills?”
4. The judgment stays yours
None of this hands judgment to a machine. The cases come back; you read them, you decide what they mean. It’s both a discipline (there are no AI summaries to lean on) and an insurance policy (you are always reading the actual text of the opinions).
My own view is that legal research is one of the places where you shouldn’t delegate what you ought to learn. Reading cases is how lawyers develop a feel for a body of law, and AI isn’t good at the hard analytical questions anyway. Let it do the annoying part — sorting the hundreds of opinions you’d otherwise skim — and you keep the interesting part.
In a market racing to make AI do more — draft, summarize, argue — the tool that does less, on purpose, is the one you can bill with a straight face and put your name on in a filing.
5. No babysitting
Align Research isn’t a chatbot you sit with and steer. A research job runs for a few hours — the agent reads thousands of opinions, sorts them, marks the relevant passages — then emails the results. Meanwhile you do other billable work.
I want to be direct about the tradeoff, because it cuts against every demo you’ve seen: you do not get an instant answer. That is the design. Reading thousands of cases properly takes time, and a system that returns something in four seconds is doing something else. Our tool is closer to a permanent researcher on staff than a search box.
Like a good researcher, it goes off and does its work. In the meantime, you do something else. That’s key. Too many times AI has made our brains less productive because we are forced to sit there babysitting the AI while it works – checking it and course-correcting it every five minutes. That’s not the right workflow. AI should work in parallel to you. That’s how a good AI agent is set up.
See how it works at Align Research. Try a case search for free here.
Sam Davidoff is the founder of litigation technology company Align, makers of Align Research. Sam is a former litigation partner at Williams & Connolly, where he practiced for twenty years.
