Hallucination Detection in NJ Legal Research: A Practical Workflow for Solo Attorneys

An AI research tool gives you a perfect-looking case citation. Proper party names, a plausible docket number, a holding that fits your argument like it was written for you. You paste it into a brief. Then opposing counsel runs it in Westlaw and finds nothing.

This isn't a hypothetical. Since 2023, federal and state courts across the country have sanctioned attorneys for citing AI-fabricated cases, and New Jersey practitioners are not insulated from the same risk. The New Jersey Rules of Professional Conduct don't carve out an exception for honest mistakes made by a chatbot. Under RPC 1.1, competence means keeping up with the benefits and the risks of relevant technology. If you use an AI tool for legal research and submit its output without verification, that's on you.

What makes this harder than it sounds is that hallucinations don't look like errors. They look like real legal authority. That's the core problem, and it's why a vague mental note to "double-check your AI" isn't a workflow.

Here's one that actually is.

Step 1: Treat Every AI-Generated Citation as Unverified Until Proven Otherwise

This sounds obvious, but the practical failure point is usually speed. You're drafting a motion, the AI drops in six citations, five of them check out quickly, and you assume the sixth is fine. Build a rule: no citation goes into a filed document until it has been confirmed in an authoritative legal database, full stop. Not skimmed. Confirmed. That means pulling the actual case, reading the relevant passage, and verifying the holding matches what the AI claimed.

For NJ practice specifically, that means checking New Jersey courts' official databases, Westlaw, Lexis, or Fastcase for any case the AI mentions. The New Jersey Judiciary's public case information system can confirm whether a docket exists at all, which takes about 90 seconds and catches pure fabrications immediately.

Step 2: Verify the Quoted Language, Not Just the Citation

AI tools frequently cite real cases but misquote or paraphrase the holding in a way that subtly shifts the legal standard. A case might exist and say something close to what the AI described, but "close" in a brief is a problem. When you pull the case, find the specific language the AI attributed to it. If the AI gave you a block quote, search for that exact phrase in the opinion. If it's not there, you have a partial hallucination, and you need to decide whether the actual holding still supports your argument.

This step catches a category of error that pure citation-checking misses entirely.

Step 3: Run a Shepard's or KeyCite Check Before You Brief It

A case can be real, accurately quoted, and still be bad law. NJ attorneys already know to Shepardize, but when AI accelerates your research volume, this step sometimes gets compressed. If an AI tool surfaces a case you wouldn't have found yourself through traditional research, treat it with extra skepticism: check its subsequent history before you rely on it. A 2019 Appellate Division opinion the AI surfaced might have been reversed or distinguished in ways that undercut your argument entirely.

Step 4: Log What the AI Told You

This step takes 60 seconds and protects you. When you use an AI tool for research on a matter, keep a brief notation in your file: what tool you used, what query you ran, and what citations it generated. If a citation later turns out to be fabricated and the issue comes up in a disciplinary context, your ability to show a systematic verification process matters. It demonstrates the kind of supervisory judgment RPC 1.1 and RPC 5.3 expect when non-lawyer tools touch client work.

Step 5: Know Which AI Tools Have Better Track Records for Legal Research

Not all AI tools carry equal hallucination risk for legal citations. General-purpose large language models (ChatGPT, Claude, Gemini) are not designed for citation accuracy and carry high fabrication risk for specific case names and dockets. Purpose-built legal research tools like Westlaw AI, Lexis+ AI, and Casetext's CoCounsel are grounded against actual legal databases, which substantially reduces (but does not eliminate) citation hallucination. That distinction matters for how heavily you verify.

If your firm is using a general-purpose AI tool for any part of legal research, the verification burden is higher, and your workflow needs to reflect that explicitly.

The Practical Takeaway

The verification workflow above adds roughly 10 to 20 minutes per research session. That's not a reason to skip AI research tools; the efficiency gains are still real. But that time is non-negotiable if you're filing anything in a New Jersey court. RPC 1.1 has always required that you understand the tools you use. What's changed is that the tool can now produce something that looks exactly like competent legal research while being entirely made up. The burden of catching that falls on you, before the brief is filed, not after opposing counsel does it for you.

If you don't have a written verification checklist for AI-generated research in your practice yet, building one this week is the most concrete single step you can take toward competent AI use under New Jersey's ethics rules.