The Truth About AI Hallucinations in NJ Legal Briefs: Most Firms Have No Workflow to Catch Them

Everyone in the legal profession has heard the cautionary tales by now. A lawyer submits a brief citing cases that don't exist. A judge's law clerk can't locate a single one of them. Sanctions follow. But those stories tend to get treated as embarrassing outliers, the attorney who trusted the chatbot too much and skipped any review. The uncomfortable reality, especially for solo and small NJ firms running lean on time and staff, is that hallucinations are a structural problem with how large language models work, not a user error you can proofread your way past with a quick skim.

This post is about building an actual workflow to catch them before they become your problem.

What Hallucination Looks Like in Practice

A hallucination isn't always a completely fabricated case name. That's the obvious version, and honestly, it's the easier one to spot. The subtler version is a real case that exists, decided by a real court, but with a holding that the AI got meaningfully wrong, sometimes inverted entirely. Or a statute cited with the correct title and section number, but from a version that was amended two legislative sessions ago. Or a quote attributed to a New Jersey Appellate Division opinion that appears nowhere in the actual text.

These don't fail a surface-level "does this look real" check. They look exactly like correct legal citations, because the formatting is right and the case actually exists. That's what makes them dangerous.

NJ attorneys have particular reason to be alert here. New Jersey's court system produces a high volume of unpublished opinions, and NJ Practice Series treatise materials are frequently paraphrased by AI tools in ways that drift from the source. If you're using a general-purpose model rather than a legal-specific one, the training data on New Jersey-specific procedural rules, local court practices, and NJLRC guidance is often thinner than federal content. The model fills gaps with plausible-sounding synthesis that has no clean citation behind it.

Why Standard Proofreading Fails Here

Most attorneys review AI-drafted work the same way they'd review a first-year associate's draft: read for flow, check the legal reasoning, fix the writing. That workflow catches a lot. It doesn't catch citation hallucinations reliably, because you're reading for sense and coherence, not independently verifying every source.

The cognitive trap is that AI-generated text is unusually fluent. A cite-checked brief from a competent associate might have a few awkward passages that signal "this needs another look." AI output often reads clean, which lowers your guard at exactly the moment it shouldn't.

RPC 3.3 requires candor to the tribunal, and under NJ precedent, the obligation to verify what goes into a court filing rests on the signing attorney, full stop. The AI vendor isn't a co-counsel of record. Competence under RPC 1.1, which the NJ Supreme Court has consistently read to include technological competence, means understanding not just how to use an AI tool but understanding its failure modes.

A Practical Verification Workflow for NJ Small Firms

You don't need a dedicated research associate to do this. You need a structured habit that runs every time AI contributes substantive legal content to a document that goes to a client, opposing counsel, or a court.

Step 1: Extract every citation before you read the brief for substance. Pull all case names, docket numbers, statute sections, and regulatory references into a separate list. Do this before you read the argument. If you read first, confirmation bias kicks in and your brain validates the citations because the argument made sense.

Step 2: Run each case citation against Westlaw, Lexis, or Fastcase, not Google. Google will find a case that almost matches. You need the official reporter or the court's own database. For NJ opinions, the Judiciary's public portal and the Appellate Division slip opinion archive are free and reliable for checking whether a citation resolves at all.

Step 3: Read the actual holding, not the headnote. Verify that what the AI says the court held is what the court actually held. This takes ninety seconds per case if the opinion is short. It takes longer for complex holdings, but it's not optional.

Step 4: For statutes, check the current codified version. NJSA online through the Legislature's site shows the current text. If you're citing a regulation, check the NJ Register for any amendments after the version the AI may have been trained on.

Step 5: Flag any citation you can't fully verify in under five minutes. If you can't locate it or verify the holding quickly, treat it as a hallucination until proven otherwise. Replace it, drop it, or do a proper manual research run. Don't include it on the assumption it's probably fine.

That process adds real time to an AI-assisted drafting workflow. For a brief with a dozen citations, budget twenty to thirty minutes for dedicated verification. That's still considerably less time than you'd spend responding to a sanctions motion, or explaining to a client why the court issued an order to show cause over a brief you filed on their behalf.

The One Practice Change That Matters Most

If you take nothing else from this: build the citation extraction step into your AI drafting process as a non-negotiable gate, not an afterthought. The attorneys who have run into serious trouble with AI hallucinations almost universally reviewed the AI output as a whole document. The ones who catch problems treat the citation list as a separate artifact requiring independent verification.

Add a line to whatever AI policy your firm uses (or build one if you haven't) that says explicitly: no AI-assisted document containing case law or statutory citations goes out the door without citation-level verification against a primary source. That one sentence, enforced consistently, eliminates most of the disciplinary exposure that comes with AI drafting.