Hallucinated Citations Got a Texas Lawyer Sanctioned — Here's the Workflow NJ Attorneys Should Build Before It Happens to Them

In May 2023, a New York attorney filed a brief in federal court citing six cases that did not exist. ChatGPT had fabricated them — with plausible-sounding names, correct-looking citations, and fictional but confident judicial reasoning. The attorney was sanctioned $5,000. In Texas, a similar incident ended with public disciplinary proceedings. These weren't reckless operators; they were practicing attorneys who trusted their tool without building a verification layer around it.

New Jersey attorneys are not insulated from this. NJ Superior Court, the District of New Jersey, and the Appellate Division all expect citation accuracy — and NJ's RPC 3.3(a)(1) prohibits knowingly making false statements of law to a tribunal. The key word is knowingly, but courts have shown little patience for "the AI told me so" as a defense when no verification steps were taken.

The problem isn't AI itself. The problem is the absence of a deliberate review workflow.

---

Why Hallucinations Happen — and Why Legal Research Is Especially Vulnerable

Large language models generate text by predicting what should come next based on training data. They don't retrieve documents; they reconstruct them from statistical patterns. That means a model trained on legal text can produce a citation that looks exactly right — the right reporter format, the right jurisdiction prefix, the right vintage — because it has absorbed thousands of real citations and is pattern-matching against them.

Legal content is particularly susceptible because it is formulaic. "247 N.J. Super. 312 (App. Div. 1991)" follows a consistent structure that an LLM can replicate effortlessly. Whether that case exists, and whether it actually supports the proposition you're citing it for, are entirely separate questions the model cannot reliably answer.

---

The Four-Step Verification Workflow for NJ Attorneys

This isn't a burdensome process. With practice, it adds roughly 10–20 minutes to a research session and eliminates the category of risk that is producing sanctions nationwide.

Step 1: Never draft citations directly from an LLM output.

Use AI to find legal arguments, identify relevant concepts, or summarize a legal standard — but treat every case name it produces as a lead, not a source. Write down the citation. Do not paste it into your brief.

Step 2: Verify every citation in a primary source database.

Open Westlaw, Fastcase, or Lexis and run the exact citation. Confirm (a) the case exists, (b) the holding matches what the AI described, and (c) the passage you intend to quote appears verbatim in the opinion. This sounds obvious, but the documentation of this step matters — see Step 4.

Step 3: Shepardize or KeyCite before you cite.

A case may exist and say exactly what you claim — and still be overruled, limited, or questioned by subsequent authority. AI tools do not reliably surface subsequent history, particularly in state courts. Running a citator check is non-negotiable for any case going into a filed document, AI-assisted or not.

Step 4: Keep a verification log.

This is the step most attorneys skip, and it may be the most important one if you're ever questioned. A simple spreadsheet — case name, reporter citation, verification date, verifying attorney's initials — creates an audit trail that demonstrates your firm exercised reasonable diligence. If a citation issue arises, a log showing your workflow is the difference between an oversight and a disciplinary matter.

---

Building This Into Your Practice Before a Filing Deadline

Solo attorneys often hit hallucination risk hardest not because they're careless, but because they're time-compressed. The pressure of a filing deadline is exactly when verification steps get skipped.

The fix is structural, not motivational. Build the verification checklist into your matter management system — whether that's Clio, MyCase, or a shared folder — so it surfaces as a required step before any motion or brief leaves the office. Make it as automatic as a conflicts check.

If you use a paralegal or legal assistant to conduct AI-assisted research, RPC 5.1 requires that supervising attorneys review their work product. That review, for AI-generated research specifically, must include citation verification — delegating the AI research task does not delegate the accuracy obligation.

---

The NJ-Specific Stakes

The New Jersey Supremeedulatory framework takes candor to tribunals seriously. An attorney who submits a brief containing a fabricated citation — even inadvertently — risks a RPC 3.3 inquiry, a judicial sanction, and reputational damage in a state bar community that is, particularly at the solo and small-firm level, genuinely small.

The good news: this risk is almost entirely preventable. AI is a powerful research accelerator, and NJ attorneys who use it well will outperform competitors who don't. But "using it well" means treating LLM output as a starting point that requires verification, not a finished product that's ready to file.

Build the workflow now, while it's a process choice — not after a judge questions your citations from the bench.