Hallucinated Citations Are Now a Sanctions Risk — Here's How NJ Litigators Should Catch Them Before Filing

It started with a brief that looked perfectly fine.

Confident citations. Clean Bluebook formatting. Case names that sounded real — Hendricks v. Morris County Bd. of Chosen Freeholders, State v. Quintero-Diaz — right up until the court clerk went to pull them and found nothing. No docket. No decision. No trace in Westlaw, Lexis, or the New Jersey courts' own public portal.

Scenes like this have played out in New Jersey courtrooms and around the country with increasing frequency since generative AI crept into legal research workflows. The lesson is not that AI legal research tools are useless — they are genuinely powerful. The lesson is that AI hallucination in legal filings is now a foreseeable, sanctionable risk, and "I didn't know the tool could do that" is no longer a credible defense.

---

Why Hallucinations Happen (and Why They Look So Good)

Large language models generate text by predicting what should come next based on patterns in training data. They don't "look up" case law the way Westlaw does — they synthesize plausible-sounding text. That means a model can produce a citation with a real court, a real year, a realistic party name structure, and a pinpoint page cite — and have none of it correspond to an actual decision.

The dangerous part is that hallucinated citations are often thematically coherent. The fake case supports the legal proposition it's cited for. It reads like good authority. A tired associate — or a solo doing everything herself at 11 p.m. — can miss it entirely.

This is not a flaw that will be patched away soon. Even purpose-built legal AI tools backed by curated databases are not immune. Retrieval-augmented generation (RAG) reduces hallucinations but does not eliminate them.

---

The New Jersey Exposure Is Real

New Jersey's RPC 3.3(a)(1) prohibits attorneys from making false statements of fact or law to a tribunal. Submitting a fabricated citation — even unknowingly — implicates this rule. Courts nationally have already sanctioned attorneys under equivalent rules, and New Jersey judges have issued pointed warnings from the bench.

New Jersey's competence standard under RPC 1.1, which the NJSBA has confirmed extends to understanding AI tools an attorney deploys, means you cannot simply offload blame to the software. If you used the tool, you own the output.

---

A Repeatable Hallucination-Check Workflow for NJ Litigators

The goal is not to abandon AI-assisted research. It's to build a verification gate between AI output and your filing. Here's a practical sequence:

1. Never let an AI tool be the first and last stop. Use AI to surface leads — relevant concepts, potential case names, doctrinal frameworks. Then independently verify every citation in a primary legal database (Westlaw, Lexis, Fastcase, or the New Jersey Courts free case search portal for state decisions).

2. Verify the citation exists, then verify it says what you claim. Two separate steps. A case can exist and still not support the proposition for which it's cited — AI tools frequently mischaracterize holdings. Pull the actual slip opinion. Read the relevant passage. Confirm the pinpoint page cite.

3. Run a "ghost citation" check before filing. Before any brief goes out the door, run every case cite through a simple checklist: Does this case appear in at least two independent sources? Is the docket number or neutral citation verifiable? Does the court and date match? If you can't answer yes to all three, the cite gets pulled or replaced.

4. Document your verification. Create a one-page verification log for each filing that lists each case cited and how it was confirmed. This takes under 20 minutes and provides a defensible record if a citation is ever challenged. It also builds the habit across your team.

5. Treat AI drafts as "unverified memos," not finished work product. The framing matters. When attorneys think of AI output as a draft memo from a junior researcher rather than finished research, they are far more likely to scrutinize it. Build this language into your internal AI policy.

---

The Supervisory Dimension

If you have paralegals or associates running AI research on your behalf, your exposure under RPC 5.1 and RPC 5.3 doesn't disappear just because you didn't personally touch the tool. The verification gate needs to be a firm-wide standard, not a personal habit.

Put it in writing. A simple internal AI research policy — even one page — that mandates citation verification before any filing is your best protection against a sanctions motion and your clearest signal to staff that AI output is never self-certifying.

---

The courts are watching. Opposing counsel is watching. And the New Jersey Disciplinary Review Board will not be sympathetic to "the AI told me so." Build the verification habit now, while it's still a best practice — before it becomes the subject of a formal ethics opinion that arrives in your inbox after the fact.