What Hallucinated Case Citations Taught One NJ Federal Court About Attorney Candor Under RPC 3.3

A federal magistrate judge in the District of New Jersey doesn't need to see a second hallucinated citation to recognize a pattern. After several high-profile incidents nationally, including the now-infamous Mata v. Avianca sanctions order from the Southern District of New York, NJ practitioners have started asking a quieter but more urgent question: what exactly does RPC 3.3 require of me when an AI tool drafts citations I didn't personally verify?

The answer isn't complicated, but the failure modes are.

What RPC 3.3 Actually Says (and What Most Attorneys Miss)

RPC 3.3(a)(1) prohibits a lawyer from making false statements of fact or law to a tribunal. RPC 3.3(a)(3) goes further: a lawyer must not offer evidence the lawyer knows to be false. These aren't novelties created for the AI era. They're foundational duties that predate word processors.

The AI wrinkle is this: the attorney who pastes a GPT-generated case citation into a brief without verifying it probably doesn't know the citation is false at the moment of filing. But "I didn't check" has never been a successful defense to candor violations. The New Jersey Supreme Court's approach to RPC 3.3, tracked through its Committee on Attorney Advertising opinions and disciplinary decisions, consistently treats negligent misrepresentation to a court as seriously as intentional misrepresentation when the attorney had clear means to verify.

In practical terms: you can't outsource your verification obligation to the tool that generated the error.

The Federal District of NJ Is Watching

The D.N.J. hasn't yet issued a standing order on AI-generated filings as sweeping as those from the Northern District of Texas or the Eastern District of Texas, but individual judges have begun asking pointed questions during conferences. At least two reported instances in 2024 involved NJ federal practitioners who submitted briefs containing citations that opposing counsel flagged as non-existent or materially misquoted.

In one instance, the court required a corrective submission and an in-chambers explanation. In the other, the issue was raised in a Rule 11 letter that never made it to formal motion, but only because counsel caught and corrected the citation before the reply deadline.

Neither became a disciplinary headline. Both could have.

The New Jersey Office of Attorney Ethics (OAE) has not yet issued specific AI guidance, but its existing framework under RPC 8.4(c), conduct involving dishonesty or misrepresentation, would apply if a pattern of unchecked AI-generated submissions were treated as reckless rather than merely negligent. The line between the two is thinner than most attorneys expect.

The Verification Workflow You Should Actually Have

The question solo and small-firm practitioners ask me most often on this topic is some version of: "How do I use AI for legal research without running into this?" The answer is a short, non-negotiable verification step between the AI output and the brief.

Here's how I'd structure it:

After any AI tool generates case citations, run every single citation through Westlaw, Lexis, or Fastcase before that citation appears in a draft that leaves your desk. Not after you've built the argument around it. Before. This sounds obvious, but the workflow failure I see repeatedly is attorneys treating the AI output as a first draft of the argument and then reverse-engineering citations to support it, never going back to independently confirm what the AI found.

The second step is quoting directly from the verified source, not from the AI's summary of it. AI tools routinely paraphrase holdings in ways that are subtly wrong, not fabricated entirely, but materially inaccurate in ways that matter for the proposition you're citing. If you can't pull the exact language from the primary source and confirm it matches what your brief says the court held, the citation isn't ready.

Third, build this into your matter management system as a closed-loop task, not an informal mental note. If you use Clio, MyCase, or even a shared Google Doc for matter tracking, create a checklist item: "All citations independently verified via primary source." It takes 30 seconds to add and creates a file-level record that you completed this step.

One More RPC 3.3 Exposure Most Attorneys Don't Think About

AI tools don't just fabricate citations. They sometimes cite real cases for propositions those cases don't support. This is arguably worse than a fully hallucinated citation, because Westlaw confirms the case exists, the attorney stops checking, and the brief goes out with a real citation standing for a false legal proposition.

Under RPC 3.3(a)(1), submitting a real case for a false proposition is still a false statement of law to the tribunal. The citation being "real" is not the end of the inquiry.

If you're using AI to draft any portion of a brief, the verification standard is: does this case exist, and does it actually stand for what I'm saying it stands for? Both questions require independent review. The second one can't be delegated back to the tool that generated the citation in the first place.

The OAE will eventually issue guidance that makes this explicit. Until it does, the existing RPC 3.3 framework is more than sufficient to reach the same result.