Candor to the Tribunal When AI Wrote the Brief: What NJ Attorneys Must Know About RPC 3.3

A federal judge in Texas fined two attorneys $10,000 after they submitted a brief packed with AI-generated citations that didn't exist. A Manhattan lawyer faced sanctions for the same problem in Mata v. Avianca. These aren't cautionary tales from some distant future, they're a pattern, and New Jersey practitioners are not immune to it.

What's less discussed is the specific ethics rule that gets activated when something false or misleading makes it into a court filing. Most bar guidance on AI has clustered around RPC 1.1 (competence) and RPC 1.6 (confidentiality). But for litigators using AI to draft briefs, motions, or certifications, the rule that should be keeping you up at night is RPC 3.3: Candor Toward the Tribunal.

What RPC 3.3 Actually Requires

Under New Jersey RPC 3.3(a), an attorney shall not knowingly make a false statement of fact or law to a tribunal, fail to disclose directly adverse controlling authority, or offer evidence the lawyer knows to be false. The word "knowingly" does a lot of work in that sentence, and it creates a false sense of comfort when AI is involved.

Here's the practical problem: when a large language model (LLM) produces a citation to State v. Whoever, 487 N.J. Super. 312 (App. Div. 2021), and you file it without checking, you didn't "knowingly" submit a false citation in the traditional sense. But the New Jersey Rules define "knows" to include when a lawyer is "aware of such facts which would cause a reasonable lawyer to inquire further." If you're using a generative AI tool and not verifying citations, a disciplinary panel is going to have a hard time believing you lacked that awareness.

The duty under RPC 3.3 is not satisfied by good intentions or by the fact that your AI tool looked confident.

The Disclosure Question NJ Courts Haven't Fully Answered

Several federal courts, including the U.S. District Court for the District of New Jersey, have adopted standing orders or local rules requiring attorneys to disclose AI-assisted drafting or certify that AI-generated citations were verified. The District of New Jersey's own standing order on generative AI (effective 2024) requires that any attorney who uses generative AI to draft a filing must certify that the content, including citations, has been reviewed for accuracy.

State court in New Jersey has been slower to formalize this. The New Jersey Superior Court does not yet have a uniform statewide standing order on AI disclosure. That absence doesn't mean you're in the clear. It means your RPC 3.3 obligations are doing the work that procedural rules haven't yet filled.

If you're practicing in both state and federal court, you may be operating under two different procedural regimes on the same file. That's not a reason to apply the stricter standard only in federal court. Consistency is both practical and defensible.

Where the Risk Actually Comes From

The candor problem with AI-drafted briefs isn't just fabricated citations, though that gets the headlines. There are subtler failure modes worth knowing:

Mischaracterized holdings. An LLM might cite a real case for a proposition the case doesn't actually support. The citation checks out on Westlaw; the holding summary doesn't. This is harder to catch because you confirmed the case exists and stopped there.

Outdated authority. AI training data has cutoff dates. A case your tool cites as good law may have been reversed, distinguished, or superseded by a later New Jersey Supreme Court decision. An attorney filing a brief is responsible for verifying precedential status, regardless of how the research was generated.

Opponent's authority you missed. RPC 3.3(a)(2) requires disclosure of directly adverse controlling authority the other side hasn't cited. If your AI-assisted research process has gaps, it may not surface the one Appellate Division case that cuts against your argument. Manual verification of adverse authority is not optional.

A Pre-Filing Review Process That Actually Holds Up

Before any AI-assisted brief leaves your office for a New Jersey court, the review should include at least these steps:

Every case citation gets independently verified in Westlaw, Lexis, or Fastcase, not spot-checked, every one. This is non-delegable if you're the signing attorney. You check the pin cite, you read the relevant passage in the original opinion, and you run it through a citator. If you've delegated this to a paralegal or associate who used AI to compile the cite list, RPC 5.1 and 5.3 put the result on you anyway.

Propositions of law get traced back to primary source. If the AI drafted a sentence stating that New Jersey follows a particular standard, you locate the case or statute that actually establishes it. Summary language from an AI tool is not authority.

Adverse authority gets a dedicated search. Run a separate research pass specifically looking for cases that cut against your position. Don't rely on the same AI-generated memo that built your argument to flag what undermines it.

If you're in the District of New Jersey, review the operative standing order for the assigned judge before filing. Several judges have individualized AI disclosure requirements beyond the general standing order.

The Professional Responsibility Frame NJ Attorneys Should Use

There's a tendency to treat AI-related ethics issues as a technology problem. They're not. RPC 3.3 was written before anyone imagined generative AI, and it maps onto the AI context with uncomfortable precision. The rule doesn't ask whether your process was reasonable, it asks whether what you told the court was accurate.

The attorney who files an AI-drafted brief without verification isn't facing a competence question in isolation. If the brief contains a false statement of law, it's a candor question. And candor violations in New Jersey carry consequences that range from admonition to suspension, depending on whether the conduct was isolated or part of a pattern.

The safest working assumption is this: sign nothing you haven't read and verified, regardless of how it was drafted. That standard predates AI by decades. It just requires more active effort now that the drafts arrive so quickly and so fluently.