Candor to the Court When an AI Wrote the First Draft: What NJ's Preliminary Guidelines Actually Require Under RPC 3.3

Every attorney who has used a generative AI tool to draft a motion or a brief has had the same quiet moment of confidence: I'll review it. I'll catch anything wrong. That review assumption carries most of the ethical weight for NJ practitioners, and it's worth spelling out exactly what happens when the assumption fails.

This isn't a hypothetical problem. It's a structural one baked into how large language models produce text. They don't retrieve law; they predict plausible-sounding text. An AI assistant might cite a case that exists, with a holding that sounds right, in a quotation that is slightly wrong. The citation clears a Westlaw search. The case is real. The quotation is a confabulation. And you filed it.

The guidelines set the floor, not the ceiling

New Jersey's Supreme Court issued its "Preliminary Guidelines on New Jersey Lawyers' Use of Artificial Intelligence" on January 24, 2024. Guideline 1 states flatly that attorneys are responsible for all work product, including AI-generated content, and that reliance on AI does not reduce or transfer that responsibility. Guideline 3 specifically addresses accuracy: attorneys must review AI-generated content for accuracy before filing or submitting it.

That language, review for accuracy, sounds simple. The problem is that "accuracy" in a brief has at least three distinct layers: factual accuracy, legal accuracy (did the case say what you think?), and quotation accuracy (are those the court's exact words?). The Preliminary Guidelines don't parse these layers, which means the burden of parsing falls on you.

Where RPC 3.3 sits in this

NJ RPC 3.3(a)(1) prohibits a lawyer from knowingly making a false statement of fact or law to a tribunal. NJ RPC 3.3(a)(3) goes further: a lawyer must not offer evidence the lawyer knows to be false and, if the lawyer later discovers that material evidence submitted was false, must take reasonable remedial measures.

The word "knowingly" does most of the work in RPC 3.3(a)(1). A subtly confabulated quotation isn't something most attorneys "know" is false the moment they file it. But RPC 3.3(a)(3)'s remedial duty kicks in the moment you do know, regardless of how the error got there. If you discover post-filing that the AI invented a quotation, the rule requires you to act, not just move on.

The Preliminary Guidelines reinforce this. They don't create a separate disclosure obligation on the face of every filing (a point many practitioners misread), but they make clear that the attorney's existing ethical duties, including candor, apply in full to AI-assisted work product. There is no AI exception written into RPC 3.3, and the Preliminary Guidelines don't create one.

The review problem in practice

Here's where the structural tension lives. The efficiency case for AI-assisted drafting depends, partly, on the assumption that a good review catches errors efficiently. That assumption holds well for logical gaps, weak arguments, and missing counterpoints. It holds less well for citation accuracy, because experienced attorneys tend to read cited cases for their general proposition rather than comparing quotations word-for-word against the original.

A practical workflow that closes this gap: treat every quoted passage in an AI-generated draft as a string that needs verification against the primary source, not against a secondary Westlaw summary. That's a different cognitive task than general review, and it needs to be a named step in your pre-filing checklist, not an implicit part of "looking it over."

For solo practitioners especially, the temptation is to compress this step when you're running against a deadline. The Preliminary Guidelines don't offer any deadline-based relief from the responsibility Guideline 1 assigns.

One more thing the guidelines don't say

The Preliminary Guidelines are explicitly preliminary. The NJ Supreme Court signaled it would revisit them as AI tools and their use in practice mature. The ACPE has not issued an AI-specific opinion as of this writing, and ACPE Opinion 701 (2006), which addresses electronic storage and cloud-based data, doesn't speak to AI-generated content or filing obligations.

That gap in formal opinion coverage means that for now, NJ attorneys are working from the intersection of the Preliminary Guidelines and the existing RPC text. That's enough to understand your obligations; it's just not enough to answer every edge case.

---

What the rule requires

NJ RPC 3.3(a)(1) prohibits knowingly making a false statement of fact or law to a tribunal. NJ RPC 3.3(a)(3) requires an attorney who learns that submitted material evidence was false to take reasonable remedial measures. The NJ Supreme Court's Preliminary Guidelines on Lawyers' Use of AI (January 24, 2024) provide in Guideline 1 that attorneys are responsible for all AI-generated work product, and in Guideline 3 that attorneys must review AI-generated content for accuracy before filing. These requirements apply cumulatively; the guidelines don't modify or limit the RPC text.

What I recommend

Build quotation verification into a named, mandatory step in your pre-filing workflow, separate from general substantive review. For every quoted passage in an AI-generated draft, open the original source and compare word-for-word before the document goes out. If you discover a confabulated quotation after filing, treat it as a RPC 3.3(a)(3) situation immediately: consult the rule text, consider whether the error was material, and document your remedial decision. Don't wait to see if opposing counsel or the court notices first.