Supervising a Paralegal Who Uses AI Better Than You Do: What NJ RPC 5.3 Actually Requires

There's a situation showing up in small NJ firms that nobody talks about openly: the paralegal knows the AI tool better than the attorney does. She set up the prompts. She built the workflow. She's faster with it than you are, and the output looks good. So you review it, sign off, and move on.

That arrangement is probably fine. But "probably" is doing a lot of work in that sentence, and RPC 5.3 is where it starts to matter.

What RPC 5.3 Actually Says

New Jersey's RPC 5.3 requires attorneys with supervisory authority over non-lawyer staff to make reasonable efforts to ensure the firm has measures in place giving reasonable assurance that the non-lawyer's conduct is compatible with the professional obligations of the attorney. It also assigns direct responsibility to a supervising attorney if they order conduct, ratify it after the fact, or fail to take reasonable remedial action once they know about a problem.

None of that is new. What's new is the surface area. When a paralegal uses generative AI to draft a motion, summarize deposition transcripts, prepare a demand letter, or run background research on opposing counsel, each of those outputs becomes something the supervising attorney is responsible for, even if the attorney didn't touch the tool that produced it.

The relevant question isn't whether the paralegal is competent with the software. The question is whether the attorney has structured supervision so that problems in the AI output get caught before they reach a client or a court.

The Specific Risks RPC 5.3 Supervision Has to Cover

When a paralegal uses a generative AI tool without a defined review structure, a few failure modes come up repeatedly.

Hallucinated authority. AI legal research tools, even the ones marketed to law firms, still fabricate citations. A paralegal who doesn't know what a correct New Jersey citation format looks like, or who hasn't been trained to spot a case number that doesn't resolve in Westlaw, may pass that citation straight to a draft brief. The attorney who signs the brief owns that under RPC 3.3, but the supervision failure started earlier.

Client data in the wrong tool. Paralegals with access to client files will use whatever tool is fastest. If the firm hasn't specifically authorized which AI tools may receive client information, and if the paralegal hasn't been told that consumer-grade ChatGPT or Claude isn't covered by the firm's data handling standards, that information is going somewhere the attorney didn't sanction. That's a confidentiality exposure that traces back to the supervisory structure, or the absence of one.

Tone and accuracy in client-facing documents. Demand letters, status update emails, intake summaries, these are often handled by paralegals, and AI makes them faster to produce. But generative AI will confidently state things that are factually wrong about a client's matter if the prompt doesn't constrain it carefully. A supervising attorney who is reviewing volume output quickly may not catch a factual error buried in paragraph four of a letter that otherwise reads fine.

What "Reasonable Measures" Looks Like in a Solo or Two-Person Firm

The NJ Supreme Court's Advisory Committee on Professional Ethics hasn't issued an opinion specifically addressing AI supervision of non-lawyer staff as of this writing. But the structure of RPC 5.3 gives enough to work with.

Reasonable measures don't require elaborate bureaucracy. In a solo or small firm context, they tend to look like this:

A written (or at minimum documented) list of which AI tools the paralegal is authorized to use for which tasks. This doubles as the answer to a bar grievance if one ever comes up.

A defined review step for any AI-assisted work product before it leaves the office. "I review everything" is not a process. "Every AI-generated draft gets a line edit against the source documents before I approve it" is a process.

A short standing instruction to the paralegal on what to do when AI output looks off: flag it, don't fix it silently, and don't submit it under time pressure without attorney sign-off.

At least one explicit conversation, ideally documented in a memo or email, covering citation verification. If the paralegal is doing AI-assisted research, she needs to know that every case must be confirmed in an authoritative source before it goes in a brief. That's not a given. In many small firms, it's never been said out loud.

The Harder Conversation

Some attorneys are, at this point, supervising paralegals whose AI fluency has outpaced their own. That's not an embarrassing situation to acknowledge. It does create a structural problem worth thinking about carefully.

If you can't evaluate whether the output of a tool your paralegal uses is reliable, you can't fulfill the supervisory obligation RPC 5.3 places on you. The fix isn't to take the tool away. It's to build a review structure that catches errors even when you don't know exactly how they were generated. Focus on the output, not the process: does the draft accurately reflect the client's facts, cite verifiable authority, and avoid statements you can't independently confirm?

That review discipline is the same one that applies to any non-lawyer work product. AI just makes the volume higher and the errors harder to spot because the prose reads so cleanly.

One practical step worth taking this week: ask your paralegal to walk you through the last three things she used AI to produce. Not to audit her, but to understand where the tool enters the workflow and where her own judgment fills in. That conversation is, in itself, supervision.