Stop Sending AI-Drafted Demand Letters Without Running This One Check First

A solo attorney in a New Jersey personal injury practice told me she had cut her demand letter drafting time from two hours to about twenty minutes using an AI writing tool. That's real. The productivity gain is genuine. Then she forwarded me one of the outputs, and I spotted a sentence that, if sent to an opposing insurer, would have been a textbook violation of RPC 8.4(c): conduct involving dishonesty, fraud, deceit, or misrepresentation.

She hadn't noticed it. The AI had inflated the treating physician's prognosis language in a way that contradicted the actual medical records she had uploaded as source material. The model didn't hallucinate a citation this time. It hallucinated a fact.

This is the specific risk that doesn't get enough attention in conversations about AI and ethics. Most of the NJ bar's guidance, and most of what's circulating in CLE content, focuses on confidentiality (RPC 1.6), competence (RPC 1.1), and supervision (RPC 5.3). Those matter. But RPC 8.4 is where demand letter AI errors actually land, because demand letters make factual representations to third parties. An inflated wage loss figure, a mischaracterized diagnosis, a fabricated treatment timeline: each of those is a potential misrepresentation, and the New Jersey Rules of Professional Conduct don't have a carve-out for unintentional ones produced by software.

What the AI actually does with your source documents

When you upload a medical record, a police report, or a wage statement into an AI drafting tool and ask it to write a demand letter, the model attempts to synthesize that material into persuasive prose. The problem is that "persuasive" and "accurate" pull in different directions. Language models are trained on enormous amounts of advocacy writing, which means they have a strong prior toward framing facts favorably. Left unchecked, that tendency produces letters that read well but overstate the record.

The specific failure modes to watch for:

• The model rounds up. A client who missed "approximately six weeks" of work becomes one who missed "more than six weeks." That's small, but it's not what your records say.
• The model interpolates. If the orthopedic note says "patient reports pain at a 7/10," the AI may write that the patient "continues to suffer significant pain," dropping the subjective qualifier entirely.
• The model borrows framing from similar cases in its training data. If your client's prognosis is genuinely uncertain, the letter may read as though future surgery is more likely than the treating doctor actually stated.

None of these are hallucinations in the sense of invented case citations. They're tone and degree errors, and they're much harder to catch on a quick read because the overall letter looks right.

The one review step most NJ solo firms skip

Before a demand letter goes out, most attorneys read it for completeness and tone. That's not enough when AI drafts it.

The check that actually catches RPC 8.4 exposure is a claim-by-claim citation audit: every factual assertion in the letter should map back to a specific line in a specific document in your file. Not generally consistent with the file, but traceable to a source.

This doesn't have to take two hours. A structured review template works well here. Set up a simple table with three columns: the factual assertion from the letter, the supporting document, and the page or paragraph. If you can't fill in column two and column three, the assertion doesn't go out. Some attorneys have started using AI tools themselves to help run this audit, asking the model to cross-reference its own output against the uploaded source documents. That's a reasonable secondary check, but it doesn't replace reading the actual records yourself. Models can miss their own errors when asked to self-review.

Why NJ specifically matters here

The New Jersey Supreme Court's standard under RPC 8.4 tracks intentional misconduct, but the Office of Attorney Ethics has pursued cases where an attorney "should have known" the representation was false. If you're signing a demand letter, you're vouching for its contents. The fact that software drafted the inflated language hasn't historically been a defense, and there's no reason to think it will become one.

The NJSBA hasn't issued specific formal guidance on AI-generated demand letters as of mid-2025, but the general ethics framework is clear: the sending attorney is responsible for what the letter says.

If you're using AI drafting tools at volume, the single most protective step you can take isn't a vendor audit or a policy document. It's building the citation review into your intake-to-demand workflow before the first AI-drafted letter ships, not after a complaint arrives.