What Information Should Lawyers Never Share with ChatGPT?

Client-identifying information

Client names, aliases, company names, or any combination that points to a specific person. This includes partial names when combined with other context — "John S., the Chicago surgeon" in a malpractice filing — that makes identification possible even without a full surname. Nicknames and professional handles also count.

The test is not whether the name is "technically" in the prompt. The test is whether the information, in combination with other details in the prompt, would allow a third party to identify the client. Combination risk is the most commonly overlooked category.

Matter-specific identifiers

Case numbers, docket references, court name combined with filing date, or any combination that links a prompt to a specific legal proceeding. Dates distinctive to the matter — a contract execution date, incident date, or hearing date — narrow the identification window even without a case number present.

Contact and location information

Client email addresses, phone numbers, home addresses, business addresses, or GPS coordinates referenced in a matter. Location data is consistently underestimated as an identifier, particularly when combined with a date and a distinctive injury type or transaction size in the same prompt.

Financial figures tied to a specific matter

Settlement amounts, damages calculations, or specific financial positions. An attorney handling a $4.3 million wrongful termination matter in a mid-size city submitting that figure — even without a name — has submitted identifying information if the matter is distinctive at that scale and in that market.

Government-issued identifiers

Social Security numbers, EINs, passport numbers, and driver's license numbers. These are the most obvious category — and also the ones most likely to be caught by a careful attorney. The prior categories are more commonly missed.

Medical information

Diagnoses, treatment records, ICD codes combined with dates, or any medical detail linked to a specific client. Medical information combined with other identifiers — a rare condition in a small town, a specific surgical date — can be identifying even when the patient's name is absent.

The automated alternative

Naosuu detects all of the above and replaces each instance with an anonymous placeholder before the prompt is sent. The attorney types normally; Naosuu handles the compliance step. The AI receives a prompt with [PERSON_1], [CASE_NO_1], and [AMOUNT_1] in place of identifying content, responds using those placeholders, and Naosuu restores the real values in the response before the attorney reads it.

The result is that the checklist above becomes irrelevant — Naosuu catches what the attorney misses, including the combination risk that manual review rarely surfaces.