The Legal Team's ChatGPT Playbook: Redact, Draft, Verify
A three-step workflow legal teams can actually adopt without violating client privilege, court rules or professional conduct obligations.
The privilege question
In most common-law jurisdictions the attorney-client privilege attaches to communications made in confidence for the purpose of legal advice. Voluntary disclosure to a third party without a confidentiality obligation waives it. A consumer LLM provider is a third party; its terms of service do not create a confidentiality obligation of the kind courts recognise. This is why bar associations from New York to California to England & Wales have issued specific guidance.
The three-step workflow
Redact
Strip client names, matter numbers, opposing party names, dates, jurisdictions, monetary amounts, and any facts that would identify the matter to a knowledgeable reader. The tokens PromptShielder emits (`[NAME_1]`, `[ORG_1]`, `[DATE_1]`) are read as placeholders by every model we've tested.
Draft
Send the redacted prompt. Ask for the structure, argument, cite check or draft. Iterate against the redacted version.
Verify
Restore the placeholders. Read the output as if a paralegal drafted it — every cite check, every factual claim, every quotation. Nothing goes to the client or the court unread.
What not to paste even after redaction
- Sealed exhibits and material under a protective order.
- In-camera submissions.
- Communications with government agencies subject to specific restrictions.
Frequently asked
Does an enterprise ChatGPT DPA restore privilege?+
It changes the confidentiality analysis but does not automatically restore privilege — courts have not yet settled the question. Redaction remains the safer default.
Can I use ChatGPT for cite checking?+
For structure, yes. For the authority itself, always confirm against a primary source or paid legal database. Models still fabricate citations.