Attorney-Client Privilege and Generative AI: A Guide to Risk and Compliance

Does Using AI Waive Attorney-Client Privilege?

The core question facing legal professionals today is whether inputting client data into a Large Language Model (LLM) constitutes a waiver of privilege.

Under the Third-Party Doctrine, disclosing confidential information to a third party generally waives privilege. However, an exception exists for "necessary intermediaries" (like translators or outside counsel paralegals).

If you paste a confidential memo into a public chatbot, you have likely waived privilege.

Why Consumer AI Tools Are High-Risk for Law Firms

To understand the risk, you must look at the Terms of Service of standard AI platforms:

• Data Retention: Platforms often store prompts for 30 days or more.
• Model Training: Unless explicitly opted out, inputs are used to train future model iterations.
• Human Review: "Safety reviews" allow third-party contractors to read logs to prevent abuse.

In the eyes of a court, this lack of privacy is a failure to take reasonable precautions to protect client confidences (ABA Model Rule 1.6).

The 4 AI Data-Handling Models: A Taxonomy for Lawyers

Not all AI is built the same. When evaluating legal tech, you will encounter four distinct data architectures. Your "privilege safety" depends entirely on which model you choose.

AI Model Type	Example	Data Retention	Human Review?	Privilege Risk
1. Consumer Chat	ChatGPT (Free/Plus)	30+ Days	Yes	High (Avoid)
2. Commercial API	GPT-4 via Azure (Default)	Configurable	Possible	Moderate (Requires DPA)
3. Zero Data Retention	Enterprise Legal AI	0 Days	No	Safe (Preferred)
4. Client-Side / Local	On-Premise LLMs	Never leaves device	No	Gold Standard

Legal Ethics Opinions on AI and Confidentiality

Recent ethics opinions from major jurisdictions, including the Florida Bar (Op. 24-1), California, and New Jersey, have converged on a single standard: Duty of Competence.

You do not need to be a coder, but you must understand:

1. Where the data is physically stored.
2. Who has access to it.
3. Whether it can be deleted.

If you cannot answer these three questions about your AI tool, you should not use it for client work.

A Practical "Safe AI" Workflow for Attorneys

If your firm permits the use of Generative AI, strictly adhere to this workflow to minimize the risk of accidental waiver.

1. The Redaction Protocol

Before prompting, strip all Personally Identifiable Information (PII) and entity names. Use placeholders:

• Instead of: "Draft a clause for the merger between Apex Corp and Beta Ltd for $50M."
• Use: "Draft a clause for the merger between [BUYER] and [TARGET] for [AMOUNT]."

2. Vendor Due Diligence (The Checklist)

Before signing a contract with a legal AI vendor, require them to answer these specific security questions:

• Zero Data Retention (ZDR): Can you guarantee prompts are deleted immediately after processing?
• Training Data: Will my data be used to train your base models? (The answer must be "No").
• SOC-2 Compliance: Do you hold a SOC-2 Type II or ISO 27001 certification?
• Subpoena Policy: Will you notify us within 48 hours if our data is subpoenaed?

3. Output Review

Treat AI output as you would the work of a first-year associate. It requires total verification. Mark all AI-generated drafts as "PRELIMINARY – AI ASSISTED" to track provenance.