Attorney-Client Privilege and AI: Complete Legal Guide 2026
Table of Contents
The Privilege Risk with AI
Attorney-client privilege is the foundation of the legal profession. Privilege protects confidential communications between attorney and client made for the purpose of obtaining legal advice. This privilege does not extend to third parties.
When lawyers use AI tools—especially cloud-based tools—they risk waiving privilege by disclosing confidential information to third parties. If you upload privileged documents to ChatGPT (which uses your data for model training), you may have waived privilege and must disclose the document to opposing counsel.
This is not hypothetical. Courts have ruled that uploading privileged documents to certain AI tools waives privilege. Lawyers must understand the privilege risks of each AI tool before use.
Privilege Doctrine Basics
Attorney-client privilege protects:
- Confidential communications between lawyer and client
- Made for the purpose of obtaining legal advice
- From disclosure to third parties
Key point: if a confidential communication is disclosed to third parties, privilege is waived—the communication loses its protection and must be disclosed to opposing counsel in litigation.
Key Privilege Risks with AI
1. Model Training on Your Data
Some AI tools use customer documents to train their models. OpenAI's default policy permits training on ChatGPT input unless you opt out. If you upload a privileged contract to ChatGPT, OpenAI may use it for model training, waiving privilege.
Safe tools: Tools that explicitly do NOT train on customer data (Harvey AI, Claude for Slack Enterprise, Microsoft Copilot for Microsoft 365 Enterprise).
2. Disclosure to Subpoena
Some AI tools do not have strong legal protections against law enforcement requests. If law enforcement subpoenas your AI tool vendor, they could obtain your documents, waiving privilege.
Safe tools: Tools operated by companies in strong jurisdictions with legal resources to fight subpoenas (Microsoft, Google, Apple, Anthropic).
3. Breach of Security
If an AI tool is breached and your privileged documents are stolen, courts may find privilege waived by lack of reasonable security.
Safe tools: Tools with SOC 2 Type II certification, encryption in transit and at rest, and strong incident response.
4. Delegation to Non-Attorneys
If a paralegal uses an AI tool without attorney supervision, privilege may be waived—the paralegal is not protected by privilege.
Safe practice: Use AI tools only under attorney direction, not delegated to support staff.
Essential Vendor Contract Requirements
Before using any AI tool with privileged documents, your firm must have written agreements with the vendor covering:
- Data Confidentiality: Vendor must maintain documents in confidence and not disclose to third parties.
- No Model Training: Vendor explicitly does not train models on customer documents.
- Data Deletion: Vendor must delete documents upon request or termination.
- Subpoena Procedure: Vendor will notify customer of subpoena and seek protective order before disclosing documents.
- Security Standards: Vendor maintains SOC 2 Type II or equivalent certification.
- Incident Response: Vendor has incident response plan and will notify customer of breaches within 48 hours.
- Privilege Acknowledgment: Vendor acknowledges that documents are attorney work product and privileged.
Bar Association Guidance
ABA Opinion 512 (2023)
The ABA's formal opinion on AI and legal practice requires lawyers to understand the AI tools they use and assess their reliability. Specifically, for confidential documents, lawyers must ensure the tool maintains privilege.
State Bar Opinions (2024-2026)
Most state bars follow ABA guidance: AI tools are permitted IF the lawyer validates the tool's data handling, maintains privilege, and understands limitations. New York State Bar, California State Bar, and Illinois State Bar have all issued favorable opinions on legal AI use under these conditions.
Key Rule: Attorney Competence
Lawyers must have sufficient knowledge of AI to use it appropriately. Blind reliance on AI without understanding its limitations is professional misconduct. This applies doubly to privilege: you must understand how the tool handles confidential information.
Safe Use Practices for Privilege Protection
Use Appropriate Tools
- For privileged work: Harvey AI, Claude for Slack Enterprise, LexisNexis, Westlaw, Microsoft Copilot for Microsoft 365 Enterprise
- For general work: ChatGPT Enterprise, Copilot
- Never for privileged work: Free ChatGPT, Gemini, Claude.ai (consumer versions)
Use Under Attorney Direction
Do not delegate AI tool use to paralegals or assistants without attorney supervision. Privilege requires attorney control and direction.
Do Not Waive Privilege Inadvertently
Avoid sharing privileged documents with opposing counsel or courts unless necessary. When disclosing documents, claim privilege on the privilege log, do not produce AI-analyzed versions that might waive privilege through selective disclosure.
Document Your Use
Maintain records of AI tool use: which tool, which documents, what was the purpose, why did you choose that tool? This documentation helps defend privilege if questioned.
Privilege Protection Checklist
Before Using AI on Privileged Documents
- Has general counsel approved this AI tool for privileged work?
- Is there a written vendor agreement covering data confidentiality?
- Does vendor agreement explicitly state documents will not be used for model training?
- Does vendor have SOC 2 Type II or equivalent security certification?
- Are you using the tool under attorney direction (not delegated)?
- Have you documented the business purpose for using this tool?
- Does vendor agreement include data deletion obligations?
- Does vendor agreement require notification of legal requests?
Frequently Asked Questions
Can we use ChatGPT for privileged work?
Not the free version. ChatGPT uses your input for model training by default. ChatGPT Enterprise has data privacy controls, but we recommend specialized legal AI (Harvey, legal research tools) for high-stakes privilege matters.
What if we accidentally upload privileged documents to ChatGPT?
Contact OpenAI immediately and request deletion. Privilege is likely waived if documents were used for training. Consider whether the disclosure requires notification to affected clients. Consult with your state bar if unsure.
Does our firm need a separate AI policy for privilege?
Yes. Your AI policy should explicitly restrict privileged document analysis to approved tools, require attorney direction, and require written vendor agreements before use.
Can we use on-premise AI to avoid privilege risks?
Yes, on-premise deployment (AI runs on your servers, documents never leave your network) eliminates most privilege risks. Harvey AI and some other tools offer on-premise options for enterprise clients.
Related: All Legal AI Tools | Enterprise AI Security Guide