Does uploading privileged documents to an AI tool waive attorney-client privilege?

Potentially yes, depending on the tool's terms and how the disclosure is made. Attorney-client privilege protects confidential communications between attorney and client made for the purpose of legal advice. The privilege can be waived if the communication is disclosed to third parties in a manner inconsistent with maintaining confidentiality. If an AI vendor's terms permit the vendor to use uploaded content for model training, or if uploaded documents are stored on vendor servers accessible to vendor employees without confidentiality protections, this could constitute a disclosure inconsistent with maintaining privilege. Review data processing terms carefully before uploading any privileged documents.

What do ethics rules require before using AI for legal work?

Model Rule 1.1 (Competence) requires lawyers to keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology. Most state bars interpret this to require lawyers using AI tools to understand how those tools work, their limitations, and when to verify outputs. Model Rule 1.6 (Confidentiality) requires lawyers to make reasonable efforts to prevent the inadvertent or unauthorized disclosure of client information. This requires vetting AI tools for data handling before use. Rules 5.1 and 5.3 require supervisory lawyers to take reasonable steps to ensure that subordinates and non-lawyer assistants comply with the rules, which extends to AI tool use.

What are the court rules on AI disclosure?

Multiple federal courts have adopted standing orders or local rules requiring disclosure of AI use in filings. The Southern District of New York requires attorneys to certify that any AI-generated content in filings has been reviewed for accuracy and that the filer takes responsibility for the content. The Northern District of California has similar requirements. Some courts require disclosure of which AI tools were used. Requirements are developing rapidly; check the specific court's local rules and standing orders before any filing in a case involving AI-assisted work.

What happened in Mata v. Avianca and what does it mean for lawyers?

In Mata v. Avianca (S.D.N.Y. 2023), attorneys submitted a brief citing multiple court decisions that did not exist. The attorneys had used ChatGPT to conduct legal research and the tool had generated plausible-sounding but entirely fabricated case citations. When opposing counsel and the court could not locate the cases, the attorneys doubled down rather than checking the citations. The court sanctioned the attorneys, finding that they had failed to verify AI-generated research outputs. The case established that relying on AI legal research without independent verification is a sanctionable failure of professional responsibility. Every AI-generated citation must be independently verified before filing.

What should a legal department look for in an AI vendor contract?

Four things are non-negotiable for legal work. First, a written commitment that the vendor will not use your uploaded data to train or fine-tune any model. Second, a data processing agreement specifying that client data is processed only on your behalf and not disclosed to other parties. Third, SOC 2 Type II certification or equivalent security certification. Fourth, a data deletion provision ensuring that all uploaded data is deleted on contract termination. Legal departments should also review the vendor's sub-processor list and confirm that any sub-processors handling confidential data have equivalent confidentiality commitments.

Can a general counsel approve AI tools for the whole legal department?

Yes, but the approval process must be documented and include substantive review of data handling terms, not just commercial terms. Best practice is for the general counsel or a designated legal AI committee to maintain an approved-tools list, review new tools before approval using a standard checklist covering privilege implications, ethics rule compliance, and security certifications, and require that all legal staff use only approved tools for client-related work. Informal or individual adoption of AI tools by lawyers without central review creates inconsistent risk management and may result in inadvertent disclosure of privileged information.

AI governance for legal teams and general c…

Legal documents and gavel on a desk

TL;DR: Legal departments using AI face three distinct risk categories: privilege waiver if confidential client documents are uploaded to AI tools with broad data terms, ethics rule violations under Model Rules 1.1, 1.6, 5.3, and sanctions exposure if AI-generated legal citations are submitted to courts without verification. Before deploying any AI tool for legal work, confirm it has a no-training commitment, a data processing agreement, and that your approved-tools list has been reviewed by the supervising attorney.

Legal departments were among the fastest adopters of AI tools in 2024 and 2025. Contract review, legal research, discovery review, and policy drafting are natural fits for AI assistance. But the legal profession operates under a distinct set of rules that most technology deployments do not encounter: confidentiality duties that attach to client information, privilege rules that protect legal advice from disclosure, ethical obligations that are personally enforceable against the lawyer, and court rules that have begun specifically addressing AI use in filed documents.

Deploying AI in a legal department without a governance framework is not just a data protection question. It is a professional responsibility question. Violations can result in bar discipline, client lawsuits, and court sanctions, not just regulatory penalties.

This guide covers the specific risks, what the rules actually require, where the case law stands, and what a legal department governance framework should include.

Why legal AI is a distinct risk category

Most corporate AI governance frameworks focus on data protection compliance, vendor security, and business risk. These are necessary but not sufficient for legal departments. Three additional risk categories apply only to legal work.

Attorney-client privilege is a rule of evidence that protects confidential communications between attorneys and clients made for the purpose of seeking or providing legal advice. The privilege belongs to the client, but lawyers are obligated to preserve it. Privilege can be waived by disclosure to third parties. When a lawyer uploads a privileged document to an AI tool, the legal analysis of whether that constitutes a waiver-triggering disclosure depends on the terms under which the tool processes data. An AI tool that uses uploaded content for model training, or that stores content in ways accessible to vendor employees, may compromise privilege even if no actual disclosure to an adversary occurred.

Confidentiality under ethics rules is broader than privilege. Model Rule 1.6 requires lawyers to protect all information relating to the representation of a client, not just privileged communications. This includes business information, strategic plans, and any other information a client shares with counsel. The duty to protect this information extends to choosing technology carefully.

Individual professional liability is the third distinct category. When a software vendor makes a mistake, the company may bear liability. When a lawyer submits a document containing AI-generated errors to a court, or advises a client based on AI-generated legal research that turns out to be wrong, the lawyer bears professional responsibility. The legal profession does not permit outsourcing of responsibility to tools.

The 6 contracts that need AI-specific language now

Most legal departments are focused on what AI tools their lawyers can use. Fewer have systematically reviewed the contracts their organization signs and whether those contracts address AI at all. These six contract types need AI-specific language added now, before the next renewal cycle.

1. Vendor agreements for AI tools. Three provisions belong in every vendor agreement where AI is a core feature of the service. First, require advance notification of model changes: 30 days notice before any update that materially affects output behavior gives your team time to re-evaluate risk before the change goes live. Second, include an explicit right to terminate if the vendor uses your data to train without your written consent, with no cure period. Third, negotiate indemnification covering IP infringement in AI-generated outputs, since vendors who use unlicensed training data may expose your organization to downstream copyright claims.

2. Employee acceptable use policies. An acceptable use policy that doesn't address AI specifically is not adequate for 2026. Add three requirements: employees must disclose when AI drafted customer-facing content, specifically in communications that a recipient would reasonably assume were written by a person; employees are prohibited from entering client data, confidential business information, or personnel data into consumer AI tools not on the approved list; and employees must disclose AI use in legal documents, court filings, and regulatory submissions, which may also be required by court rules independently of your internal policy.

3. Customer-facing terms of service. If your products or services use AI in customer interactions, your terms should now explicitly disclose that use, specify what categories of customer data the AI processes, and include an AI limitation clause stating that AI-generated outputs are not professional advice and that customers should not rely on them for consequential decisions without independent verification. The FTC's deceptive practices framework makes undisclosed AI use in consumer contexts a live enforcement risk.

4. NDAs and confidentiality agreements. Standard NDAs were written assuming that confidentiality means not sharing information with other people or companies. They do not contemplate what happens when an employee enters confidential information into an AI platform that processes it on third-party servers. Add explicit language treating AI platforms as third parties for confidentiality purposes: entering confidential information into an AI tool that is not covered by an appropriate data processing agreement constitutes disclosure to a third party under the NDA. This language protects the organization when employees use consumer AI tools with confidential counterparty information.

5. Data processing agreements. Any vendor who processes personal data on your behalf needs a DPA. For AI vendors, the standard DPA template almost certainly doesn't cover the specific ways AI systems process data, because AI-specific processing, including inference, fine-tuning, and model improvement, is distinct from conventional data processing. Review each AI vendor DPA to confirm it covers the actual processing the system performs, names all subprocessors including the underlying AI API providers (OpenAI, Anthropic, Google, etc.), and explicitly prohibits using your data for model training outside the scope of the services you are paying for.

6. Insurance riders. Most organizations have not notified their insurers that they now use AI tools in professional work. That notification matters. Errors and omissions and directors and officers policies written before 2023 may not clearly cover AI-related claims, including claims arising from AI-generated errors in professional deliverables or from AI-assisted decisions that are later challenged. Check whether E&O and D&O policies cover AI-related claims explicitly. Emerging AI-specific liability products from Lloyds syndicates and AIG cover hallucination-based professional liability, IP indemnification gaps, and AI system failures. This is a new and still-developing insurance category, but the exposure is real.

General counsel who wait for AI-specific legislation to require these updates are already behind. Clients and counterparties are asking about AI use in negotiations today.

Attorney-client privilege and AI tools

The privilege waiver analysis for AI tools turns primarily on the vendor's data handling terms.

Under the third-party disclosure doctrine, privilege may be waived when privileged information is shared with a third party in a manner inconsistent with maintaining confidentiality. Courts have found privilege waiver where clients shared confidential information with accountants, public relations firms, and other consultants without maintaining appropriate confidentiality safeguards.

For AI tools, the key questions are: Does the vendor's terms of service permit the use of uploaded content for model training? Does the vendor store uploaded content on servers accessible to vendor employees? Does the vendor share uploaded content with sub-processors who do not have equivalent confidentiality commitments? If any of these questions receive an affirmative answer, there is material privilege risk from uploading privileged documents.

Consumer-tier AI tools typically have the most concerning terms. OpenAI's default ChatGPT product historically included provisions permitting use of inputs for model improvement. Enterprise or API products typically have different terms with explicit no-training commitments. The same AI technology can be deployed under very different data handling terms depending on the tier.

The practical rule for legal departments: do not upload client documents to any AI tool without first reviewing the data processing terms at the level of the specific product tier being used, not just the vendor's general website.

Person reviewing legal documents on a laptop

Model rules of professional conduct and AI

The ABA's Model Rules create three specific obligations relevant to AI use.

Rule 1.1 (Competence) requires keeping abreast of changes in law and practice, including the benefits and risks of relevant technology. State bar opinions applying this to AI consistently hold that competent use requires understanding the tool's limitations, knowing when outputs need verification, and not delegating professional judgment to the system.

Rule 1.6 (Confidentiality) requires reasonable efforts to prevent inadvertent or unauthorized disclosure of client information. Bar opinions from New York, California, Florida, and other states have applied this to AI tools, concluding that lawyers must vet AI vendors for data handling before using them for client work.

Rules 5.1 and 5.3 (Supervision) require supervising lawyers to ensure subordinates and non-lawyer assistants comply with conduct rules. Bar opinions have concluded that AI tools are the functional equivalent of non-lawyer assistants for Rule 5.3 purposes: supervising lawyers are responsible for AI-generated output used by their staff.

State bar guidance and ABA Formal Opinion 512

The American Bar Association issued Formal Opinion 512 in 2024, providing the most comprehensive bar-level guidance on AI use in legal practice. Key holdings include: lawyers must understand the AI tool they are using well enough to assess its reliability for the specific task; confidential client information may not be shared with AI tools without appropriate protections; AI-generated work product must be reviewed and verified by a lawyer before reliance; and lawyers cannot disclaim responsibility for AI errors by attributing them to the tool.

State bars have followed with their own guidance. New York State Bar Association Opinion 1253 addressed confidentiality requirements for AI legal tools, concluding that using AI tools that do not provide adequate data protection for confidential information violates Rule 1.6. The Florida Bar and California State Bar have issued similar guidance. Texas has addressed AI use in litigation specifically, emphasizing verification requirements for AI-generated legal research.

The trend in state bar guidance is consistent: AI tools are permitted, but they are subject to the same professional responsibility framework as any other aspect of legal practice. There is no technology exception to confidentiality, competence, or supervision requirements.

AI hallucination risk and the Mata v. Avianca sanctions

In Mata v. Avianca (S.D.N.Y. 2023), attorneys submitted a brief citing six court decisions that did not exist. They had used ChatGPT for legal research; the tool generated plausible-sounding citations with realistic docket numbers and party names. When opposing counsel could not locate the cases, the filing attorneys doubled down. The court sanctioned all of them for failing to verify the accuracy of their court submissions.

The lesson is that AI legal research outputs, particularly case citations, must be verified in Westlaw or Lexis before use in any filing or client advice. AI language models generate plausible text, not verified facts. Since Mata, at least a dozen additional sanctions orders have targeted attorneys for unverified AI-generated citations.

Court-specific AI policies

A growing number of federal courts have adopted specific AI disclosure requirements or standing orders. As of mid-2026, the following applies.

The Southern District of New York requires attorneys to certify in filings involving AI-generated content that the content has been reviewed for accuracy and that the filing attorney takes full responsibility for it. Several SDNY judges have individual standing orders with additional requirements.

The Northern District of California has similar certification requirements. Multiple California federal judges have issued individual orders requiring disclosure of the specific AI tools used and confirmation of independent verification.

The Fifth Circuit and several district courts within it have standing orders addressing AI in appellate briefs and district court filings.

Requirements are not uniform. Local rules and individual judge standing orders vary. The practical obligation before any filing is to check the court's current local rules and the assigned judge's standing orders specifically for AI provisions. This is not a one-time check; standing orders are updated frequently.

What to look for in a legal AI platform

Not all AI tools carry the same risk profile for legal work. Purpose-built legal AI platforms with enterprise data protections present materially different risks than general-purpose consumer AI tools.

The minimum acceptable data handling commitments for legal work are: a written no-training commitment specifying that the vendor will not use your inputs or outputs to train, fine-tune, or improve any model; a data processing agreement that specifies the purposes for which your data is processed and the security measures in place; SOC 2 Type II certification demonstrating that the vendor has independently verified security controls; and a contractual obligation to delete all data on contract termination.

Additional factors to evaluate: whether the platform is purpose-built for legal use cases and incorporates citation verification into the workflow; whether it provides source citations with responses so that verification is straightforward; whether it processes data in jurisdictions consistent with your clients' data transfer restrictions; and whether it offers privileged data processing under a formal legal services wrapper.

Vendors who offer legal AI products include Harvey, Lexis AI, Westlaw Precision, CoCounsel (Thomson Reuters), and enterprise configurations of Claude and GPT-4o with appropriate data processing agreements. Evaluate each specifically for your data handling requirements.

Legal department governance checklist

Before deploying any AI tool for legal work, work through this governance checklist:

Designate a legal AI governance lead, typically the general counsel or a senior attorney, who is responsible for the approved-tools list and vendor review.
Establish an approved-tools list covering all AI tools that may be used for client-related work. Require that all legal staff use only tools on this list for any work involving confidential client information.
For each approved tool, document the data handling review: confirm no-training commitment, data processing agreement, security certification, and deletion provisions.
Establish a verification requirement: AI-generated legal research (cases, statutes, regulations) must be verified in an authoritative database before any use in court filings or client advice. Document this requirement in the department's AI policy.
Train all legal staff on the verification requirement and on confidentiality obligations with respect to AI tools. Document the training.
Check court-specific AI disclosure requirements before each filing in any court with AI rules or standing orders.
Review the department's malpractice insurance coverage for AI-related errors. Confirm that coverage applies to claims arising from AI-assisted legal work.
Establish an incident response process for AI-related errors: what happens if an AI-generated error is discovered after a filing is submitted or client advice is provided.

Legal documents and gavel on a desk

TL;DR: Legal departments using AI face three distinct risk categories: privilege waiver if confidential client documents are uploaded to AI tools with broad data terms, ethics rule violations under Model Rules 1.1, 1.6, 5.3, and sanctions exposure if AI-generated legal citations are submitted to courts without verification. Before deploying any AI tool for legal work, confirm it has a no-training commitment, a data processing agreement, and that your approved-tools list has been reviewed by the supervising attorney.

This guide covers the specific risks, what the rules actually require, where the case law stands, and what a legal department governance framework should include.

Designate a legal AI governance lead, typically the general counsel or a senior attorney, who is responsible for the approved-tools list and vendor review.
Establish an approved-tools list covering all AI tools that may be used for client-related work. Require that all legal staff use only tools on this list for any work involving confidential client information.
For each approved tool, document the data handling review: confirm no-training commitment, data processing agreement, security certification, and deletion provisions.
Establish a verification requirement: AI-generated legal research (cases, statutes, regulations) must be verified in an authoritative database before any use in court filings or client advice. Document this requirement in the department's AI policy.
Train all legal staff on the verification requirement and on confidentiality obligations with respect to AI tools. Document the training.
Check court-specific AI disclosure requirements before each filing in any court with AI rules or standing orders.
Review the department's malpractice insurance coverage for AI-related errors. Confirm that coverage applies to claims arising from AI-assisted legal work.
Establish an incident response process for AI-related errors: what happens if an AI-generated error is discovered after a filing is submitted or client advice is provided.

AI governance for legal teams and general counsel: privilege, confidentiality, and risk 2026

Why legal AI is a distinct risk category

The 6 contracts that need AI-specific language now

Attorney-client privilege and AI tools

Model rules of professional conduct and AI

State bar guidance and ABA Formal Opinion 512

AI hallucination risk and the Mata v. Avianca sanctions

Court-specific AI policies

What to look for in a legal AI platform

Legal department governance checklist

AI governance for legal teams and general counsel: privilege, confidentiality, and risk 2026

Why legal AI is a distinct risk category

The 6 contracts that need AI-specific language now

Attorney-client privilege and AI tools

Model rules of professional conduct and AI

State bar guidance and ABA Formal Opinion 512

AI hallucination risk and the Mata v. Avianca sanctions

Court-specific AI policies

What to look for in a legal AI platform

Legal department governance checklist