When a Lawyer Sends Privileged Information to AI, Is It Still Protected?
Third-party AI platforms, confidentiality, privilege waiver, work-product protection, and the legal consequences of allowing outside systems to process client information.
Analysis by
Founder, Lex Wire Journal • Technology, Governance & Sovereignty Strategist
Consider a corporate attorney conducting due diligence for a proposed acquisition.
Among the documents are confidential communications between company executives and outside counsel concerning regulatory exposure, potential liabilities, and negotiation strategy.
The attorney uses an AI platform to summarize documents, identify contractual risks, and organize the information for further review.
The platform is operated by a third-party provider. Its terms govern how information is processed, retained, accessed, and potentially reused within the service.
The client retains ownership of the documents. The provider may even promise contractual confidentiality.
But does either fact establish that the communications remain protected by attorney-client privilege?
No.
Data ownership, ethical confidentiality, evidentiary privilege, and technical control are related but legally distinct concepts.
An agreement stating that a customer owns its information does not, by itself, establish who can access that information or whether a particular disclosure preserves privilege.
Likewise, processing information through third-party technology does not necessarily mean that privilege has been waived.
The legal consequences depend on the applicable law, the nature of the communication, the role of the provider, and the circumstances of the disclosure.
In my continuing legal education programs addressing artificial intelligence, professional responsibility, and client communication, I emphasize the importance of examining where client information travels and who can access it.
Those questions have become more consequential as lawyers integrate generative AI into research, document review, drafting, and litigation preparation.
Federal court decisions issued in 2026 have brought these issues into sharper focus.
In United States v. Heppner, a federal court rejected claims of attorney-client privilege and work-product protection over materials a represented defendant independently created using a consumer AI platform.
In Warner v. Gilbarco, Inc. and Morgan v. V2X, Inc., other federal courts recognized work-product protection in different circumstances involving self-represented litigants.
These cases do not establish a single national rule governing every AI platform.
They reveal the importance of identifying precisely which legal protection is being asserted.
The deeper question is not simply whether a law firm owns its client information.
It is who can access the information, for what purposes, under what authority, and with what consequences for the legal protections surrounding the communication.
The Bottom Line
Uploading confidential client information to an AI platform can create attorney-client privilege and professional responsibility risks, but waiver is not automatic in every circumstance.
Lawyers must distinguish ethical confidentiality under ABA Model Rule 1.6, evidentiary attorney-client privilege, and work-product protection. They must also evaluate the actual information-handling arrangements of the relevant AI provider.
The most useful professional question is not merely who owns the data, but who can access, retain, reuse, or disclose it, and whether those arrangements preserve the applicable legal protections.
Three Legal Protections That Must Not Be Confused
Discussions of artificial intelligence often use the terms confidentiality and privilege interchangeably.
That is a mistake.
The distinction matters because the legal consequences of processing client information through an AI platform depend on the protection involved.
Three separate doctrines are particularly important.
| Protection | General Scope | Principal Authority |
|---|---|---|
| Ethical Confidentiality | Information relating to a client representation, subject to applicable authorizations and exceptions. | ABA Model Rule 1.6 and applicable state professional rules. |
| Attorney-Client Privilege | Qualifying confidential communications made for the purpose of obtaining or providing legal advice. | Applicable federal or state privilege law. |
| Work-Product Protection | Qualifying materials prepared in anticipation of litigation or for trial. | Federal Rule of Civil Procedure 26(b)(3) and applicable doctrine. |
Ethical Confidentiality
ABA Model Rule 1.6 generally prohibits lawyers from revealing information relating to the representation unless disclosure is authorized by informed consent, is impliedly authorized, or falls within an applicable exception.
This ethical obligation extends beyond communications that qualify for privilege.
It may encompass business records, factual information, client identity information, and other material obtained during representation, depending on the circumstances.
Attorney-Client Privilege
Attorney-client privilege generally protects qualifying confidential communications involving a lawyer and client for the purpose of obtaining or providing legal advice.
Privilege is an evidentiary protection and can prevent compelled disclosure of qualifying communications.
Its scope and application depend on governing law. Federal Rule of Evidence 501 establishes important choice-of-law principles for privilege in federal proceedings.
Work-Product Protection
Work-product protection serves a different purpose.
It protects qualifying materials prepared in anticipation of litigation or for trial, including certain mental impressions, legal theories, and litigation strategy materials.
Its scope and waiver rules differ from those governing attorney-client privilege.
A communication may be confidential under Rule 1.6 without satisfying the elements of attorney-client privilege.
A document may also qualify for work-product protection even though it does not constitute a privileged attorney-client communication.
The first step in evaluating an AI-related disclosure is therefore to determine which protection applies.
Rule 1.6 and the Lawyer’s Duty to Safeguard Client Information
The ethical analysis begins with ABA Model Rule 1.6 .
Rule 1.6(a) governs disclosure of information relating to the representation.
Rule 1.6(c) requires lawyers to make reasonable efforts to prevent inadvertent or unauthorized disclosure of, or unauthorized access to, that information.
The standard is one of reasonable efforts, not a guarantee that no security incident can ever occur.
The reasonableness of safeguards depends on circumstances, including the sensitivity of the information, the likelihood of disclosure, and the availability and burden of additional protective measures.
AI systems introduce additional information flows into that analysis.
A document submitted to an AI application may be processed through infrastructure controlled by a provider.
Depending on the product and configuration, information may be retained, logged, accessed by authorized personnel, or processed through other services.
Those arrangements must be evaluated rather than assumed.
ABA Formal Opinion 512, Generative Artificial Intelligence Tools , issued July 29, 2024, addresses these concerns directly.
The opinion explains that some generative AI systems present risks of revealing information relating to a representation to unauthorized recipients.
In the circumstances described by the opinion, informed client consent is required before entering protected information into systems presenting those disclosure risks.
The opinion also recognizes that not every AI use requires informed consent.
For example, general brainstorming that does not involve information relating to a representation can present a different confidentiality analysis.
The professional obligation is therefore to understand the particular tool, the information being processed, and the safeguards governing the proposed use.
A lawyer should not assume that a product is ethically appropriate merely because it is marketed to law firms.
Nor should the lawyer assume that all third-party AI processing is necessarily prohibited.
When Does Third-Party Access Threaten Attorney-Client Privilege?
Attorney-client privilege is generally grounded in the confidentiality of communications made for obtaining or providing legal advice.
Disclosure to an unrelated third party can undermine the confidentiality necessary to maintain the privilege.
However, the involvement of a third party does not automatically defeat privilege in every circumstance.
Some third parties assist counsel in providing legal advice, and privilege doctrine has long recognized that their participation may be consistent with protected communications.
United States v. Kovel: The Third-Party Assistance Principle
In United States v. Kovel, 296 F.2d 918 (2d Cir. 1961) , the Second Circuit considered privilege in the context of an accountant assisting a law firm.
The decision recognized that communications involving a nonlawyer assistant may remain protected where the assistant’s role facilitates the provision of legal advice.
The principle is important for evaluating modern legal technology.
But Kovel does not establish that every technology vendor becomes an agent whose participation preserves privilege.
It did not involve generative AI.
Applying its reasoning to a particular AI arrangement requires careful attention to the provider’s role, the purpose of the processing, confidentiality arrangements, and governing law.
A provider performing functions integral to delivering legal services under appropriate restrictions may present a different factual situation from a consumer platform receiving information under broadly permissive usage terms.
The critical question is not whether a contract uses the word confidential.
It is whether the actual relationship and disclosure preserve the legal conditions necessary for privilege.
The 2026 AI Privilege Decisions: What Courts Have Actually Held
During 2026, federal courts began addressing disputes involving AI-generated litigation materials and claims of privilege or work-product protection.
Three decisions illustrate the significance of separating those doctrines.
The decisions involve different facts, parties, procedural settings, and legal questions.
They should not be interpreted as establishing a single categorical rule for every AI application.
United States v. Heppner: Consumer AI and the Limits of Privilege
In United States v. Heppner, No. 25-cr-00503-JSR, a represented criminal defendant independently used Anthropic’s Claude to analyze his legal situation.
The defendant entered information he had learned from counsel and generated materials discussing potential defense strategies.
His lawyers did not direct him to conduct those AI interactions.
The court rejected attorney-client privilege and work-product claims over the materials.
In the privilege analysis, the court emphasized that Claude was not the defendant’s attorney, that the materials were not attorney-client communications, and that the applicable platform arrangements undermined the claimed confidentiality.
The court also rejected work-product protection under its analysis of counsel involvement and the materials’ creation.
The opinion discussed, without deciding, whether a system used at counsel’s direction might present a different agency analysis.
The significance of Heppner is not that every interaction with generative AI necessarily waives privilege.
The case involved a represented individual who independently used a consumer AI platform, with particular privacy terms, without direction from counsel.
Those facts mattered to the court’s analysis.
The practical lesson is that clients may themselves create confidentiality and discovery risks by entering legal advice or strategy into outside AI systems.
Attorneys should consider whether appropriate client guidance is necessary, particularly during active litigation or investigations.
Warner v. Gilbarco: AI-Assisted Work Product in Pro Se Litigation
In Warner v. Gilbarco, Inc., No. 2:24-cv-12333, a self-represented litigant faced discovery requests seeking information about her use of AI tools in the litigation.
The court denied the relevant motion to compel, concluding that the information sought was not discoverable in the circumstances.
It also recognized that the litigant could assert work-product protection over qualifying AI-assisted materials.
Importantly, the court distinguished attorney-client privilege waiver from work-product waiver.
It reasoned that use of an AI tool did not necessarily constitute disclosure to an adversary or in a manner likely to place protected material in an adversary’s hands.
The court did not establish that all ChatGPT communications are attorney-client privileged.
Its analysis concerned the discoverability and work-product protection of materials associated with a pro se litigant’s preparation.
Warner illustrates why the waiver analysis cannot stop at the observation that information was entered into a computer system.
Work-product protection serves a different purpose from attorney-client privilege, and its waiver standard is not identical.
Morgan v. V2X: Work Product, Platform Identification, and Protective Orders
In Morgan v. V2X, Inc., No. 25-cv-01991-SKC-MDB, a self-represented civil litigant used AI in connection with litigation preparation.
The dispute concerned proposed modifications to a protective order and disclosure of AI platforms used with information designated confidential.
The court recognized that Federal Rule of Civil Procedure 26(b)(3) can protect a pro se litigant’s mental impressions and litigation preparation.
But it distinguished protected mental impressions from the identity of an AI platform used to process information subject to a protective order.
The court required disclosure of the platform identity in the specified circumstances.
It also modified the protective order to impose AI-specific confidentiality restrictions, including restrictions on submitting protected information to certain mainstream AI tools.
Morgan is especially important because it separates two questions:
whether AI-assisted litigation analysis may be protected work product, and whether the use of a particular platform is consistent with a court’s confidentiality order.
Those questions can produce different answers in the same case.
What the Three Decisions Establish
The decisions should not be characterized as three courts answering an identical legal question in conflicting ways.
Heppner involved a represented criminal defendant acting independently of counsel.
Warner involved a self-represented civil litigant and discovery into AI-assisted litigation preparation.
Morgan involved a self-represented civil litigant, work-product protection, and specific court-imposed confidentiality restrictions.
The common lesson is more useful than a simplistic claim that AI either preserves or destroys privilege.
Different legal protections have different elements, different purposes, and different waiver standards.
The use of AI does not eliminate those distinctions.
Legal Protection Versus Technical Control: The Question Beneath the Contract
A major weakness in many discussions of AI confidentiality is the assumption that contractual ownership of information resolves the question of protection.
It does not.
A contract may state that the law firm or client retains ownership of all uploaded documents and outputs.
The same agreement may authorize the provider to process information, retain certain records, or permit access by designated personnel and subprocessors.
Those provisions are not inherently inconsistent.
Ownership is one legal relationship. Authorized processing is another.
Technical control presents a third.
This distinction is central to my broader work on digital sovereignty and institutional intelligence.
An institution may possess contractual rights over information while depending on an outside provider for the ability to access, process, export, or delete it.
In the legal context, that dependency may affect more than operational flexibility.
It may also influence the factual circumstances relevant to confidentiality and privilege analysis.
The distinction can be understood through three questions.
Who Has Legal Rights?
Contracts determine what the parties are authorized or obligated to do.
Relevant provisions may address data ownership, confidentiality, permitted uses, deletion, disclosure, and liability.
Who Has Technical Access?
System architecture determines which people and services can actually access information during processing, administration, support, logging, or storage.
Technical access may be limited by encryption, access controls, isolation, or other protective mechanisms.
But the presence of a security feature does not, by itself, establish who controls the keys or can access information during authorized processing.
What Are the Evidentiary Consequences?
Neither ownership nor technical access alone determines privilege.
The legal analysis depends on whether the elements of the asserted protection are satisfied under governing law.
The relevant contract and system configuration may provide important facts, but they are not substitutes for that analysis.
The practical consequence is that vendor due diligence should examine contractual rights, technical capabilities, and legal protections separately.
The Lex Wire Five Layers of AI Information Control
To organize that assessment, I propose a practical model: the Five Layers of AI Information Control .
The framework is designed to help lawyers identify what information is involved, how a system handles it, and which legal protections may be affected.
It is not a separate ethical rule, a judicially adopted privilege test, or a guarantee of compliance.
A proposed framework for evaluating confidentiality, privilege, third-party access, and operational control.
Identify attorney-client communications, work product, other confidential client information, and material subject to contractual or legal restrictions.
Determine where information travels, which systems process it, and which personnel, providers, and subprocessors may access it.
Evaluate retention of documents, prompts, responses, logs, derived records, and the contractual and technical arrangements governing reuse or model improvement.
Apply Rule 1.6, privilege doctrine, work-product principles, contractual confidentiality, and relevant court orders to the actual arrangement.
Assess export, deletion, termination, continuing access, retained records, and the firm’s ability to verify relevant contractual obligations.
These five layers help identify information-control risks before confidential material enters an AI system.
They also encourage a more precise professional inquiry.
Rather than asking whether a product is generally secure, the lawyer asks whether its particular operation is appropriate for the information and legal work involved.
“Are you using AI entirely inside systems your firm controls, or are you sending client information to a third party’s platform?”
Jeff Howell, Esq. | Client Communication in the Age of AI, Original CLE Script
That question comes directly from my client communication CLE materials.
It provides a useful starting point, although local processing does not automatically establish security or privilege, and third-party processing does not automatically establish waiver.
What matters is understanding the actual arrangement.
Vendor Contracts: What Lawyers Should Examine Before Uploading Client Information
The terms governing an AI platform may be central to evaluating confidentiality risks.
Lawyers should understand more than a provider’s marketing description.
Important provisions include the following.
| Contract Provision | Principal Question |
|---|---|
| Data Rights | What rights does the provider receive to process, retain, or use inputs and outputs? |
| Model Training | May submitted information be used for model improvement or other training? |
| Retention | What records remain stored, for how long, and under which exceptions? |
| Subprocessors | Which additional entities process the information? |
| Access Controls | Who may access submitted information for support, administration, or other purposes? |
| Confidentiality | What restrictions limit use and disclosure? |
| Incident Response | When must the provider notify the firm about security incidents? |
| Termination | What happens to records and access rights after the service ends? |
These issues cannot always be resolved from the contract alone.
The actual service configuration may materially affect how the terms operate.
For example, an agreement prohibiting training on customer information does not necessarily mean the provider retains no logs or has no authorized operational access.
Likewise, encryption may reduce certain risks without eliminating all access during processing.
Formal Opinion 512 advises lawyers to read and understand the terms of use, privacy policies, and related contractual arrangements of the AI tools they use.
Technical consultation may be appropriate when the lawyer cannot independently evaluate the system’s material characteristics.
The objective is not to demand an impossible guarantee that no system can ever fail.
It is to develop a reasonable understanding of the risks and safeguards relevant to the proposed use.
Does Uploading Privileged Information to ChatGPT or Another AI Tool Waive Privilege?
It can create privilege risks, but waiver is not automatic in every circumstance.
The answer depends on the governing law, the communication’s privileged character, the purpose of the disclosure, the provider’s role, the applicable confidentiality protections, and other relevant facts.
A lawyer should reject two overly broad propositions.
The first is that every disclosure to a third-party technology provider necessarily waives privilege.
The second is that a confidentiality clause necessarily preserves privilege regardless of the arrangement.
Neither proposition adequately accounts for the governing legal doctrines.
In Heppner, the absence of attorney direction, the nature of the AI communications, and the platform’s privacy terms were significant.
In other circumstances, a provider may function under materially different contractual, operational, or agency arrangements.
The legal result may therefore differ.
Lawyers should assess those differences before submitting privileged communications to outside systems.
When the analysis identifies material risks affecting the client, appropriate consultation or informed consent may be required.
That communication framework is discussed in Client Communication in the Age of AI: Transparency, Informed Consent, and the Attorney’s Duty to Explain .
AI-Assisted Work Product and Inadvertent Disclosure
Privilege disputes are only part of the legal risk presented by AI-assisted litigation work.
Work-product protection may also apply to qualifying litigation materials prepared using AI.
Under Federal Rule of Civil Procedure 26(b)(3) , qualifying documents and tangible things prepared in anticipation of litigation or for trial receive protection subject to specified exceptions.
The doctrine also provides particular protection for attorney mental impressions, legal theories, and litigation strategy.
As Warner and Morgan demonstrate, courts may recognize work-product protection in AI-assisted litigation preparation.
That does not establish that every prompt or AI response qualifies.
The purpose and circumstances of creation remain important.
Inadvertent Production and Federal Rule of Evidence 502
A separate concern arises when AI-assisted document review incorrectly classifies materials that are later produced to opposing counsel.
For example, an AI review tool might overlook legal advice embedded within a longer business email chain.
If those materials are inadvertently produced, the lawyer may need to evaluate potential waiver, corrective action, and applicable discovery procedures.
Federal Rule of Evidence 502(b) establishes conditions under which inadvertent disclosure in a federal proceeding or to a federal office or agency does not operate as waiver.
Those conditions include reasonable steps to prevent disclosure and prompt reasonable steps to rectify the error.
Rule 502(d) also permits federal courts to issue nonwaiver orders for disclosures connected with litigation pending before them.
These protections can be important in large-scale discovery involving technology-assisted review.
But they do not create a universal safe harbor for voluntary disclosures of client information to AI vendors.
The scope of the applicable rule, court order, and circumstances must be examined.
The practical lesson is that privilege review requires appropriate verification, testing, and oversight.
A lawyer should not assume that an AI system’s classification of a document resolves its legal status.
Four AI Scenarios and Their Different Legal Risks
The following examples are hypothetical.
They illustrate differences in information handling, privilege concerns, and operational control.
Consumer AI Platform
A lawyer uploads a confidential attorney-client email into a general-purpose AI assistant.
The applicable service terms permit certain forms of provider access, retention, or reuse.
The lawyer should evaluate Rule 1.6, the relevant privilege doctrine, and the actual terms before proceeding.
The risk cannot be determined from the product name alone.
Different subscriptions, settings, and contracts may produce materially different information-handling arrangements.
Enterprise AI with Contractual Protections
A firm uses an enterprise AI application under an agreement containing confidentiality restrictions, limits on secondary use, and specified retention controls.
These safeguards may materially improve the confidentiality risk assessment.
They do not automatically establish that all communications processed through the system retain privilege.
The lawyer must still evaluate the provider’s role and the governing privilege requirements.
Locally Controlled AI
A law firm operates an AI system within infrastructure under its direct control.
Client information does not require transmission to an external model provider for the specified processing task.
The arrangement may reduce certain third-party exposure risks.
But the firm must still assess local access permissions, storage, backups, security, and authorized personnel access.
Local processing should not be confused with automatic compliance.
It is one architectural choice that may affect the risk analysis.
Client Independently Uses an AI Assistant
A client enters information learned from legal counsel into an outside AI platform to explore possible strategies.
The client then sends the AI-generated analysis to the attorney.
As Heppner illustrates, attorney-client privilege does not automatically attach to independently generated AI conversations merely because they concern legal matters or are later shared with counsel.
This final scenario deserves particular attention.
Lawyers often focus on the firm’s use of AI.
But clients may independently use AI to analyze legal advice, contracts, disputes, and litigation strategies.
Depending on the circumstances, that activity may introduce risks the attorney did not create and may not initially know about.
Client education can therefore become an important part of protecting confidential legal communications.
Who Controls the Information When the AI Engagement Ends?
Information protection does not end when a law firm finishes using a particular AI application.
Contracts may permit retention of specified records after termination.
Information may exist in backups, logs, or other records subject to documented retention policies.
A firm should understand whether it can export information, request deletion, discontinue access, and determine what exceptions apply.
These questions concern operational control rather than privilege doctrine alone.
But they are relevant to the broader obligation to protect client information throughout its lifecycle.
A firm’s practical ability to exercise control over sensitive information matters when evaluating long-term vendor relationships.
This connects directly to my analysis in Who Controls Your Law Firm’s AI Intelligence Layer? , which examines the institutional knowledge, workflows, and AI-enabled capabilities that can accumulate inside technology platforms.
The issues are related but distinct.
Privilege concerns the legal protection surrounding qualifying communications.
Institutional sovereignty concerns the broader ability to govern, maintain, transfer, and control digital assets and capabilities.
The Bottom Line: Protect the Communication, Not Merely the Data
Artificial intelligence has created new ways to process confidential client information.
But the underlying legal protections have not become interchangeable merely because the technology is new.
Ethical confidentiality under Rule 1.6, attorney-client privilege, and work-product protection each require their own analysis.
The 2026 decisions in Heppner, Warner, and Morgan reinforce the importance of that distinction.
A lawyer should not assume that a platform preserves privilege merely because the provider promises confidentiality or recognizes customer ownership of information.
Nor should the lawyer assume that every use of outside technology necessarily destroys protection.
The relevant inquiry is what information is involved, why it was created, who may access it, the legal basis for the claimed protection, and the consequences of disclosure.
“Legal ownership, contractual confidentiality, and technical control are separate forms of protection. Responsible AI adoption requires lawyers to understand how they interact.”
Jeff Howell, Esq. | Lex Wire Journal, Professional Analysis
This is the central principle behind the Five Layers of AI Information Control.
The framework encourages lawyers to move beyond general assurances of privacy and examine the actual legal and technical arrangements surrounding client information.
Attorney-client privilege exists to protect confidential legal communications, not merely the files in which those communications are stored.
Responsible AI use requires lawyers to preserve the legal protections surrounding client information while understanding who can access it, for what purposes, and under what authority.
Related Lex Wire Journal Analysis
- The Ethics of Artificial Intelligence in Legal Practice: Professional Judgment in the Age of Automation examines the lawyer’s responsibility to independently evaluate AI-assisted legal work.
- Client Communication in the Age of AI addresses transparency, informed consent, Rule 1.4, and meaningful client participation.
- Building an AI Governance Framework for Law Firms develops six pillars of operational AI accountability.
- Building Your Law Firm’s AI Playbook provides practical implementation procedures for internal AI use.
- Who Controls Your Law Firm’s AI Intelligence Layer? explores institutional knowledge, vendor dependency, and digital sovereignty.
Legal Authorities and Source Materials
Professional Responsibility
- American Bar Association, Model Rule 1.6: Confidentiality of Information .
- American Bar Association, Formal Opinion 512: Generative Artificial Intelligence Tools (July 29, 2024).
- American Bar Association, Model Rules of Professional Conduct , particularly Rules 1.1, 1.4, 1.6, 5.1, and 5.3.
Judicial Authorities
- United States v. Kovel, 296 F.2d 918 (2d Cir. 1961). Read opinion .
- United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y. Feb. 17, 2026). Read opinion .
- Warner v. Gilbarco, Inc., No. 2:24-cv-12333 (E.D. Mich. Feb. 10, 2026). Read order .
- Morgan v. V2X, Inc., No. 25-cv-01991-SKC-MDB (D. Colo. Mar. 30, 2026). Read order .
Federal Evidence and Discovery Rules
- Federal Rule of Evidence 501 , privilege in federal proceedings.
- Federal Rule of Evidence 502 , limitations on waiver of attorney-client privilege and work-product protection.
- Federal Rule of Civil Procedure 26(b)(3) , materials prepared in anticipation of litigation or for trial.
Original CLE Instructional Materials
- Jeff Howell, Esq., Ethics in the Age of AI: Presentation Outline and Script , continuing legal education materials.
- Jeff Howell, Esq., Client Communication in the Age of AI: Building Trust Through Transparency, Ethics, and Oversight , Lawline CLE instructional materials.