AI and the Work Product Doctrine: Privilege Risks When Lawyers Use ChatGPT and Claude
Team Sonomos · Sonomos
The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.
The attorney-client privilege and the work product doctrine are the two bedrock protections of the attorney-client relationship. Both are threatened — not destroyed, but threatened — by undisciplined AI tool use. When lawyers paste deposition transcripts, draft briefs, client memos, or litigation strategy notes into ChatGPT, Claude, or Gemini on consumer accounts, they risk waiving protection that took decades of case law to establish. This guide explains what the work product doctrine actually protects, where AI tool use creates waiver risk, and the workflow that lets lawyers benefit from AI without surrendering privilege.
The work product doctrine: what it protects
The work product doctrine, established in Hickman v. Taylor (1947) and codified in Federal Rule of Civil Procedure 26(b)(3), protects materials prepared in anticipation of litigation by or for a party or its representative. It comes in two forms:
Ordinary work product — documents and tangible things (investigation notes, interview summaries, chronologies, witness lists) that can be obtained by an opposing party only by showing substantial need and inability to obtain the equivalent without undue hardship.
Opinion work product — an attorney's mental impressions, conclusions, opinions, or legal theories. This receives near-absolute protection; courts are highly reluctant to order its disclosure regardless of need.
Work product doctrine is distinct from attorney-client privilege:
| Attorney-client privilege | Work product doctrine | |
|---|---|---|
| What it protects | Confidential communications between attorney and client | Materials prepared in anticipation of litigation |
| Who holds it | Client (and, derivatively, attorney) | Attorney or party |
| Waiver | Disclosure to third parties not in the privileged relationship | Voluntary disclosure to adversaries |
| Duration | Survives termination of the representation | Survives the litigation for which the materials were prepared (FTC v. Grolier, 1983) |
Both protections can be implicated when AI tools are used in legal practice. The analysis is different for each.
How AI tool use implicates attorney-client privilege
Attorney-client privilege protects confidential communications made for the purpose of obtaining legal advice. The privilege can be waived by voluntary disclosure to a third party outside the privileged relationship.
The question for AI tools: Is disclosing a client communication to ChatGPT, Claude, or Gemini a disclosure to a "third party" that waives privilege?
The first court to decide the question, in United States v. Heppner (S.D.N.Y. Feb. 17, 2026), held that a defendant's exchanges with a consumer AI tool were not privileged: the tool is not an attorney, the vendor's consumer privacy policy defeated any expectation of confidentiality, and privileged information he typed in was waived "just as if he had shared it with any other third party." The court left open whether the result would differ if counsel had directed the use, and no court has yet addressed an enterprise tool used under a confidentiality agreement.
Why consumer tools create waiver risk: When an attorney submits a client communication to a consumer AI tool, they are disclosing that communication to OpenAI, Anthropic, or Google under consumer terms that, by default or by the user's setting, permit the vendor to retain, review, or train on the data. The communication is no longer held in confidence within the attorney-client relationship. Courts applying privilege analysis ask whether disclosure was intended to be confidential and whether it was made in a manner consistent with maintaining confidentiality. Submitting client communications to a vendor under terms permitting data use is inconsistent with confidentiality.
Why enterprise tools reduce waiver risk: Enterprise tiers of ChatGPT (Enterprise), Claude (Claude for Work), and Microsoft 365 Copilot under Microsoft's DPA include contractual commitments that restrict the vendor's use of the data to service delivery. The vendor is acting as a processor, not a recipient with independent data rights. Whether courts will treat contracted AI vendors as the equivalent of other service providers (like legal research databases, document review platforms, or cloud storage providers) is the key open question. Ethics opinions do not decide privilege, but they analyze contracted AI vendors the way they have long analyzed cloud storage and outsourced services: the lawyer must confirm an enforceable confidentiality obligation, understand retention and training terms, and be notified of breaches or legal process. Whether courts will treat such vendors as the equivalent of other service providers for privilege purposes remains open.
Work product doctrine and AI: the three scenarios
Scenario 1: Attorney drafts a brief with AI assistance.
The attorney writes the mental impressions, conclusions, and strategy; the AI assists with drafting language. The final brief is opinion work product whether or not AI helped write it. The AI is acting as a sophisticated drafting tool. If the attorney pastes their own strategy notes into the AI to assist with drafting, those strategy notes remain work product — the attorney has not disclosed them to an adversary, only to a service provider assisting in the work. Several courts so held in 2026: in Tate Group Automotive v. Legacy Automotive Capital (Tex. Bus. Ct., June 3, 2026) the court, agreeing with the federal decisions in Warner v. Gilbarco and Morgan v. V2X that it quoted, held that a party's ChatGPT conversations were work product and that disclosure to the AI platform was not disclosure to an adversary. The Heppner court reached the opposite result on different facts, because the documents there were not prepared by or at the behest of counsel.
Risk assessment: Low, if using an enterprise AI tool with confidentiality terms. Higher risk on consumer tools — the vendor's data use rights are inconsistent with work product confidentiality.
Scenario 2: Attorney pastes deposition transcripts into AI for summarization.
Deposition transcripts are not work product themselves — they are discovery materials. But the attorney's selection of which depositions to summarize, the questions asked of the AI, and any annotations or analysis generated in response may be opinion work product. If the AI generates a summary that reflects the attorney's litigation theory, that summary may be protectable.
Risk assessment: Moderate. The transcript itself is not privileged. The derivative analysis may be. Consumer AI tools are inappropriate; enterprise tools with confidentiality terms are preferable.
Scenario 3: Attorney uploads a case strategy memorandum to AI for analysis.
A case strategy memo is paradigmatic opinion work product. If submitted to a consumer AI tool, the vendor receives the mental impressions, conclusions, and litigation theories of the attorney. Under consumer terms, the vendor typically has data rights (Anthropic's consumer plans now let the user turn training off; the default position of that setting should be checked). This is the highest-risk scenario.
Risk assessment: High. Consumer AI tools should not receive case strategy memos. Enterprise tools with strong contractual confidentiality obligations should be evaluated carefully and their terms reviewed by ethics counsel before use.
State bar ethics opinions on AI and confidentiality
Multiple state bars have issued formal or informal opinions on attorney use of AI tools. The consensus across opinions issued through 2026:
- Duty of confidentiality applies. Model Rule 1.6(c) (and its state equivalents) requires attorneys to make reasonable efforts to prevent unauthorized disclosure of, or access to, client information. Using consumer AI tools with client data may violate this duty.
- Competence duty extends to AI. Model Rule 1.1 Comment 8 requires keeping abreast of "the benefits and risks associated with relevant technology." Attorneys who use AI tools without understanding their data practices may be violating the competence rule.
- Supervision obligation. When attorneys supervise non-lawyers or subordinate lawyers using AI, they must ensure AI use complies with professional responsibility rules (Rule 5.1/5.3).
- Vendor selection matters. Opinions turn on the tool's terms — whether inputs are shared, retained or used for training — which in practice separates consumer products from contracted business tiers. ABA Formal Opinion 512 frames the test as a fact-driven risk assessment, with informed consent required before client information enters a self-learning tool.
No state bar has categorically prohibited AI use, but several (the New York City Bar in Formal Opinion 2024-5, the State Bar of California's Practical Guidance, Florida Bar Opinion 24-1, Texas Opinion 705) have emphasized that the attorney's duty of confidentiality does not pause when the tool is an AI.
The practical framework: AI in legal practice without privilege risk
Step 1: Identify what information is entering the AI context.
Before using any AI tool for legal work, classify the information: Is this client communication? Work product? Publicly available information? Information that is already in the record (filed briefs, public filings, case law)? The privilege analysis depends on what is being shared.
Step 2: Use enterprise-tier tools with confidentiality terms.
For any work involving client information or work product, use enterprise AI tools — ChatGPT Enterprise, Claude for Work (Teams/Enterprise), or Microsoft 365 Copilot on a commercial M365 plan with appropriate DPA terms. Obtain the vendor's confidentiality commitments in writing and review them with your firm's ethics counsel. For Claude in particular, Is Claude Confidential? sets out what Anthropic can see, store, and train on at each tier.
Step 3: Never paste raw client communications into consumer AI.
Emails from clients, voicemail transcriptions, intake forms, and any other direct client communication should not enter consumer AI tools. The confidentiality of these communications is the core of the attorney-client privilege, and consumer tool terms are inconsistent with maintaining that confidentiality.
Step 4: Document your AI governance decisions.
Law firms increasingly face discovery requests and bar investigations. Documenting which AI tools are approved, under what conditions, and with what confidentiality terms provides evidence that the firm took reasonable measures to protect client data. This documentation is also useful for firm-wide acceptable use policies — see the AI Acceptable Use Policy template for the seven sections an effective policy needs.
Step 5: Train attorneys on classification before submission.
The failure mode is not malice — it is habit. An attorney who uses ChatGPT to draft emails at home and then opens a client matter without switching to the firm's approved enterprise tool will make the mistake. Training, technical controls (SSO that defaults to the enterprise account), and firm policy reduce this risk.
Frequently asked questions
Does using AI to help write a brief waive work product protection?
No, as a general matter. The work product doctrine protects materials prepared in anticipation of litigation; the manner of preparation — whether by hand, dictation, or AI assistance — does not determine protectability. What matters is whether the material reflects the attorney's mental impressions and was prepared in anticipation of litigation. AI-assisted drafts can be opinion work product. In 2026 the Texas Business Court in Tate Group, agreeing with two federal district courts, held that a party's ChatGPT conversations prepared for litigation were work product and that sharing them with the platform did not waive the protection; the Heppner court denied protection on different facts because the documents were not prepared by or at the behest of counsel.
What about AI-generated summaries of discovery documents?
It depends. A summary of a deposition transcript is ordinary work product if prepared in anticipation of litigation. If the summary reveals the attorney's selection criteria, litigation focus, or legal theory, it may be opinion work product. The underlying deposition transcript is not work product — it is a discovery document. The derivative summary may be.
Can opposing counsel obtain the prompts we submitted to an AI?
Potentially. If the AI tool retains prompts and responses, those records may be discoverable if not protected by privilege or work product. In Heppner the government obtained the defendant's Claude files by search warrant and the court ruled them unprotected; in Tate Group the court reviewed the ChatGPT conversations in camera and ordered the pages that were not work product produced. Consumer AI tools retain conversation history; enterprise tools may or may not retain logs depending on configuration. This is a reason to (a) use enterprise tools where retention can be controlled and (b) structure prompts as opinion work product (incorporating the attorney's analysis) rather than neutral requests.
Does using AI for legal research affect privilege?
Legal research is generally not privileged or work product — the law is public. Using AI to synthesize case law, research statutes, or summarize regulatory guidance carries no privilege risk if no client information or attorney strategy is included in the prompt. The risk arises when legal research is combined with case-specific facts or strategy.
Our firm uses an AI legal research tool (Harvey, Westlaw AI, etc.). Different analysis?
Purpose-built legal AI tools like Harvey AI, Westlaw AI, and Lexis+ AI are designed for law firm use and typically include attorney-specific confidentiality terms in their contracts. The analysis is similar to enterprise general-purpose AI tools: review the confidentiality terms, confirm the vendor's data use restrictions, and ensure the DPA is executed before using the tool with client data. See AI Tools for Legal Research for the comparison.
The bottom line
Attorney-client privilege and the work product doctrine are not automatically waived by AI tool use — but consumer AI tools create genuine waiver risk that enterprise tools can substantially reduce. The practical answer in 2026 is to treat AI tools like any other legal technology vendor: conduct due diligence on confidentiality terms, execute appropriate agreements, prohibit consumer tool use for client work, and document your governance decisions. The lawyers who get this right will be able to use AI productively; those who ignore it risk ethics complaints and the loss of privilege, as the 2026 Heppner ruling shows.
Sources and primary references
Reviewed November 2, 2026. Regulations are cited to their official text. Vendor statements are cited to the vendor's own documentation as it read on that date; vendors revise these pages, so confirm before relying on them.
- Hickman v. Taylor, 329 U.S. 495 (1947) (Cornell LII) — origin of the work-product doctrine.
- Federal Rule of Civil Procedure 26(b)(3) (Cornell LII) — substantial-need standard for ordinary work product and protection of mental impressions.
- FTC v. Grolier Inc., 462 U.S. 19 (1983) (Cornell LII) — work-product protection survives the end of the litigation for which the materials were prepared.
- Federal Rule of Evidence 502 (Cornell LII) — scope of waiver and the reasonable-steps safe harbor for inadvertent disclosure.
- United States v. Heppner, Memorandum, No. 25 Cr. 503 (JSR) (S.D.N.Y. Feb. 17, 2026), ECF 27 (copy hosted by Law Business Research) — consumer AI exchanges not privileged or work product; privileged inputs waived; counsel-directed use left open.
- Tate Group Automotive, LLC v. Legacy Automotive Capital, LLC, Court Minute Entry Regarding ChatGPT Materials, No. 25-BC11B-0020 (Tex. Bus. Ct. June 3, 2026) (copy from re:SearchTX hosted by J.L. Ellis) — ChatGPT conversations protected as work product under Tex. R. Civ. P. 192.5(a)(1); quotes Warner v. Gilbarco and Morgan v. V2X on waiver; disagrees with Heppner.
- ABA Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024) — Rule 1.6 analysis, cloud and outsourcing analogy, Rules 5.1 and 5.3.
- New York City Bar Formal Opinion 2024-5 (Aug. 2024) — "open" versus "closed" systems; supervision and training.
- Florida Bar Ethics Opinion 24-1 (Jan. 19, 2024) — lawyers may use generative AI but must protect confidentiality; cloud-computing analogy.
- Texas Professional Ethics Committee Opinion 705 (Feb. 2025) — reasonable precautions before entering confidential information.
- D.C. Bar Ethics Opinion 388 (Apr. 2024) — vendor terms and whether inputs are visible to strangers to the attorney-client relationship.
- Microsoft Learn, Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat — Microsoft as data processor under the DPA; no training on prompts and responses.
- Anthropic Privacy Center, Is my data used for model training? — commercial inputs and outputs not used for training by default.
- OpenAI, Data controls in the OpenAI platform — API retention and zero-data-retention terms.
- Google, Gemini Apps Privacy Hub — consumer retention and human review.
Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works
This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.
Reference pages for this post
Related Articles
Best Privacy-Focused AI in 2026: Chatbots That Keep Your Data Private
The most private AI runs locally, via Ollama or LM Studio. Lumo, Brave Leo and Venice say they keep no chat logs; Duck.ai hides your IP. Compared, Nov 2026.
Can My Employer See My AI Chats? ChatGPT, Claude, Gemini and Copilot
Yes, if your employer controls the account, the computer or the network. Work accounts on ChatGPT, Claude, Gemini and Copilot let it read or keep chats. A side-by-side table of what each vendor documents, as of November 2026.
Can My Employer See My Copilot Chats? Work vs Personal Accounts
Copilot chats on a work account sit in your mailbox, where your employer can search, hold and export them with Purview. A personal account sits outside the tenant, but a work laptop can still capture it. What Microsoft documents, as of November 2026.