A companion article — “Attorney-Client Privilege and Chatbot Usage” — addresses whether conversations with AI tools like Claude®, ChatGPT®, or Gemini® can be protected by attorney-client privilege. This article addresses a related but distinct question: whether AI-assisted legal work can be protected under the work product doctrine. The two doctrines are often discussed together, but they have different elements, different strengths, and different implications for AI usage. For professionals who use AI tools in legally sensitive contexts, both questions matter — and the work product analysis may ultimately be the more favorable one.
What the Work Product Doctrine Protects
The work product doctrine, established by the Supreme Court in Hickman v. Taylor, 329 U.S. 495 (1947), and codified in Federal Rule of Civil Procedure 26(b)(3), protects materials prepared by or for an attorney in anticipation of litigation. Its purpose is to preserve the integrity of the adversarial process by ensuring that attorneys can work with a zone of privacy — developing legal strategy, analyzing weaknesses, and preparing for trial without fear that their mental processes will be handed directly to the opposing side.
The doctrine covers two categories of material. Ordinary work product — factual research, document compilations, and similar materials — can be overcome by an opposing party that demonstrates substantial need and inability to obtain the equivalent by other means. Opinion work product — an attorney’s mental impressions, conclusions, opinions, and legal theories — receives near-absolute protection under Upjohn Co. v. United States, 449 U.S. 383 (1981), and its progeny. That near-absolute protection is not easily displaced, regardless of how or where the opinion work product was created.
How Heppner and Gilbarco Define the Current Landscape
Two decisions issued in February 2026 — one denying work product protection, one upholding it — together define the current state of the law on AI tools and work product. Both applied identical legal principles to different facts and reached different outcomes. Taken together, they confirm that AI tools have not changed the underlying doctrine; they have only added new factual variables to the existing analysis.
In United States v. Heppner, No. 25-cr-503 (S.D.N.Y. Feb. 17, 2026), Judge Rakoff rejected work product protection for AI-generated materials on the narrowest possible ground: Heppner — a non-attorney — had prepared them entirely on his own initiative, without any direction from his counsel. His lawyers had not asked him to use Claude®. They had not supervised the process. They had not reviewed the prompts. The documents were Heppner’s independent project, and the mere fact that he later shared them with his attorneys did not transform them into attorney work product.
That holding is factually specific in a way that severely limits its reach. The “direction of counsel” requirement — the element Heppner turned on — is satisfied automatically when the attorney is the one using the AI tool. When a lawyer directly uses Claude®, ChatGPT®, or a comparable platform to research legal theories, draft litigation strategy memoranda, analyze case weaknesses, or prepare for depositions, there is no question about direction of counsel. The attorney is counsel. The work product doctrine’s foundational requirement is met from the first prompt.
In Warner v. Gilbarco, Inc. (E.D. Mich., Feb. 10, 2026), decided one week before Heppner’s written opinion, a pro se plaintiff used generative AI tools to prepare materials related to her employment discrimination litigation. The defendants argued that the materials lost work product protection because the plaintiff had disclosed them to the AI provider. The court rejected that argument directly: AI platforms are “tools, not persons,” and disclosure to a tool is not disclosure to an adversary. Work product waiver requires disclosure that is inconsistent with maintaining secrecy from the opposing party — and using software to process your own thoughts does not meet that bar.
A Perkins Coie analysis of these companion decisions put it well: “The arrival of generative AI has not changed the law governing privilege and work product protections. Both courts applied long-established principles to a new technology, and the differing outcomes flow from the differing facts.”
The Opinion Work Product Argument: AI Prompts as Protected Mental Impressions
The most powerful work product argument in the AI context is one that Heppner did not reach and that has received insufficient attention in the commentary that has followed: the prompts that attorneys type into AI tools are themselves strong candidates for opinion work product protection.
Consider what an attorney’s prompts actually reveal. When a litigator asks an AI tool to analyze the weaknesses in a particular legal theory, or to identify the strongest counterarguments to a proposed claim construction, or to compare the facts of a case to a line of precedent, the prompt discloses which issues the attorney has identified as important, which theories the attorney is developing or testing, which weaknesses the attorney has recognized, and how the attorney is framing the legal landscape. These are precisely the mental impressions and legal theories that opinion work product doctrine is designed to protect.
The outputs — the AI’s responses — are a more complex question, and their protection will depend on the extent to which they reflect and are shaped by the attorney’s direction and judgment. But the inputs, the prompts, are the attorney’s work in a straightforward sense, and courts have long recognized that materials reflecting attorney thought processes receive the strongest protection the doctrine affords. This argument is available regardless of subscription tier and regardless of the platform’s terms of service. Opinion work product protection attaches to the attorney’s mental impressions — it does not depend on where those impressions were recorded or what service processed them.
The Confidentiality Threshold for Work Product Is Lower Than for Privilege
One of the most important distinctions between attorney-client privilege and the work product doctrine — and one that is particularly significant in the AI context — is that work product protection has a lower confidentiality threshold.
Attorney-client privilege requires that the communication itself be made in confidence, and courts closely examine whether the circumstances of the communication were consistent with an expectation of privacy. Work product doctrine, by contrast, is not destroyed by disclosure to third parties generally. It is only defeated by disclosure that is “inconsistent with maintaining secrecy from the opposing party” — the adversary in litigation. Gilbarco made this point explicitly: sharing materials with an AI tool is not sharing them with the opposing party.
This distinction matters for AI usage because the confidentiality concerns that animated Heppner’s attorney-client privilege analysis — the AI platform’s rights to use content for training, to disclose to governmental authorities — bear on whether the communication was confidential in the attorney-client privilege sense. They have considerably less force in the work product analysis, where the relevant question is not whether the platform might access the content, but whether sharing the content with the platform was inconsistent with keeping it from the adversary.
The Cloud Services Parallel
The work product implications of AI tool usage cannot be analyzed in isolation from the broader landscape of cloud-based professional services that attorneys routinely use without anyone suggesting that work product is forfeited.
Attorneys use cloud-based legal research platforms that receive their queries and return analytical results. They use e-discovery platforms that process privileged documents and return relevance rankings and privilege log recommendations. They use document management systems hosted on third-party servers. They draft briefs in Google Docs® and Microsoft Word Online®, storing their work product on servers they do not control, under terms of service they have largely not read. None of this has been held to forfeit work product protection.
AI chatbots used for legal research and analysis occupy the same functional category — tools that receive attorney input, perform computation, and return results that inform legal strategy. The work product doctrine has never required attorneys to use only tools they personally own or control. It has required that materials be prepared in anticipation of litigation at the direction of counsel. Those requirements are satisfied by how the attorney uses the tool, not by the identity of the tool’s manufacturer or the terms of its hosting agreement.
What This Means Across Different User Types
The work product question is most straightforwardly favorable for attorneys who directly use AI tools in their legal practice — the doctrine’s requirements are met by the attorney’s own use, without the complications that arise when non-attorneys use AI tools independently.
For non-attorney users — executives, inventors, in-house business staff — work product protection is available but requires more careful structuring. Materials prepared by non-attorneys can qualify as work product when prepared at the direction of counsel and in anticipation of litigation. The key is the relationship between the non-attorney user and supervising counsel at the time the AI-assisted work is done, not after the fact. The retroactive-sharing problem that defeated Heppner’s claim — preparing documents independently and later giving them to lawyers — is avoidable with appropriate planning.
For inventors working with patent counsel, the work product question intersects with patent prosecution privilege issues that are themselves unsettled. The scope of protection available for AI-assisted invention disclosures, claim drafting assistance, and prosecution strategy work is a question that deserves careful analysis tailored to the specific facts. The third article in this series addresses those patent-specific issues in depth.
The Developing Law
Heppner is a first word, not a final one. The work product implications of AI tool usage have barely been litigated, and Gilbarco’s “tool, not a person” framing offers courts a path to more balanced outcomes that Heppner’s reasoning alone did not take. The courts will need to address questions that neither opinion resolved: how attorney prompts are to be classified; whether the confidentiality analysis differs between the two doctrines, as it should; and how the supervision and direction requirement applies to AI-assisted legal work in practice.
What is clear is that the work product doctrine — particularly its opinion work product component — provides a more robust framework for protecting AI-assisted legal work than attorney-client privilege does in many scenarios. The doctrine’s requirements map more naturally onto how attorneys actually use AI tools, and its confidentiality threshold is more readily satisfied.
The Questions You Should Be Asking
Whether you are a litigator who uses AI to research case theory, a patent attorney who uses AI to assist with prosecution strategy, a corporate counsel who uses AI to analyze regulatory exposure, or an executive whose lawyers are using AI in matters affecting your business, the work product implications of those AI-assisted workflows are worth understanding before a dispute arises.
The relevant questions include: Who is using the AI tool — an attorney, or a non-attorney? Is that use occurring under attorney direction and supervision? What is the purpose of the AI-assisted work, and is it fairly characterized as being in anticipation of litigation or legal proceedings? What are the terms governing the platform being used, and do those terms affect the confidentiality analysis under the specific doctrine being invoked?
John Ogilvie of Ogilvie Law Firm advises clients on intellectual property matters and the legal risks associated with emerging technology, including the privilege and confidentiality implications of AI usage in professional settings. If you have questions about how the work product doctrine applies to your AI-assisted legal workflows, use the Contact form to get in touch.