In February, a judge in the Southern District of New York issued an opinion ruling in United States v. Heppner that a criminal defendant’s exchanges with a publicly available generative AI platform were not protected by the attorney-client privilege or the work product doctrine. The reasoning was simple: a publicly available AI platform is neither confidential nor a lawyer, and thus, there is no reasonable expectation of privacy. In addition, the criminal defendant’s counsel had not asked the client to prepare litigation documents, so the work product doctrine was not applicable. The Heppner decision has been followed by a series of federal rulings that cumulatively provide clarity on the issue of discoverability of AI-generated materials.
Particularly for non-lawyers, it is critical to understand these rulings and create guardrails around AI use in the workplace.
Different Results in the Civil Context
Within weeks of Heppner, two federal courts reached more protective conclusions in civil matters involving individuals who are representing themselves, formally known as pro se litigants. In Warner v. Gilbarco, Inc., the Eastern District of Michigan held that AI-generated materials a pro se plaintiff prepared in anticipation of employment-discrimination litigation were entitled to work product protection. The court characterized generative AI as a tool, not a person, and rejected the argument that using such a tool waived work product protection, reasoning that a work product waiver requires disclosure to an adversary or in a manner likely to result in disclosure to an adversary.
The District of Colorado adopted similar reasoning in Morgan v. V2X, Inc., holding that AI interactions do not automatically compromise work product protection because the materials are unlikely to reach an adversary unless there is a discovery process compelling production. The court further observed that there is no support for making work product protection conditional on attorney involvement in the text of Federal Rule of Civil Procedure 26(b)(3), which extends to materials prepared “by or for another party or its representative” rather than only by counsel.
Read together, Heppner, Warner and Morgan suggest that the question of whether AI use defeats privilege or work product turns on the same factors that have always controlled the analysis: third-party disclosure, expectation of confidentiality, attorney direction and likelihood of adversarial exposure.
However, Warner and Morgan should not be read as broad endorsements of unrestricted AI use. Morgan also issued what appears to be the first reported AI-specific protective order, requiring the plaintiff to disclose the identity of any AI platform used in connection with confidential discovery materials and barring the parties from inputting confidential information into commonly available consumer AI tools, including standard versions of ChatGPT, Claude and Gemini. In other words, the identity of the AI tool itself may now be treated as discoverable in connection with potential protective-order violations.
Discoverability of AI-Prepared Materials
It is important to consider litigation implications beyond privilege when using generative AI to conduct business. Courts have treated AI prompts and outputs as electronically stored information. Practically speaking, means that if it is relevant, it is discoverable in a future litigation. As such, it is important to ask the following questions: Why do I want this document to exist? Did I verify its accuracy? Did I save it in such a way that ensures the label accurately reflects its origin?
Consider the following scenario. A non-lawyer employee decides to prepare a personal timeline of events using communications from their email inbox in order to document a difficult negotiation with a counterparty. The timeline tracks various changes that were made to the applicable agreement, and each party’s stated reasoning for requesting such changes. Fast forward two years, and the parties are in the middle of a litigation concerning the meaning of a certain contract provision. That document has been produced in litigation, as it is both responsive to a document request by the opposing party and relevant to the parties’ dispute. As it turns out, the timeline created by the AI tool was not properly vetted and reflects inaccurate information. Such scenarios have litigators acutely concerned about prolific AI-use without guardrails.
Practical Guidance for Organizations
For business leaders, it is best to assume that anything typed into a generative AI tool could one day appear in a court filing or a regulator’s file. Internalized across an organization, that assumption should drive most of the right behaviors without requiring a working knowledge of privilege doctrine. Beyond that, there are three principles to consider:
- Choose your organization’s tools deliberately. Enterprise platforms whose contracts prohibit training on inputs and restrict third-party access offer materially better protection than free consumer versions. The Morgan court treated that distinction as legally consequential and sophisticated users should treat it as operationally consequential well before any litigation is on the horizon.
- Treat prompts and outputs as business records. Write prompts as though they may someday be read aloud. Verify outputs before acting on them and save files with names, locations and metadata that honestly reflects how the document was produced. These do not necessarily need to be complex and can be as simple as naming conventions, or a policy that a disclaimer must be added to documents generated with artificial intelligence.
- Know when to involve counsel. The moment a project touches actual or foreseeable litigation, a regulatory inquiry, an internal investigation or a legal issue, the calculus shifts. Privilege and work product protection can often be preserved when AI is used at counsel's direction with appropriate labeling and access controls—but only if that structure is built in at the outset, not after a subpoena arrives.
Ultimately, Heppner, Warner and Morgan do not establish new law so much as confirm that familiar doctrines apply to new technology, and they are being applied in largely predictable ways. The companies that find themselves in trouble over AI in the coming years will not be those that adopted it—they will be those that adopted it casually. A modest amount of discipline at the start will buy substantial protection later.