Turning Chats Into Trial Exhibits: Litigation Risks of Generative AI Use

By Jackson Moore, Andrew Atkins and Carina Arellano

Could your personal interactions with ChatGPT, Gemini, Claude, and other generative AI tools be discovered in litigation? Could a court force production of these communications against your wishes?

Similar to risks we discussed in our alert, "The Silent Guest in Your Meetings: Legal Risks of AI "Note-Takers," chats with publicly available generative AI tools may result in involuntary disclosure of sensitive information and the waiver of privilege protections. Courts are struggling to apply the ancient attorney-client privilege and the younger work product doctrine to these new technologies.

AI Interactions May Become Trial Evidence 

Generative AI tools offer users fast, plain-English explanations of complex issues and can summarize vast amounts of information. At the same time, parties in litigation may use AI tools at their peril. As we recently explained, generative AI tools are language prediction machines; they are not truth machines, they do not consult legal libraries, they do not ‘think’ as a human does, and their answers to legal questions can be convincing—and convincingly wrong: one database has tracked 1,277 legal decisions where generative AI had produced hallucinations, including fabricated case citations, fake quotations, and misstated arguments. Even generative AI tools marketed by established legal technology providers have answered legal queries inaccurately and hallucinated at a high rate.

In addition, a user’s exchanges with a generative AI tool can become powerful evidence in a later lawsuit or trial. Users often interact with these tools using an iterative, question-and-answer format. If not approached thoughtfully, this dialogue can become an exhibit used by an opposing party to demonstrate the user’s motivations, state of mind, and intentions behind certain actions.

Claude is Not Your Lawyer: Generative AI Use and the Attorney-Client Privilege

The attorney‑client privilege protects legal advice and related confidential communications between an attorney and a client. This privilege is narrowly construed by courts and often waived when confidential information is disclosed to third parties.

Parties dialoguing with a generative AI tool, particularly an open-source generative AI tool, may be forced to produce their interactions in a later lawsuit. In general, an open-source generative AI tool may retain user prompts, store summaries of personal interactions, and use that data to train future models, thereby compromising a user’s expectation of confidentiality.

Moreover, sharing legal advice or other confidential communications with a generative AI tool risks waiving the attorney-client privilege that might attach to such material. This is especially true if it is done using an open-source model and/or in the absence of an attorney's direction.

Recently, a federal court required production of a party’s interactions with Claude. In United States v. Heppner, 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026), Judge Jed S. Rakoff of the Southern District of New York required the production of a criminal defendant’s dialogue with Claude, a generative AI tool owned by Anthropic. The defendant had asked Claude to outline a potential defense strategy after the defendant had received a grand jury subpoena. He performed these searches without any input or direction from his attorney.

The defendant argued that he had used generative AI tools to prepare for his meeting with his counsel and had shared with Claude information that he had learned from his lawyer. The Heppner Court found that no attorney-client privilege attached to these documents and ordered their production.

The Court made three findings. First, the defendant’s interactions with Claude were not attorney-client communications because Claude was not an attorney. The Court rejected the suggestion that generative AI interactions were similar to cloud-based applications such as word processors. The Court emphasized that the attorney-client privilege requires a "trusting human relationship" with a licensed professional. "No such relationship exists, or could exist, between an AI user and a platform such as Claude."

Second, Judge Rakoff found the communications with Claude were not confidential. The Claude interactions were not akin to notes a client prepared on his own but instead were communications with a non-lawyer, third-party platform. The Court cited to Claude’s terms and conditions; these required a user to consent to Anthropic’s use of a user’s "inputs" and Claude’s "outputs" to train Claude, and to disclose such data to third parties, including governmental regulatory authorities. In short, Claude was an open-source platform, and the user did not change privacy settings that might have better protected the information exchanged. The Court concluded that the defendant had no reasonable expectation that his interactions with Claude would be kept confidential under these facts.  

Third, the Court found that the defendant did not use generative AI to obtain legal advice because he did not interact with Claude at his attorney's suggestion or direction. The Court admitted that this factor presented a "closer call" because the generative AI tool may have been consulted to help the client prepare to meet with his counsel. However, because the defendant asked Claude questions on his own, and with no guidance or direction from his attorney, the defendant could not analogize Claude to a "highly trained professional who may ask as the lawyer’s agent within the protection of the attorney-client privilege."  

The Work Product Protection / Distinguishing Heppner

The Heppner Court also refused to protect the defendant’s generative AI interactions as "work product."  The work product doctrine can protect documents and tangible things prepared in anticipation of litigation, either by a party or the party’s counsel. See Hickman v. Taylor, 329 U.S. 495 (1947); Fed. R. Civ. P. 26(b)(3)(A). The Heppner Court rejected the defendant’s work product arguments because the generative AI searches were conducted "on his own volition," and the resulting output by Claude were not "prepared by or at the behest of counsel."  

Judge Rakoff acknowledged that non-lawyer generated documents might qualify as "work product" but said the doctrine was intended to protect lawyer mental processes, and that a client’s independent queries of a generative AI tool, by definition, were not lawyer mental processes. Judge Rakoff reached this conclusion, even though the defendant apparently included his lawyer’s legal advice as part of his generative AI interactions.

Later Cases Might Distinguish Heppner 

The Southern District of New York is an influential court, and Heppner offers an early analysis of the discoverability of a party’s generative AI-usage. Parties and lawyers must read and understand this case for these reasons alone.

At the same time, Heppner might approach privileges and protections too narrowly. Are personal notes discoverable unless they are written at counsel’s direction or instructions? Are a litigant’s questions, prepared before an initial lawyer meeting, discoverable? Would Heppner be decided differently had Claude’s privacy policies shielded the defendant’s inputs and Claude’s outputs from Anthropic? How do Claude’s privacy policies compare to other agreements governing other cloud-based technology (such as Google Mail)?

Two federal civil cases decided after Heppner reached contrary results and found a pro se party’s generative AI searches to be work product. In Warner v. Gilbarco, 2026 WL 373043 (E.D. Mich., Feb. 10, 2026) (Patti, Mag. J.), a pro se plaintiff successfully defended a motion to compel the production of her usage of ChatGPT in connection with her lawsuit.

The Court found her searches to be work product because they were prepared in anticipation of litigation or trial under Fed. R. Civ. P. 26(b)(3)(A), a civil procedure rule not applicable to Heppner, a criminal case. The Court found the plaintiff’s use of ChatGPT did not waive the work-product protection because waiver requires disclosure of work product to an adversary, which ChatGPT was not. The Court also declared ChatGPT to be a tool, not a person as Heppner had analogized.

The second decision, Morgan v. V2X, 2026 WL 864223 (D. Colo. Mar. 30, 2026) (Braswell, Mag . J.), was also an employment discrimination and wrongful termination case between a pro se plaintiff and his former employer. The Court protected the pro se plaintiff’s use of AI tools as work product.

The Morgan Court found Heppner distinguishable because Heppner was a criminal matter, while the Morgan case was governed by the express work product protections set forth in Rule 26(b)(3) of the Rules of Civil Procedure. The Morgan Court also found no "gap" between the pro se litigant’s use of generative AI tools and the attorney because "[a] pro se litigant is simultaneously the party and the advocate."

Finally, the Morgan Court analogized generative AI tools to other third-party systems (computers, phones, internet devices, etc.) that process otherwise protected information. The Court asked: "Does this mean that anyone with a Gmail account has forfeited all rights to confidentiality and privacy?" The Court concluded it did not.

The Morgan Court also entered a protective order to address the use of generative AI tools in the lawsuit. The Court prohibited the parties from submitting confidential information to generative AI tools unless those tools were contractually prohibited from:

1. storing or using inputs to train or improve its model; and
2. disclosing inputs to any third party except where such disclosure is essential to facilitating delivery of the service. Where disclosure to a third party is essential to service delivery, any such third party shall be bound by obligations no less protective than those required by this Order. In addition, the AI provider must contractually afford the party or authorized recipient the ability to remove or delete all Confidential Information upon request. A party intending to use AI that it contends meets these requirements must retain written documentation of these contractual protections.

The Court recognized that these restrictions would disadvantage a pro se plaintiff who lacked ready access to more expensive, more protective, enterprise-tier generative AI accounts. The Court noted that low-cost generative AI tools could be used in ways that did not involve uploading confidential information, and "the Court cannot ignore the real risks associated with mainstream tools that persistently collect and store data and could compromise confidentiality. . . . What this Order requires is that Confidential Information not be entrusted to platforms that lack the contractual safeguards described above, regardless of the sophistication or apparent trustworthiness of the tool."          

While Heppner may have applied the attorney-client privilege and work product doctrine narrowly, arguably Warner and Morgan may have applied the doctrines too broadly. We expect courts will refine these emerging issues in future decisions, as they better understand generative AI technology, how AI tools work, and how they are governed contractually.

Conclusion and Takeaways

These early federal decisions provide conflicting guidance on whether a party’s generative AI use can be protected from disclosure in a later lawsuit. There is no consensus regarding how federal and state courts will treat these tools. Courts that choose to follow Heppner may offer no protection at all to parties who interact with generative AI tools on their own.

For parties who want to use these tools, consider the following:

• Assume your generative AI prompts and the resulting output will be discoverable in a later lawsuit.
• Avoid uploading confidential data to open-source generative AI platforms. Instead, ask hypotheticals and use anonymized facts.
• Read and understand your generative AI contracts. Some platforms will disclose inputs and outputs while other platforms will contractually commit to not use customer data for training or otherwise.
• Consider Involving your attorney. In Heppner, the client’s work product claims failed, in part, because he acted alone in using generative AI tools. Documented attorney supervision and direction over a client’s generative AI use can strengthen work product arguments and limit overuse.
• Generative AI tools are here to stay. Individuals and companies need to thoughtfully approach how they use these tools to avoid the perils and pitfalls which may accompany their imprudent use.

As courts continue to address how traditional privilege and work product principles apply to generative AI, the safest course is to assume that prompts, outputs, and related exchanges may later be scrutinized in discovery. Until the law becomes more settled, parties and counsel should approach AI use with the same care they would apply to any potentially discoverable communication or litigation-related document.

If you have questions about these developments, please contact Jackson Moore, Andrew Atkins, Carina Arellano or your regular Smith Anderson attorney.