case-lawprivilege-discoveryFeatured

Judicial Split: Is Generative AI a 'Tool' or a 'Third Party'? How Two Courts Reached Opposite Conclusions on Work Product Protection

Decision & Law Editorial Team

March 1, 2026

12 min read

2800 words

work-product-doctrineattorney-client-privilegegenerative-aidiscoveryprofessional-responsibility

Decided

Warner v. Gilbarco / United States v. Heppner

2:24-cv-12333 / 25-cr-00503-JSR

E.D. Mich. / S.D.N.Y.

February 10–17, 2026

AI Tool: ChatGPT / Claude

Key Issue

AI as tool vs. third party — work product doctrine split

Key Takeaways for Practitioners

Two federal courts issued incompatible frameworks for AI and work product protection just seven days apart in February 2026.
Warner (E.D. Mich.): AI is a 'tool, not a person' — using ChatGPT does not waive work product protection.
Heppner (S.D.N.Y.): AI is a third party — using Claude without attorney supervision destroys confidentiality and waives protection.
The split turns on one question: Is the AI an instrument the user employs, or a recipient to whom the user discloses information?
Until a circuit court resolves the split, practitioners in the Second Circuit face materially higher waiver risk than those in the Sixth Circuit.
The safest strategy: lawyer-in-the-loop workflows + enterprise AI platforms with contractual confidentiality terms.

Two Courts. Seven Days. Opposite Outcomes.

In February 2026, two federal courts issued decisions on the same question — whether AI-generated legal work is protected from discovery — and reached diametrically opposite conclusions.

Warner v. Gilbarco, Inc., No. 2:24-cv-12333 (E.D. Mich. Feb. 10, 2026): protection granted. United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y. Feb. 17, 2026): protection denied.

The gap between these outcomes is not a matter of degree. It reflects two incompatible theories of what generative AI fundamentally is in the eyes of the law — and that theoretical divide will define AI discovery practice for years.

Warner v. Gilbarco: AI as Tool

Facts

In Warner, an employment discrimination action, a pro se plaintiff admitted to using ChatGPT to draft her legal filings and perform internal case analysis. Defendants moved to compel production of her prompts and the AI's outputs, arguing she had waived any protection by voluntarily disclosing information to ChatGPT — a third party.

The Court's Framework

Magistrate Judge Anthony P. Patti denied the motion to compel. Because the plaintiff was pro se, attorney-client privilege was unavailable. The court instead applied the work product doctrine under FRCP 26(b)(3), finding that the AI outputs constituted internal mental impressions and litigation strategy — opinion work product receiving the highest level of protection.

The court rejected the waiver argument with a functional characterization of the technology:

"ChatGPT (and other generative AI programs) are tools, not persons, even if they may have administrators somewhere in the background."

Adopting the defendant's theory, the court warned, would "nullify work-product protection in nearly every modern drafting environment."

United States v. Heppner: AI as Third Party

Facts

In Heppner, a criminal securities fraud prosecution, the defendant independently used the consumer version of Claude to query his defense strategy while under investigation. FBI agents seized his devices and found thirty-one Claude-generated documents. Defense counsel asserted attorney-client privilege and work product protection.

The Court's Framework

Judge Jed S. Rakoff granted the government's motion to access the materials. Claude is not an attorney. Anthropic's privacy policy — which permits data collection, model training, and disclosure to regulators — destroyed any reasonable expectation of confidentiality. Because the defendant used the tool without attorney supervision, the interaction constituted a disclosure to a third party, waiving both protections.

"Claude is not an attorney . . . all recognized privileges require a trusting human relationship."

The Doctrinal Split: A Direct Comparison

| Factor | Warner v. Gilbarco | U.S. v. Heppner | |---|---|---| | Court | E.D. Mich. (6th Cir.) | S.D.N.Y. (2nd Cir.) | | AI Characterization | Tool (like a word processor) | Third party (recipient of disclosure) | | Attorney Supervision | None (pro se) | None (client acted alone) | | Privacy Policy Role | Not analyzed | Dispositive | | Outcome | Protection granted | Protection denied |

The tension is not in the outcomes — it is in the reasoning. The courts adopted incompatible frameworks for how AI tools relate to privilege. Until the Second or Sixth Circuit resolves the split, practitioners face genuine jurisdictional uncertainty.

ISSUE: Which Framework Should Govern?

Rule

The work product doctrine protects "documents and tangible things that are prepared in anticipation of litigation or for trial by or for another party or its representative." FRCP 26(b)(3). Opinion work product — reflecting "mental impressions, conclusions, opinions, or legal theories" — receives near-absolute protection. Hickman v. Taylor, 329 U.S. 495, 510–11 (1947). Waiver occurs when protected materials are disclosed to a third party in a manner inconsistent with maintaining confidentiality.

Application

The tool framework (Warner) is grounded in function: what the user does with AI, not what AI is. Under this view, AI prompts are analogous to notes in a legal pad — the fact that a third-party company manufactured the pad does not make the notes discoverable. The key question is whether the work reflects litigation strategy, not whether it was generated by a human or machine.

The third-party framework (Heppner) is grounded in structure: who receives the information. Under this view, the relevant fact is that the user transmitted information to a commercial platform operating under terms that permit government disclosure. The attorney-client privilege and work product doctrine were designed to protect confidential relationships — and a commercial AI platform with a permissive privacy policy cannot anchor such a relationship.

The Kovel doctrine — which extends privilege to non-attorney agents assisting counsel — points toward a potential synthesis: AI used under attorney direction and supervision may retain protection even under the Heppner framework, because the key defect in Heppner was the absence of attorney oversight, not the use of AI per se.

Conclusion

Neither framework is clearly correct as a matter of existing doctrine. Warner's functional approach better serves the policies underlying work product protection. Heppner's structural approach better maps onto existing confidentiality doctrine. The resolution will likely turn on whether appellate courts analogize AI to software tools (favoring Warner) or to third-party consultants (favoring Heppner).

Practical Guidance Until the Split Is Resolved

1. Maintain a lawyer-in-the-loop for all litigation AI use. Even under Heppner, attorney supervision is the key variable. Document that AI was used at counsel's direction, in anticipation of litigation.

2. Use enterprise AI platforms with contractual confidentiality. Avoid consumer-facing tools with permissive privacy policies. Enterprise configurations that prohibit training on user data and include enforceable confidentiality terms create the reasonable expectation Heppner found missing.

3. Preserve AI work product under litigation holds. Update preservation protocols to explicitly address AI prompts and outputs. FRCP 37(e) spoliation exposure applies.

4. Update retainer agreements. Authorize attorney-directed AI use expressly. Obtain informed client consent regarding the risks of independent AI usage — the Heppner scenario (client using AI without counsel) is the highest-risk pattern.

5. In the Second Circuit: assume the strictest standard. Until the circuit resolves the split, practitioners in S.D.N.Y. and other Second Circuit courts face materially higher waiver risk. Treat all consumer AI interactions as presumptively discoverable.

What Comes Next

No Supreme Court decision addresses this issue. The Second and Sixth Circuits may ultimately define the rule. Practitioners should monitor AI-related privilege logs and discovery disputes in both circuits closely through 2026.

The EU AI Act's data governance requirements may also influence U.S. courts as they develop frameworks for AI-generated evidence — an irony not lost on practitioners navigating the Brussels Effect from American courtrooms.