AI, Legal Privilege, and the Evolving Landscape of Discovery

The integration of generative AI (GenAI) tools into legal and business workflows is rapidly accelerating, prompting courts to grapple with the application of long-standing legal principles – attorney-client privilege and the work-product doctrine – to AI-generated data. This includes prompts, outputs, and activity logs. While recent rulings confirm that GenAI data is discoverable as electronically stored information (ESI), a critical question remains: when is this data protected from disclosure? The answer, according to emerging case law, lies in how, why, by whom, and under what conditions these tools are utilized.

Attorney-Client Privilege in the Age of AI

Attorney-client privilege safeguards confidential communications between lawyers and their clients made for the purpose of seeking or providing legal advice. While, GenAI systems themselves are not lawyers or clients. Communications with AI tools are not automatically privileged, even if the subject matter is legal in nature. Privilege may apply when GenAI is used under the direct supervision of counsel to facilitate legal advice, functioning similarly to a non-lawyer assistant – but only if a reasonable expectation of confidentiality exists and is maintained.

The Heppner Ruling and Expectations of Confidentiality

The case of United States v. Heppner (No. 25-cr-00503-JSR ECF 27, S.D.N.Y. Feb. 17, 2026) provides crucial insight. In Heppner, the defendant used a publicly available GenAI tool, entering both factual and legal prompts to assess potential legal exposure. He then shared the AI-generated analyses with his defense counsel. When federal agents seized his computer and the ESI it contained, the government sought to compel production of the AI-generated content.

Judge Jed S. Rakoff ruled that the AI-generated content was not protected by either attorney-client privilege or the work-product doctrine, emphasizing three key points:

• The GenAI platform was a third-party tool with no expectation of confidentiality.
• The materials were not created at counsel’s direction and, weren’t created to facilitate legal advice.
• Sharing AI-generated content with an attorney after the fact did not retroactively confer privilege or protection.

This decision reinforces that attorney-client privilege applies only to confidential communications between a lawyer and client for the purpose of legal advice – it doesn’t extend to documents that later grow useful to counsel.

The Work-Product Doctrine and GenAI

The work-product doctrine protects materials prepared by or at the direction of counsel in anticipation of litigation, offering heightened protection for materials reflecting counsel’s mental impressions, conclusions, or legal strategy.

Courts are beginning to differentiate between:

• GenAI data created at counsel’s direction to analyze claims, defenses, or litigation strategy, which may qualify as work product.
• GenAI data created independently for business or exploratory purposes, which generally does not.

In Heppner, the court denied work-product protection because the AI-generated materials weren’t prepared at counsel’s direction and didn’t reflect defense counsel’s strategy. The ruling clarifies that GenAI data isn’t automatically work product simply because it addresses legal issues.

Contrasting Outcomes: Tremblay v. OpenAI

The case of Tremblay v. OpenAI, Inc. (No. 23-cv-03223-AMO, N.D. Cal. Aug. 8, 2024) presented a different scenario. Plaintiffs alleging copyright infringement conducted targeted pre-suit testing of ChatGPT to evaluate potential claims. They produced the prompts used in their complaint but refused to produce additional prompts and outputs, arguing they revealed counsel’s mental impressions and litigation strategy.

The court partially agreed, holding that unused prompts, account data, and testing results constituted opinion work product prepared in anticipation of litigation. Importantly, the court limited waiver to the specific prompts and outputs affirmatively relied upon in the pleadings, preventing a blanket waiver of all related materials.

Privilege Waiver Considerations

Both Heppner and Tremblay primarily address whether privilege or work-product protection initially applies. Equally crucial is the risk of waiver. Confidentiality must be maintained to preserve such protection. Loading sensitive data into GenAI tools that retain data, reuse it, or use it for training significantly increases the risk of waiver.

Practical Steps to Preserve Privilege and Work-Product Protection

• Use Secure GenAI Tools: Opt for closed, enterprise platforms with terms of service that restrict the service provider’s ability to store, review, or use user inputs for training purposes.
• Supervise and Document AI Use: Treat GenAI as a supervised assistant. Prompts and outputs should be generated under counsel’s direction and reviewed by counsel.
• Limit and Label: Avoid including privileged information in prompts and clearly label protected materials as privileged or work product protected (though labeling alone isn’t definitive).
• Remember the Metadata: GenAI activity logs and metadata can reveal litigation strategy and raise work-product concerns.
• Consider Nonwaiver Agreements: Address GenAI data in ESI agreements and seek Rule 502(d) orders to mitigate waiver risk.
• Prevent Privilege Challenges: Privilege logs should detail the creation of AI-generated data, including how, by whom, and under what confidentiality controls.

Looking Ahead

As Heppner demonstrates, courts are applying established discovery doctrines to these new technologies. Privilege disputes involving GenAI data will hinge on supervision, purpose, and reasonable expectations of confidentiality. Litigators should proactively address these issues, collaborate with e-discovery and information-governance teams, and advise clients that casual or unsupervised GenAI use can generate discoverable – and unprotected – material.