Tara Emory and Maura Grossman Address Protection of Attorney Prompts in “GenAI TAR”

Tara S. Emory and Maura R. Grossman have authored “GenAI Prompts in eDiscovery: Protected Work Product or Not?,” which will be published in Columbia Sci. and Tech. L. Rev., Vol. 28, No. 1 (forthcoming Dec. 2026).

This excellent and comprehensive paper addresses whether draft and final prompts written by attorneys to instruct a large language model are protected under the work product doctrine.  It relates to searches for responsive documents in eDiscovery. The conclusion is nuanced and this blog hardly does it justice.  It is well-worth reading the article in full.

The authors discuss a new[1] concept—“GenAI TAR”—which they describe as responsiveness classification based on natural-language prompts.  They describe it as typically a TAR 1.0 workflow with a large language model.

The paper defines attorney prompts as “discrete, attorney-crafted instructions.”  It suggests that these prompts share characteristics with search terms, traditional TAR processes, and document-review protocols for managed review.

However, the authors suggest that GenAI TAR is distinct from other eDiscovery search methods.  They write: “In GenAI TAR, judgment is typically embedded in the prompt itself.”  The authors explain that “effective prompts typically require iterative refinement….”  The paper states that “effective prompts require attorneys to write, test, and refine them, which serves to improve their recall and precision, improving the outcome for both sides.”  Thus, they discuss the concepts of “draft” prompts and “final” prompts.

The paper draws a marked and logical factual distinction between keyword search terms—which are unprotected—and prompts, particularly iterative prompts that contain attorney thought processes. They write: “Prompts can be more revealing than simple search terms.” In short, the paper explains that refined, iterative prompts can reveal an attorney’s thought process.  The authors view them as more akin to a protected TAR 1.0 attorney-selected “seed set,” but add that prompts may deserve greater protection than seed sets because prompts, unlike seed sets, disclose the selection rationale.

The authors suggest a general rule.  Absent agreement by the parties (cooperation and transparency are encouraged) or a demonstrated flaw in production (the general rule for discovery-on-discovery), “prompts should ordinarily not be subject to disclosure.”

The proposed general rule is subject to an exception, if discovery and disclosure is necessary to address material deficiencies in a production.  In that case, some discovery of prompts is permissible.  However, even where those deficiencies exist, the authors propose that discovery of strategic thinking, mental impressions, or case theory (opinion work product) should be generally protected.  In cases of “outright misconduct,” however, that limitation may fall.[2]

To repeat only one of the paper’s examples of a “material deficiency,” evidence of “some human error in categorization” that caused “gaps” in production would likely support discovery-on-discovery, i.e., limited discovery of prompts.  In my words, if there is a showing of a failure to produce responsive, proportionate ESI, that failure may be caused by flaws in the prompts, making at least some of those prompts discoverable.

Part of the paper analogizes principles of “discovery-on-discovery.”  Generally, satellite or culpability discovery is disfavored and permitted only upon some showing of a deficiency in an opponent’s responses.[3]  The authors suggest that search methodology generally does not fall within Fed.R.Civ.P. 26(b)(1)’s scope of “relevant to any party’s claim or defense.” They write: “Like search terms and seed sets, prompts are tools for finding responsive documents, not in themselves relevant to the claims or defenses at issue in the litigation.” [emphasis in original].  Where, however, the tools did not find the responsive documents, the manner in which they were used may become relevant.

The article discusses three types of attorney-designed prompts in the document review context: 1) final prompts; 2) draft prompts; and, 3) prompts that merely mirror an RFP.  The first are highly protected because, as the product of an iterative process, they reflect attorney thoughts and strategy.  The second deserve “stronger protection” because they show “how attorneys’ thinking evolved.” Further, the authors state: “Draft prompts, used for testing, but not for final classification, are even further removed from production and even less likely to be useful or necessary.”  The third do not reveal mental processes “and therefore deserve little or no protection.”

Because GenAI TAR is different, the authors suggest an interesting wrinkle to traditional cooperation.  Generally, the Sedona Conference Cooperation Proclamation encourages early cooperation.  Here, however, with GenAI TAR the article states: “If a party is considering whether or not to share their prompts, they may prefer to wait until the prompts are reasonably well-developed before deciding.” The article suggests that early disclosure may lead to generic prompts.

However, the authors’ general “rule” is that “transparency should ordinarily remain voluntary,” because it may have drawbacks. Those potential drawbacks are discussed in detail in the paper.

On the other side, the authors suggest that “prompt protection may benefit both the requesting and responding parties.” The article provides a detailed rationale and a hypothetical example.

Instead of disclosing prompts, the article suggests that validation using traditional ESI metrics is a better mousetrap.  While I skip over the details, the authors wrote “other quality checks, combined with sound sampling and recall metrics, may address concerns about production adequacy without exposing prompts that clearly contain work-product-protected content.”

Also, the paper suggests that, when “genuine disputes” are presented, “judges or court-appointed neutrals and special masters can resolve them efficiently by reviewing the prompts in camera….”

The article addresses tangentially the issue of discovery of prompts that are outside the attorney-review process. For more information on some of the cases that the authors cited and some related principles, please see:

• Deponent’s Use of A.I. to Answer Deposition Questions Barred; ChatGPT Was Not an Attorney (May, 2, 2026).
• “Colorado policy could shield AI from complaints regarding unauthorized practice of law” (Apr. 1, 2026).
• Important A.I. Work Product and Protective Order Decision (Mar. 31, 2026).
• Does OpenAI Provide Legal Assistance? (Mar. 11, 2026).
• Discovery Permitted About Development and Use of AI Program (Mar. 10, 2026).
• “Two Courts, Two Answers: When Does Using AI Waive Privilege?” (Mar. 3, 2026).
• A.I. Privilege, Heppner, and How Did the Court Learn About the Absence of Certain Attorney-Client Communications Between Mr. Heppner and His Attorneys? (Mar. 2, 2026).
• A.I. Documents Deemed Not Privileged (Feb. 12, 2026).
• “Against an AI Privilege” – Are Prompts Discoverable?  Is Output? (Jan. 2, 2026).

______

[1] See T. Emory, J. Pickens, and W. Louis,  “TAR 1 Reference Model: An Established Framework Unifying Traditional and GenAI Approaches to Technology- Assisted Review,” 25 Sedona Conf. J. 109 (2024)(discussing “GenAI TAR 1”).

[2] For a similar suggestion concerning the propriety and scope of general “discovery-on-discovery,” see The Hon. P. Grimm (ret.), M. Berman, et al., “Discovery about Discovery: Does the Attorney-Client Privilege Protect All Attorney-Client Communications Relating to the Preservation of Potentially Relevant Information?, 37 U. Balt. L. Rev. 413 (2008).

[3] See  Guidelines for “Discovery on Discovery” (Aug. 5, 2025); “Discovery on Discovery” Permitted (Sep. 4, 2025); “Discovery on Discovery” Denied in “It Ends With Us” Lawsuit (Sep. 5, 2025);  Limited “Discovery on Discovery” Permitted (Oct. 13, 2025); Judicial Definitions of “Hit Report” and “Families” Held Unnecessary (Oct. 13, 2025); “Discovery on Discovery” Ordered After Amazon’s Flawed Implementation of Litigation Hold (Oct. 15, 2025).