GenAI Prompts in eDiscovery: Are They Protected Work Product? eDiscovery Trends

Two Experts Say GenAI Prompts Deserve Work-Product Protection — No Court Has Agreed Yet

Tara Emory and Maura R. Grossman have published a paper arguing that prompts attorneys write to run GenAI-based document review should ordinarily be shielded from discovery as opinion work product. eDiscovery Today, via Doug Austin, first reported it in 2026. The authors' position: absent party agreement or a "demonstrated deficiency in the production," prompts reflecting strategic judgment or case theory shouldn't be subject to compelled disclosure. They lean on two early AI-litigation rulings, Tremblay v. OpenAI and Concord Music Group v. Anthropic, as signs of a judicial trend, and walk through a hypothetical matter, Regional Health Systems v. Medi Staff, to show how a prompt evolves from generic to fact-based to strategic — for instance, once witness interviews surface a named individual such as "Marcus Chen," instructing the model to flag language showing support for that person.

Are GenAI review prompts actually protected work product?

Not as settled law — two respected authors make the case, pointing to early AI rulings, but no eDiscovery court has squarely decided this question.

Emory and Grossman's argument rests on a distinction: search terms and general descriptions of methodology have long been treated as discoverable, but prompts, they say, "can be more revealing than simple search terms" because they can encode facts learned through investigation or an attorney's theory of the case. That's a persuasive framing, but it's an advocacy paper, not a ruling — and the cases it cites arose out of AI copyright disputes, not fights over discoverability of review prompts themselves.

Why should buyers care about the old "TAR tax" comparison?

Because forced disclosure once made parties avoid technology-assisted review altogether, and the authors warn GenAI could see the same chilling effect.

The "TAR tax" was the burden of exhaustive methodology disclosure that made some litigants skip TAR rather than defend it in discovery fights. If courts start compelling prompt disclosure by default, the paper argues, parties may either avoid GenAI review or write deliberately generic prompts to stay safe — which the authors say would itself reduce recall and precision. That's a real adoption risk worth tracking, but it's currently a prediction, not a documented outcome.

If prompts stay hidden, what proves the review actually worked?

Outcome-based validation — recall and precision metrics plus null-set sampling — instead of exposing the prompts that produced the results, the authors propose.

That's a reasonable substitute in principle, but it shifts the real question onto the producing party: who defines "recall," who runs the sampling, and who checks the null set for false negatives? A validation report generated and interpreted entirely by the party doing the review is not independent evidence. Any buyer relying on this defense should ask their eDiscovery vendor how those statistics are produced, whether the methodology is disclosed even if the prompts aren't, and who would audit them if challenged.

Should firms already be treating prompts as protected work product?

Treat prompts as attorney work in progress — version them and separate fact-driven edits from strategic ones, the paper's hypothetical implies.

The Regional Health Systems walkthrough is useful less as precedent and more as a drafting discipline: distinguishing a prompt update that merely reflects a newly learned fact from one that encodes legal theory helps a firm later argue for protection on a document-by-document basis, if it ever comes to that. It's a sensible internal practice regardless of how courts eventually rule.

Frequently asked questions

Is this paper a binding legal standard?

No. It's a position paper by two eDiscovery experts, Tara Emory and Maura R. Grossman, not a court decision, and no eDiscovery-specific ruling has adopted their framework.

What court decisions does the paper rely on?

Tremblay v. OpenAI and Concord Music Group v. Anthropic, both from AI-related litigation, which the authors read as leaning toward protecting attorney-crafted prompts — not rulings decided on eDiscovery prompt-disclosure disputes.

Does this change how search terms are treated?

No. Courts already treat search terms and general search-methodology descriptions as unprotected; the authors argue GenAI prompts are different because they can reveal more than a term list.

Source: eDiscovery Today, reporting by Doug Austin.