Your Client's AI Prompts Are Now Discoverable: A Litigator's Guide to E-Discovery in the Age of Generative AI
10/25/20265 min read
Your Client's AI Prompts Are Now Discoverable: A Litigator's Guide to E-Discovery in the Age of Generative AI
[Editorial note — pending human legal review before publication. Educational only; not legal advice; not individualized guidance. Describes recent trial-court and magistrate decisions and draft standards guidance in a fast-moving area; verify current authority for your jurisdiction. Jurisdiction-specific rules distinguished from general commentary.]
Generative AI has put litigators in a strange position: it's simultaneously the most powerful document-review tool they've ever had and an entirely new category of discoverable evidence. Those two roles create two distinct sets of duties, and most litigation teams have thought carefully about neither.
One question — is my client's AI interaction privileged or waived? — we've covered separately. This is the other question, and it's just as consequential: what must you preserve and produce, and what are your obligations when you use AI to run discovery itself?
Part 1: AI as a discovery target
There is no "AI privilege." Courts are treating AI prompts and outputs as electronically stored information under the ordinary rules. Whether a particular interaction happens to be protected by privilege or work product is a separate analysis; as a baseline, AI-generated content is ESI, and if it's relevant and non-privileged, it's discoverable like any other ESI. No special shield attaches simply because a machine was involved.
The duty to preserve now reaches AI tools. Once litigation is reasonably anticipated, the duty to preserve extends to discoverable AI data. In practice, that means your litigation holds should name AI prompts, outputs, and relevant logs directly, rather than assuming a generic "all ESI" hold captures them. The stakes are the familiar ones: under Federal Rule of Civil Procedure 37(e), a failure to take reasonable steps to preserve ESI that results in prejudice can support serious sanctions, including adverse-inference instructions.
The duty to produce is bounded by relevance and proportionality. This is where litigators have real room to work. Production obligations extend only to AI materials that are relevant to a claim or defense and proportional to the needs of the case under Rule 26(b)(1). Two recent rulings in the copyright litigation against OpenAI illustrate both directions. In one, Magistrate Judge Ona Wang ordered OpenAI to produce roughly 20 million output logs — including user prompts and AI outputs — rejecting arguments that the logs were irrelevant and that production was disproportionate, subject to anonymization. In a related ruling, the court declined to compel production of a party's internal AI-tool prompts and outputs, finding them not relevant to the core issues and not proportional given the substantial review and privilege burden. The lesson: AI data is discoverable, but it isn't automatically discoverable. Relevance and proportionality objections remain entirely consistent with the rules.
The emerging standard. The Sedona Conference's Working Group 1 circulated draft guidance in April 2026 (still in draft, for discussion) taking the sensible position that generative AI information should be treated like other forms of discoverable ESI — preserved, collected, and produced through reasonable, proportional efforts, with the familiar inquiries of relevance, privacy, proportionality, and burden applied to prompts, grounding data, and outputs.
The practical move is to think of AI interactions a bit like internet search history: potentially revealing, often sensitive, and best handled through targeted, case-specific analysis rather than either blanket preservation of everything or reflexive refusal to produce anything.
Part 2: AI as a discovery tool
The other half of the duty runs to how you use AI in the review itself.
Defensibility hasn't changed its shape. Technology-assisted review has been accepted for years, and courts are extending that acceptance to generative-AI review — one 2026 decision declined to give GenAI review special scrutiny and treated it like TAR. What courts require is what they've always required of TAR: a reasonable, transparent, validated process. Rule 26(g) governs the certification of your discovery methods, and proportionality under Rule 26(b)(1) actually cuts in AI's favor — AI-driven culling and prioritization can be strong evidence that you took a reasonable, cost-proportionate approach.
Your review prompts are usually protected — usually. Work by Tara Emory and Maura Grossman on "GenAI TAR" reflects the developing consensus that, absent party agreement or a demonstrated flaw in the production, attorney review prompts "should ordinarily not be subject to disclosure." They're treated like the mental impressions behind a review protocol. But that protection has a limit familiar from discovery-on-discovery: if there's a showing of a material deficiency — gaps traceable to human error in categorization, for instance — a court may permit limited discovery of the prompts behind the process. And prompts outside the attorney-review context are a different matter: one 2026 ruling found an expert's AI prompts fair game under Rule 26.
Human oversight isn't optional. Agentic and generative review introduces a failure mode traditional TAR didn't have: an early misclassification can quietly propagate through hundreds of downstream decisions, and fluent models can produce confident, wrong justifications. Endpoint metrics won't catch it. A defensible GenAI-review process keeps a human attorney in the loop at calibrated checkpoints, not just at the finish line.
The competence thread tying it together
Both roles collapse into a single professional obligation: you can't preserve, produce, protect, or defend what you don't understand. Knowing where your client's discoverable AI data lives, and knowing how your own review tool reaches its conclusions, is now part of litigation competence — not an e-discovery vendor's problem to worry about alone. (California litigators face parallel preservation and production duties under the state's Civil Discovery Act and its ESI provisions; the analysis tracks the federal framework closely.)
A litigator's AI e-discovery checklist
Name AI tools, prompts, and outputs explicitly in litigation holds once litigation is reasonably anticipated. Map where your client generates discoverable AI data — enterprise tools, public chatbots, agentic systems — before a request arrives. Preserve in a targeted, proportional way; document your reasoning rather than capturing everything. Be ready to argue relevance and proportionality against overbroad AI requests — and to make them narrowly when you're the requesting party. When using AI/TAR to review, build a transparent, validated, Rule 26(g)-defensible process. Keep attorney review prompts protected by running a clean process; a material deficiency is what opens them up. Insert human checkpoints to catch cascading misclassifications before they reach a privilege log.
FAQ
Are AI prompts and outputs discoverable? Yes, when they're relevant to a claim or defense and proportional to the case. Courts treat them as ESI under the Federal Rules; there is no special "AI privilege."
Do we have to preserve ChatGPT or similar conversations? The duty to preserve extends to discoverable AI data once litigation is reasonably anticipated. Litigation holds should name AI tools, prompts, and outputs directly rather than assuming a generic hold covers them.
Does that mean every AI interaction must be produced? No. Relevance and proportionality still govern. Courts have both compelled large AI-log productions and declined to compel AI data that wasn't relevant or was disproportionate to produce.
If I use AI to review documents, are my review prompts discoverable? Generally not. Attorney review prompts are usually protected like a review protocol's mental impressions — unless there's a demonstrated material deficiency in the production, which can open limited discovery into them.
Is AI-assisted review defensible? Yes, when the process is reasonable, transparent, validated, and supervised. Courts increasingly treat generative-AI review like established TAR under Rule 26(g).
AI will touch nearly every matter you handle, on both sides of the discovery line. The litigators who get ahead of it — targeted holds, proportional preservation, defensible review processes — will spend far less time fighting avoidable discovery battles. If your team wants a practical AI e-discovery protocol built around your actual tools and matters, that's exactly the kind of readiness we help litigators put in place.