Live Webinar · Aug 10 Gen AI for In-House Counsel: how we post-train an LLM to run Document Review at under 5¢/document. Save Your Seat →
Short answer

Do I have to disclose GenAI in the ESI protocol?

If the protocol already mentions TAR, assume it covers GenAI. Raise it at the Rule 26(f) conference before you spend the money.

If your ESI protocol already requires you to disclose TAR, assume a court may treat generative AI review as TAR and require the same disclosure. That is what happened in Schulte v. LinkedIn, 2026 WL 1905851 (N.D. Cal. June 30, 2026).

The Federal Rules still do not say “GenAI.” There is no new nationwide duty that names the model. The practical duty is the one you already have: reasonableness, Rule 26(g), and whatever you agreed to in the protocol.

What courts have actually required

TAR cases have long required “some level of transparency”: the tool, the workflow, and how you will validate. Seed sets are sometimes withheld. In Schulte, LinkedIn’s protocol required TAR disclosure. The court applied that term to the GenAI tool and then refused a deeper audit of metrics without a production problem.

Prompts are the open issue. A prompt can look like a search term (often exchanged), a seed set (sometimes withheld), or attorney work product (rarely shared). Do not litigate that for the first time in a Rule 37 motion after the review is done.

What to put on the 26(f) agenda

Say that you may use generative AI for responsiveness, privilege, or both. Say whether it will make a final call or only a first pass. Agree what you will disclose (tool, no-seed-set design, sampling plan) and what you will not (prompts, raw metrics) unless a deficiency showing is made. Write it down. An agreed protocol is cheaper than a fight and is the record you will need later.

A minimum disclosure that usually works

Name the platform and that it is generative, not CAL. State whether humans still review a sample from each prediction bucket. State that prompts are treated as work product unless the protocol says otherwise. Offer to discuss validation methodology, not raw document-level scores, unless a deficiency is shown.

ABA Formal Opinion 512 still sits next to the protocol: competence, confidentiality (do not put the corpus in a consumer model that trains on prompts), supervision, and a client conversation when the means matter. Disclosure to the other side and ethics duties to the client are different lists. Do both.

Ready to see DecoverAI in action?

Book a demo and see classification, privilege review, and production in one pass.

Book a Demo →