Three More A.I. Protective Orders

AI protective orders are spreading, but open-versus-closed is a weak confidentiality test

Writing for EDRM in October 2026, Michael D. Berman added more protective orders on AI use to his argument that such clauses are becoming the standard of care. In Hoffman v. Univ. of Nebraska Medical Ctr. (D. Neb., 2026 WL 2969621), a jointly requested order bars recipients from giving confidential material, deposition testimony or anything extracted from them to open-source AI platforms without the producing party's written consent. It permits closed-source tools and names Harvey, Westlaw, LexisNexis and Servient's AI Canvas. In ANR Pipeline Co. v. B&B Metal Processing Co. (E.D. Wis., 2026 WL 2970050), entered the same day, the court accepted the parties' stipulated order but modified it to set limits on AI use with confidential documents. Berman's title promises three orders, but the text we had cuts off partway through the Wisconsin one.

What do these orders actually forbid?

Nebraska bars uploading confidential material to open-source AI services without the producing party's consent, whatever privacy promises the service makes. Wisconsin allows AI tools only in a secure setting.

The Nebraska clause reaches information extracted from the material, not just the files. On its wording, an AI-built summary or chronology drawn from a confidential production is treated like the production itself. It defines open-source by example, pointing at services that use uploads to train or improve models.

Wisconsin took a different route. The parties asked for a conventional order protecting confidential technical, commercial or business information, and the judge added AI parameters, including a "secure environment" test. A stipulation drafted without AI terms can come back from the bench with them attached.

Is open-source versus closed-source a meaningful line?

Not by itself. Open or closed describes how a model is licensed, not where your data goes, who retains it or whether it trains anything. Nebraska's order leans on the label anyway.

For open-source platforms, Nebraska discounts the vendor's own confidentiality assurances. For the closed-source side, the excerpt shows a list of brand names plus anything the parties agree to. It shows no required terms on retention, training or access. Westlaw and LexisNexis are research products, and Harvey is a general legal assistant. A brand on a list tells you little about what happens to a particular upload.

Has a standard of care really formed?

Not yet proven. Berman argues so, but he is a commentator citing orders he chose to post about. Two orders sharing a date, one judge-modified, signal a trend, not a rule.

No outlet we saw offers a count of how many protective orders are entered without AI language, so there is no denominator. Berman also cites a consent order in a Google matter from September 2026, which shows the clauses appear in negotiated settings. The Wisconsin modification shows judges will add them anyway. Plan for both.

What should I ask my vendor and my own team?

Ask where data is processed, whether anything is retained or used for training, and what evidence proves it. Then ask whether the tool qualifies under the order your matter will carry.

Get the answers in contract language, not a sales deck. Check whether outputs such as summaries and chronologies are stored and where. Get the producing party's consent in writing before using any tool outside a named list. Architecture is evidence you can show a judge. Relevant e-Discovery, our own platform, runs single-tenant or inside a firm's AWS account under its own keys, and you should discount that mention accordingly. Test any vendor's claim against the order's wording, not its marketing.

Frequently asked questions

Do these orders ban AI in discovery?

No. Nebraska expressly permits closed-source tools. Wisconsin permits AI use under conditions, though we could not read all of them.

Do AI-generated summaries fall under the Nebraska clause?

Its wording covers information extracted from confidential material, so derived outputs sent to an open-source platform would appear to be caught. Ask counsel for a reading on your facts.

What about the third order?

We did not see it. The source text ends mid-sentence in the Wisconsin order, so read the full orders before relying on this note.

Sources: EDRM, Michael D. Berman, "Three More A.I. Protective Orders."