Dunn: An Interesting AI Protective Order Analysis Uses a Functional Definition

“MD Judiciary’s online dispute resolution tool is back”
November 15, 2026

A.I. Protective Orders govern how the recipient of discovery materials may—or may not—use A.I. to analyze information produced to it. A.I. Protective Orders Are Becoming the Standard of Care (Nov. 4, 2026).

“Knowing how confidential information is handled [by an opponent] is a legitimate interest.”  Dunn v. LexisNexis Risk Solutions, Inc., 2026 WL 2568346 (D. Col. Aug. 31, 2026). The court explained:

When a party uploads their own documents and information [into an AI model], they accept the risk (however small) that their own confidential information could become compromised. By contrast, when a party turns over confidential information to another, the producing party cannot manage their risk absent agreement or court order. Given the pace of change in consumer-grade AI, and the limited visibility into how these tools process data, it is reasonable for a party to request that confidential information be kept out of consumer-grade AI tools that do not offer the same assurances and configurations found in certain enterprise offerings.  [emphasis in original].

However, there is a wide variety of AI Protective Orders. Some address all discovery materials.  Others are limited to confidential discovery materials.  Some specify permitted and prohibited platforms by name.  Others specify functionality.  A few do both.  Some are a page long; others are only a paragraph.  Motion to Permit Limited A.I. Review of “Confidential” Video Was Denied (Oct. 31, 2026).

Phil Favro has asked “Will Courts Enforce Protective Order Restrictions on AI Use?” (Nov. 12, 2026), and offered a drafting suggestion, citing Phil Favro, Will Courts Enforce Protective Order Restrictions on AI Use? (Favro Law Sep. 29, 2026).

I have also echoed Craig Ball’s cautions and concerns that equity be considered and that these orders not become an obstacle to unrepresented or indigent litigants. See  A.I. Protective Orders Are Becoming Routine (Aug. 27, 2026), and  Another A.I. Protective Order – Part 10 (Aug. 21, 2026), both citing Craig Ball, The AI Protective Order Double Standard | Ball in your Court (Jul. 27, 2026).

The court’s analysis in Dunn v. LexisNexis Risk Solutions, Inc., 2026 WL 2568346 (D. Col. Aug. 31, 2026), is helpful in approving a functional definition of permitted or prohibited AI.  The parties agreed to “almost every aspect” of a protective order, with one exception, the “Additional Disputed Language.”

The agreement “prohibits the uploading of confidential information to certain AI platforms but permits the use of secure platforms, so long as they limit training and disclosure, and allow deletion.”  That is a functional definition. The factors are: 1) training; 2) disclosure; and, 3) deletion.  The agreement goes further and also “also identifies the tools the parties intend to use.”  However, the Dunn court wrote:

Still, the parties could not reach agreement on the Additional Disputed Language: one sentence at the end of the proposed protective order that would require disclosure, consent, and amendment each time a party seeks to use a new AI tool with confidential information. Defendants support this requirement. Plaintiff objects.

The Dunn court rejected the request for future disclosure of any new AI tool, and the “Additional Disputed Language.” It viewed the “consent-and-amend” process as cumbersome, requiring a new court order “every time a party adopts a new tool.”  It also held that “the purpose for … transparency has been served” by the agreement  that “sets clear, tool-agnostic requirements: no model training on inputs, no third-party disclosure beyond what is essential for service delivery, and the ability to delete information.”  That is what Craig Ball suggested.

The court wrote:

Now that the parties have specific parameters, they do not need to return to this Court every time they adopt a new tool. Both parties are represented by competent counsel, and they are each capable of evaluating whether a new tool satisfies the standing requirements.

In my words, the functional description of the agreed-upon language provided sufficient protection.  The court wrote that: “[I]t is reasonable for a party to request that confidential information be kept out of consumer-grade AI tools that do not offer the same assurances and configurations found in certain enterprise offerings.” [emphasis added].

For more information on AI Protective Orders, please click on the “AI Protective Orders” link in the “Categories” tab at the upper right side of this post.

Share