Court Evaluates AI-Related Provisions in Discovery Protective Orders

As generative AI becomes more common in litigation workflows, courts are increasingly being asked to address how AI tools may be used with confidential discovery materials.

A recent decision from the District of Colorado provides insights into one court’s approach to generative AI in discovery. The ruling distinguishes between public consumer AI tools and enterprise platforms with specified safeguards, while addressing proposed provisions requiring approval before new AI tools may be used.

In this consumer dispute, the parties agreed that discovery material was confidential. After a hearing, they even agreed on a cutting-edge artificial intelligence provision for their protective order.

The agreed-upon terms established standard safeguards:

• No public dumping: A strict ban on uploading confidential data to generic, consumer-grade AI platforms.
• Secure tools only: Permitting platforms that explicitly block model training, prevent third-party disclosure, and allow complete data deletion.

The parties, disagreed, however, about additional proposed language, with defendants demanding “Additional Disputed Language” requiring that every time a party wanted to use a new AI tool, they had to secure opposing counsel’s consent and formally amend the court’s protective order. Plaintiff opposed the proposed provision.

The Magistrate Judge rejected plaintiff’s claim that AI risks are “identical” to everyday cloud storage. While the court previously noted in Morgan v. V2X, Inc.[1] that routing data through AI resembles routing it through email, the Judge emphasized that AI systems still present heightened risks. They are nascent, operate under less mature frameworks, and process inputs with a unique opacity. As the court put it, that opacity “changes the calculus when someone else’s data is at stake.”

This distinction—your data vs. your opponent’s data—drove the final ruling:

1. Use of a party’s own information: A party uploading its own proprietary files to a generative AI tool accepts whatever data risk that follows.
2. Use of an opposing party’s confidential information: A party handing an adversary’s confidential data cannot manage that risk without strict court-enforced guardrails.

Because of this, the court agreed that blocking consumer-grade AI tools was highly reasonable. However, the court ruled that the defense’s “consent-and-amend” clause went too far. Forcing parties to trigger motion practice every time a firm updates its software suite is inefficient. Because the protective order already established strict, tool-agnostic standards (no training, no disclosure, deletion capability), competent counsel can evaluate new tools against those rules without running back to the judge.

This decision offers several takeaways, including:

• AI-specific provisions are appearing more frequently in protective orders: AI-specific provisions are no longer a novelty. The decision will be of interest to practitioners who are considering whether and how to address AI use in protective orders.
• Importance of data ownership considerations: The court placed significant emphasis on the distinction between a party’s own information and an opposing party’s confidential information.
• The court favored criteria-based requirements over tool-specific references: In this case, the court declined to require provisions that would have required approval and amendment of the protective order whenever a new AI tool was introduced.