New York Federal Court's OpenAI Discovery Orders Provide Key Insights For Companies Navigating AI Preservation Standards

Duane Morris Takeaways: In a series of discovery rulings in the case of In Re OpenAI, Inc. Copyright Infringement Litigation, No. 23 Civ. 11195 (S.D.N.Y.), Magistrate Judge Ona T. Wang issued a series of orders that signal how courts are likely to approach AI data, privacy, and discovery obligations. Judge Wang's orders illustrate the growing tension between AI system transparency and data privacy compliance – and how courts are trying to balance them.

For companies that develop or use AI, these rulings highlight both the risk of expansive preservation demands and the opportunity to share proportional, privacy-conscious discovery frameworks. Below is an overview of these decisions and the takeaways for in-house counsel, privacy officers, and litigation teams.

Background

In May 2025, the U.S. District Court for the Southern District of New York issued a preservation order in a copyright action challenging the use of The New York Times' content to train large language models. The order required OpenAI to preserve and segregate certain output log data that would otherwise be deleted. Days later, the Court denied OpenAI's motion to reconsider or narrow that directive. By October 2025, however, the Court approved a negotiated modification that terminated OpenAI's ongoing preservation obligations while requiring continued retention of the already-segregated data.

The Court's Core Rulings

1. Forward-Looking Preservation Now, Arguments Later

On May 13, 2025, the Court entered an order requiring OpenAI to preserve and segregate output log data that would otherwise be deleted, including data subject to user deletion requests or statutory erasure rights. See id., ECF No. 551. The rationale: once litigation begins, even transient data can be critical to issues like bias and representativeness. The Court stressed that it was too early to weigh proportionality, so preservation would continue until a fuller record emerged.

2. Reconsideration Denied, Preservation Continues

A few days later, when OpenAI sought reconsideration or modification of preservation order, the Court denied the request without prejudice. Id., ECF No. 559. The Court noted that it was premature to decide proportionality and potential sampling bias until additional information was developed.

3. A Negotiated "Sunset" and Privacy Carve-Outs

By October 2025, the parties agreed to wind down the broad preservation obligation. On October 9, 2025, the Court approved a stipulated modification that ended OpenAI's ongoing preservation duty as of September 26, 2025, limited retention to already-segregated logs, excluded requests originating from the European Economic Area, Switzerland, and the United Kingdom for privacy compliance, and added targeted, domain-based preservation for select accounts listed in an appendix. Id., ECF No. 922.

This evolution — from blanket to targeted, time-limited preservation — shows courts' willingness to adapt when parties document technical feasibility, privacy conflicts, and litigation need.

Implications For Companies

1. Evidence vs. Privacy: Courts Expect You to Reconcile Both

These rulings show that courts will not accept "privacy law conflicts" as a stand-alone excuse to delete potentially relevant data. Instead, companies must show they can segregate, anonymize, or retain data while maintaining compliance. The OpenAI orders make clear: when evidence may be lost, segregation beats destruction.

2. Proportionality Still Matters

Even as courts push for preservation, they remain attentive to proportionality. While early preservation orders may seem sweeping, judges are open to refining them once the factual record matures. Companies that track the cost, burden, and privacy impact of compliance will be best positioned to negotiate tailored limits.

3. Preservation Is Not Forever

The October 2025 stipulation illustrates how to exit an indefinite obligation: offer targeted cohorts, geographic exclusions, and sunset provisions supported by a concrete record. Courts will listen if you bring data, not just arguments.

A Playbook for In-House Counsel

1. Map Your AI Data Universe

Inventory all AI-related data exhaust: prompts, outputs, embeddings, telemetry, and retention settings. Identify controllers, processors, and jurisdictions.

2. Build "Pause" Controls

Design systems capable of segregating or pausing deletion by user, region, or product line. This technical agility is key when a preservation order issues.

3. Update Litigation Hold Templates for AI

Traditional holds miss ephemeral or system-generated data. Draft holds that instruct teams how to pause automated deletion while complying with privacy statutes.

4. Propose Targeted Solutions

When facing broad discovery demands, offer alternatives: limit by time window, geography, or user cohort. Courts will accept reasonable, well-documented compromises.

5. Build Toward an Off-Ramp

Preservation obligations can sunset — but only if supported by metrics. Track preserved volumes, costs, and privacy burdens to justify targeted, defensible limits.

Conclusion

The OpenAI orders reflect a new judicial mindset: preserve broadly first, negotiate smartly later. AI developers and data-driven businesses should expect similar directives in future litigation. Those that engineer for preservation flexibility, document privacy compliance, and proactively negotiate scope will avoid the steep costs of one-size-fits-all discovery — and may even help set the industry standard for balanced AI litigation governance.

Disclaimer: This Alert has been prepared and published for informational purposes only and is not offered, nor should be construed, as legal advice. For more information, please see the firm's full disclaimer.