By Kelly Twigger

Podcast | Transcript

Welcome to the final Case of the Week for 2025! As we close out the year and look toward a fantastic 2026, we are doing something a little different. Rather than focusing on a single order, we are diving deep into the OpenAI Copyright MDL in the Southern District of New York.

This consolidated litigation—involving The New York Times, Microsoft, and OpenAI—has delivered the ruling we’ve been waiting for: Generative AI prompts and outputs are now squarely treated as core, discoverable ESI.

If you are building strategy based on assumptions that AI chats are too private or too voluminous to produce, you might be building on sand. Here is the reality check from the Southern District of New York and how Minerva26 helps you stay ahead of it.

The Core Conflict: Preservation vs. Privacy Promises

The central question in this MDL is one every modern litigator faces: How do you preserve and produce chat logs at scale when that collides with privacy promises and global data protection laws?

OpenAI argued that they deleted logs to protect user privacy and comply with GDPR. The Court’s response was swift and decisive: Preserve it anyway.

In May 2025, Magistrate Judge Ona Wang entered a preservation order requiring OpenAI to segregate and hold all output log data that would otherwise be deleted. The lesson? Once data is potentially relevant, courts will override your normal deletion and privacy practices.

The 20 Million Log Dispute

By late 2025, the parties agreed that reviewing every log was impossible. The solution was sampling.

  • The Proposal: OpenAI proposed a sample of 20 million consumer logs, de-identified using their own internal tools to strip PII (names, emails, account IDs).
  • The Pivot: Once the sample was prepared, OpenAI tried to limit production to only logs containing specific “search term hits,” arguing that producing the full set was disproportionate and intrusive.
  • The Ruling: On December 2, 2025, the Court rejected OpenAI’s argument. Because OpenAI’s own defense relies on showing “substantial non-infringing use” of the models, the entire sample—even chats having nothing to do with the plaintiffs’ content—became relevant to the merits of the case.

3 Strategic Takeaways for 2026

If you are navigating discovery for clients using ChatGPT, Copilot, Gemini, or Claude, these are the new rules of the road:

1. Prompts are Squarely in the Duty to Preserve

Do not wait for a judge to tell you ChatGPT prompts are relevant. If you wait, you may already face spoliation sanctions.

  • System-Based Discovery: While you may need individual custodian exports for key witnesses (like Sam Altman’s texts), Generative AI logs push us toward system-based discovery. You need to identify the platforms (Enterprise, API, Team accounts) and preserve at the tenant level, not just the user level.

2. If You Propose the Sample, You Own It

OpenAI proposed the 20 million figure and their own de-identification tool as the “reasonable” solution. When they later tried to retreat to search terms, the Court held them to their word.

  • Strategy Tip: Never treat numbers in ESI protocols or joint letters as throwaway talking points. Assume the Court will view your “reasonable proposal” as a binding commitment.

3. Privacy is a Dial, Not an Off Switch

This is the most critical doctrinal point from Judge Wang: Privacy is not a seventh factor in the Rule 26(b)(1) proportionality analysis.

  • Privacy is part of the “burden” analysis.
  • Courts mitigate this burden through engineering: sampling, de-identification, and protective orders (AEO).
  • Privacy concerns will limit how discovery proceeds, but they will not prevent it.

How Minerva26 Helps You Manage the Chaos

The OpenAI MDL teaches us that you cannot figure out your data map after the preservation letter arrives.

Minerva26 is built to be the strategy layer that sits on top of your tools. It helps you:

  • Map Data Sources: Track which systems (Copilot, Gemini, internal LLMs) are logging prompts and where those logs live.
  • Understand Retention: Know which “privacy knobs” and retention settings are active before a duty to preserve triggers.
  • Devise Strategy: Decide proactively what to push for or resist based on the latest rulings, rather than scrambling when a court starts talking about “de-identification pipelines.”

Minerva26 isn’t another system to feed; it’s the intelligence layer that ensures you aren’t playing catch-up.

There is so much more in the full recording, including the nuance of the “Fair Use” defense and why OpenAI’s request for a stay was denied.

Listen to the Full Episode

From all of us at Case of the Week, thank you for a wonderful 2025. Here’s to a little rest, a lot of joy, and a strategic start to 2026.

The post Discovery Wake-Up Call: AI Prompts Are Discoverable (And Privacy Won’t Save You) first appeared on Minerva26.