Background
On November 6, 2026, United States Magistrate Judge Susan van Keulen signed and entered the stipulated protective order, as modified by the court, in Gavalas v. Google LLC et al., No. 5:26-cv-1849-EKL-SVK, in the Northern District of California, San Jose Division. The caption identifies Joel Gavalas, acting as personal representative of the Estate of Jonathan Gavalas, as plaintiff and Google LLC and Alphabet Inc. as defendants. Counsel executed the stipulation on October 29, 2026, and the order states that each party agreed to be bound upon counsel's execution even before court entry. The document does not decide the underlying claims.
The parties explained that discovery was likely to involve confidential, proprietary, or private information warranting protection from public disclosure and from use outside the litigation. They based their proposal on the district's model order for litigation involving patents, highly sensitive confidential information, and trade secrets. The court's redline modifications principally direct unresolved confidentiality, expert-access, and modification disputes through Judge van Keulen's standing-order procedures and joint statements. The order does not confer blanket confidentiality or an automatic right to seal filings; Civil Local Rule 79-5 continues to govern sealing requests.
The Court’s Holding
The court ordered a two-tier confidentiality framework covering material designated "CONFIDENTIAL" or "HIGHLY CONFIDENTIAL – ATTORNEYS' EYES ONLY," together with copies, extracts, summaries, compilations, and communications revealing that material. Protected material may be used only to prosecute, defend, or settle this litigation, not for business purposes. Access is limited to specified qualified recipients, generally subject to necessity and execution of Exhibit A's acknowledgment. The definition of an expert addresses employment by parties or competitors, expressly including certain developers of large language models, AI chatbots, AI assistants, or competing AI services. Those restrictions are not absolute bars: written approval or court review may permit access, and later objections remain available for subsequently produced material or newly discovered expert information.
Section 12.9 specifically prohibits protected material and information derived from it from being inputted, uploaded, or submitted to open, publicly accessible, freely available, consumer-grade large language models or AI tools. A receiving party may use only an enterprise-grade "Confidential AI Tool" employing industry-standard safeguards after ensuring that its provider contract prohibits use or storage of inputs to train, refine, or improve the model; restricts disclosure of inputs, metadata, and generated information to third parties, subject to a protected service-delivery exception; and permits deletion of protected material and generated information on request. The receiving party must retain written documentation of those contractual protections.
AI-derived inputs, outputs, models, embeddings, indexes, logs, and other artifacts containing protected material remain subject to the litigation-only use restriction. Submission to a noncompliant AI tool is expressly treated as unauthorized third-party disclosure under Section 10, triggering notice, retrieval, and corrective duties. At final disposition, protected material and AI-generated outputs must be returned or destroyed within 60 days, with written certification, although counsel may retain specified archival litigation materials that remain protected. Confidentiality obligations and the court's enforcement jurisdiction continue after final disposition, except for material that becomes public record.
Key Takeaways
- The order bars protected discovery material and derived information from being submitted to consumer-grade or publicly accessible AI tools.
- Enterprise AI use is permitted only after contractual safeguards addressing model training, third-party disclosure, security, and deletion are confirmed and documented.
- A noncompliant AI submission constitutes unauthorized disclosure to a third party and triggers the order's disclosure-response obligations.
- AI-generated outputs and residual artifacts remain protected, must serve only this litigation, and are included in the final return-or-destruction process.
- The order governs discovery confidentiality and does not adjudicate the parties' underlying claims or authorize blanket sealing.
Why It Matters
The decision provides a detailed, court-entered framework for using generative AI during litigation without treating AI processing as outside discovery confidentiality obligations. It addresses not only initial uploads but also training use, metadata, integrated third-party tools, generated outputs, residual embeddings and logs, deletion rights, and end-of-case certification. Its express classification of noncompliant submission as unauthorized disclosure supplies a concrete enforcement consequence rather than a general admonition to use AI carefully.
The order also integrates AI-related competitive concerns into expert access while preserving approval, objection, and judicial-review procedures. Its protections operate alongside ordinary limits on designations, privilege handling under Federal Rule of Evidence 502(d), secure storage and transmission requirements, and continuing jurisdiction. The ruling's scope is substantial for management of confidential discovery and AI-assisted review, but it does not establish liability for AI misuse or resolve any substantive claim against Google or Alphabet.
Your browser cannot display this PDF inline.