On June 30, 2026, Magistrate Judge Laurel Beeler of the Northern District of California issued a ruling that practitioners in the eDiscovery community had been anticipating for years. In Schulte v. LinkedIn Corp., No. 22-cv-00237-HSG (LB) (N.D. Cal. June 30, 2026), the court became the first federal tribunal to directly confront a party's use of generative AI — specifically, Relativity aiR — as a document review tool. The question before the court was elemental: does generative AI in discovery require its own legal framework, or does the existing technology assisted review ("TAR") doctrine absorb it?
Judge Beeler's answer was unambiguous. Generative AI review is simply a form of TAR, subject to the same reasonableness and proportionality standards that have governed predictive coding and machine learning tools since Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012). No special rules are required. The ruling has been recognized by Arnold & Porter and WilmerHale as a watershed moment signaling growing judicial confidence in AI-assisted review workflows. This article examines the court's reasoning, identifies the practical obligations the decision imposes, and provides concrete guidance for litigators and in-house counsel who are either already deploying or considering AI-assisted review.
The Court's Reasoning: Continuity, Not Revolution
The plaintiffs in Schulte pressed two principal objections to LinkedIn's use of Relativity aiR. First, they argued that LinkedIn's decision to pre-cull documents with keyword search terms before running the generative AI tool was impermissible — that it arbitrarily narrowed the universe of documents the AI could assess. Second, they demanded disclosure of additional validation metrics, including elusion estimates and the number of human reviewers involved in quality control, beyond what LinkedIn had already disclosed under the operative ESI protocol.
Judge Beeler rejected both arguments. On pre-culling, the court found that applying keyword filters before running a generative AI model is a reasonable, accepted workflow step, not a methodological defect. The logic is familiar: search terms have long been used as a first-pass filter in TAR 1.0 and TAR 2.0 environments, and their use alongside a more sophisticated AI layer does not undermine the overall reasonableness of the process. On the metrics question, the court declined to compel disclosure of elusion estimates or reviewer counts where the requesting party had identified no specific deficiency in the production itself. Speculation about what the AI might have missed is not, standing alone, grounds for discovery into the producing party's review methodology.
The court's framework is, in essence, the principle that opposing counsel cannot force "discovery on discovery" into AI workflows absent a concrete showing that the production is deficient. That principle — rooted in proportionality under Federal Rule of Civil Procedure 26(b)(1) — applies to generative AI precisely as it applies to any other review tool.
What the Decision Does Not Resolve
Schulte is a significant first step, not a comprehensive roadmap. The court did not address how courts should evaluate generative AI review when no ESI protocol is in place, nor did it set minimum validation thresholds applicable to all AI-review deployments. It did not speak to the obligations of parties who use AI tools that operate as black boxes, where the producing party cannot explain, at even a general level, how the model was trained or calibrated for the matter at hand. Those questions remain open, and litigators should not read Schulte as a license to deploy AI tools without documentation, transparency, or human oversight.
Practical Guidance for Litigators and In-House Counsel
The following steps will position counsel to defend AI-assisted review workflows consistent with Schulte and the broader TAR doctrine.
1. Negotiate an ESI protocol that specifically addresses AI review before the project begins. The court's denial of the plaintiffs' metrics demands rested in significant part on LinkedIn's compliance with its existing disclosure obligations under the operative protocol. The ESI protocol is the primary shield. Counsel should address, at minimum: the type of AI tool to be used; the general workflow, including any pre-culling steps; and what validation information will be shared voluntarily.
2. Document the workflow at each stage. Record which search terms were applied in pre-culling, how the generative AI model was configured for the matter, and what quality-control steps — including human review — were applied to the AI's outputs. This documentation is the foundation of any proportionality defense.
3. Retain a forensic or eDiscovery expert to validate the process. An independent technical assessment of the AI workflow, conducted before disputes arise, provides objective support for the reasonableness of the methodology. Validation metrics need not be disclosed to opposing counsel as a matter of course; they should, however, exist internally.
4. Do not treat judicial approval of AI review as approval of any particular AI deployment. The court evaluated Relativity aiR as used in a specific workflow with specific disclosures. A different tool, a different protocol, or a different level of documentation will produce a different record — and potentially a different outcome.
5. Respond to opposing counsel's challenges with specificity, not generality. If an adversary demands additional validation metrics, require that they identify a concrete production deficiency before engaging. Schulte confirms that speculative challenges to AI methodology do not, without more, justify compelled disclosure.
Conclusion
In conclusion, Schulte v. LinkedIn establishes that the courts are prepared to evaluate generative AI-assisted document review under the same doctrinal architecture that has governed TAR for more than a decade. That is, on balance, a stabilizing development: practitioners now have a framework rather than a vacuum. The practical imperative, however, is preparation. A well-negotiated ESI protocol, a documented workflow, and defensible validation practices are not optional refinements — they are the conditions on which judicial deference to AI-assisted review will continue to rest. As generative AI tools become more capable and more widely deployed in litigation, the courts' confidence in those tools will track the rigor of the practitioners who use them.