In brief
In one of the first rulings to examine the use of AI tools in discovery, the court in Schulte v. LinkedIn Corp., No. 22-cv-00237-HSG (LB) (N.D. Cal. July 1, 2026) provides guidance on how courts may evaluate generative AI tools used in document review.
The court permitted LinkedIn’s use of Relativity aiR, a generative AI-powered document review product for Relativity, over multiple challenges raised by the plaintiffs. The court applied traditional discovery principles, finding LinkedIn’s keyword filtering and AI-assisted responsiveness review reasonable and proportional, while declining to require broader disclosures about the AI tool absent a specific deficiency in the production.
Key takeaways
The district court’s ruling offers several practical takeaways:
- Courts may evaluate generative AI-review platforms under traditional proportionality and reasonableness standards rather than impose AI specific obligations.
- Parties can use appropriate search terms to “pre-cull” results prior to using an AI-review tool.
- Parties are unlikely to be ordered to engage in “discovery on discovery” such as disclosing performance metrics or otherwise revealing details on their discovery process, absent some showing of a specific deficiency in the production.
- Courts are embracing the use of generative AI-tools and focusing on discovery processes using those tools, rather than whether the tools themselves are appropriate.
In more detail
LinkedIn disclosed to plaintiffs that it would use Relativity aiR to filter nonresponsive documents from its products, use aiR to make final responsiveness determinations, and employ search terms to pre-cull documents before aiR review. In response, plaintiffs sought three forms of relief from the court: (1) an order prohibiting the use of search terms to pre-cull documents; (2) an order compelling LinkedIn to run Relativity aiR across all custodial files; and (3) an order compelling disclosure of additional metrics used for validation. The court rejected each of the plaintiffs' requests.
First, it found that LinkedIn’s use of search terms before running documents through the AI-review process was reasonable and proportional, particularly because the plaintiffs had not shown that the search terms themselves were defective or excluded relevant documents. Second, the court declined to require LinkedIn to process all custodial data through the AI-system, noting the significant burden and expense that would result. In other words, forcing a party to run GenAI across raw, multi-terabyte custodial repositories without prior keyword filtering was found to impose an undue burden regarding processing, hosting, and validation costs. Third, the court refused to order further disclosure regarding the AI-tool’s performance. The court emphasized that requests to scrutinize an opponent’s discovery process are generally disfavored unless there is evidence of a specific deficiency in the production. Because LinkedIn disclosed its use of AI, including that the technology was making final responsiveness calls (accompanied by human quality control sampling), and the plaintiffs could not identify a concrete problem with the results, additional disclosures were not required.
Conclusion
The decision represents an important early precedent for litigants employing generative AI in discovery. While courts are unlikely to grant AI-tools special treatment, Schulte suggests that well-designed AI review processes will be evaluated under familiar discovery principles governing reasonableness, proportionality, and quality control. As organizations continue to deploy generative AI in litigation workflows, parties should focus on building defensible review protocols rather than expecting courts to create a separate framework for AI-enabled discovery.