Court greenlights generative AI for document review: Key takeaways from Schulte v. LinkedIn

In a first-of-its-kind ruling, a federal judge permitted LinkedIn’s use of Relativity aiR, a generative artificial intelligence (AI)-powered document review tool, in responding to discovery requests. The court’s decision in Schulte v. LinkedIn Corp., No. 22-cv-00237-HSG (LB), 2026 WL 1905851 (N.D. Cal. July 1, 2026), is the first federal court decision accepting the use of generative AI to make final responsiveness determinations in litigation discovery. Notably, the use of AI for final responsiveness determinations was neither challenged by the plaintiffs nor questioned by the court, which may reflect growing acceptance of generative AI in e-discovery workflows.

The decision also fits within a longer line of federal authority approving the reasonable use of advanced review technologies in discovery. In Da Silva Moore v. Publicis Groupe, 868 F. Supp. 2d 137 (S.D.N.Y. 2012), this author (Peck) issued the first judicial opinion approving technology-assisted review (TAR), holding that predictive coding was an available and judicially approved tool in appropriate large-volume cases, provided counsel implemented a reasonable process with proper quality-control and proportionality safeguards.

Viewed against that history, Schulte is less a departure than an extension of the principles first articulated in Da Silva Moore. If Da Silva Moore served as the landmark decision that validated the use of TAR in discovery, Schulte addresses the next generation of review technology, considering how generative AI review tools should be evaluated when deployed as part of a documented, proportionate workflow.

This alert examines the court’s reasoning and highlights practical takeaways for organizations managing complex e-discovery workflows.

Background

Schulte v. LinkedIn is a putative class action filed in the Northern District of California.

On May 15, 2026, LinkedIn disclosed that it would employ Relativity aiR, a generative AI-powered review platform, to filter non-responsive documents from its production. LinkedIn disclosed that: 

1. No seed or training set was used
   
   
2. Relativity aiR was being used to make final responsiveness determinations
   
   
3. Quality control was conducted through human review of samples from each responsiveness category. 

LinkedIn also provided 25 search strings to be applied to custodial documents as a pre-culling step before AI review. 

Plaintiffs raised three distinct issues concerning the scope and methodology of LinkedIn’s production. None challenged LinkedIn’s use of Relativity aiR itself. The court denied all three motions.

The court’s reasoning

Generative AI document review 

Plaintiffs sought three forms of relief: 

1. An order prohibiting LinkedIn from using search strings to pre-cull documents before submitting them to Relativity aiR
   
   
2. An order compelling LinkedIn to run Relativity aiR across all custodial files
   
   
3. An order compelling disclosure of additional validation metrics, including elusion estimates, document error rates, and the number of human reviewers validating the AI tool’s predictions

Pre-culling with search strings 

The court found that plaintiffs “did not argue or show that LinkedIn’s 25 search strings were deficient.” Citing the TAR decisions in Livingston v. City of Chicago, No. 16-cv-10156, 2020 WL 5253848 (N.D. Ill. Sept. 3, 2020), and In re Biomet M2a Magnum Hip Implant Prods. Liab. Litig., No. 3:12-MD-2391, 2013 WL 1729682 (N.D. Ind. Apr. 18, 2013), the court noted that “courts have held that using search terms to pre-cull documents before technology-review platforms satisfies the reasonableness and proportionality standards of Rules 26(b) and 34(b)(2).” Schulte, 2026 WL 1905851, at *2.

The court also emphasized the practical realities of discovery at scale. Two custodians alone had over 800 gigabytes of data, and, across 19 custodians, LinkedIn would need to feed multiple terabytes into Relativity aiR, resulting in significant processing, hosting, and human review costs. The court ordered the parties to meet and confer within 21 days regarding the adequacy of the search strings. Id. at *2.

Validation metrics and “discovery on discovery” 

The court applied the “discovery on discovery” standard from Taylor v. Google LLC, No. 20-CV-07956-VKD, 2024 WL 4947270, at *2 (N.D. Cal. Dec. 3, 2024), under which such discovery is disfavored and requires a “showing of specific deficiency” before it will be ordered. The court found that LinkedIn had satisfied its disclosure obligations under Paragraph 5(a) of the Interim ESI Order, which required parties to disclose their intent to use TAR tools. Plaintiffs’ only identified deficiency – that the target population comprised 204,444 documents – was deemed insufficient to warrant further disclosure.

Key takeaway

The court treated generative AI review tools like Relativity aiR no differently than traditional TAR and predictive-coding platforms for purposes of reasonableness analysis. That approach echoes Da Silva Moore, which made clear that TAR could be judicially approved in appropriate cases when supported by a reasonable process, quality control, and proportionality. Indeed, the court in Schulte described Relativity aiR as “a form of technology-assisted review,” Schulte, 2026 WL 1905851, at *3, signaling that generative AI tools may be viewed as an evolution of the earlier-generation TAR tools courts have approved for more than a decade. 

The decision also indicates that responding parties are not required to disclose detailed performance metrics absent a showing of specific deficiency in the production. At the same time, organizations using AI-assisted review tools may continue to maintain robust validation processes. Metrics such as elusion, precision, and recall remain important components of a defensible AI review workflow, even where opposing parties cannot compel their disclosure based on speculation alone.

More broadly, Schulte reflects a shift in how courts are evaluating AI-assisted document review. The focus appears to be moving away from whether generative AI may be used in discovery and toward more operational questions, such as how the document population was constructed, what information was disclosed to the opposing party, and whether the review process was proportionate. In that respect, the issues addressed in Schulte resemble those the courts have long considered in connection with traditional TAR, applied in the context of newer AI-enabled review tools.

Practical implications

This decision carries several implications for organizations and litigation teams managing e-discovery:

1. Generative AI review tools may be used in production workflows. The court’s refusal to impose special restrictions on Relativity aiR signals that generative AI tools may be used for final responsiveness determinations without additional validation requirements beyond those applicable to traditional TAR. In that respect, Schulte continues the trajectory that began with Da Silva Moore: courts may be willing to accept newer review technologies when the underlying process is reasonable, transparent, and proportionate. 
   
   
2. Pre-culling with search terms remains a commonly accepted approach. The court’s approval of combining keyword search with AI review supports the use of multi-layered culling strategies designed to reduce costs and focus review efforts on a specific document population. This may be particularly relevant in matters involving large data volumes.
   
   
3. The “discovery on discovery” standard applies to AI review workflows. Parties using generative AI for review are not automatically subject to heightened disclosure obligations. Requesting parties must still demonstrate a specific deficiency before the court will order production of validation metrics, error rates, or other performance data.
   
   
4. Transparency and early disclosure remain important considerations. LinkedIn’s proactive disclosure of its AI methodology, including the tool used, the absence of training data, and its quality-control process, appears to have supported its position. Organizations may wish to consider early and appropriate disclosure of AI-assisted review processes consistent with applicable ESI protocols.
   
   
5. Proportionality remains central to discovery disputes. Across all three rulings, the court consistently grounded its analysis in Rule 26(b)(1)’s proportionality requirements. The burden and cost of proposed discovery – whether involving additional custodians, expanded data sources, or further disclosure of validation metrics – must be justified by a showing of likely benefit. Parties seeking expansive discovery bear the burden of demonstrating that the incremental information is both relevant and not reasonably available elsewhere.

Conclusion

Schulte v. LinkedIn may represent a milestone in the judicial treatment of generative AI tools in federal discovery practice, much as Da Silva Moore did for predictive coding and TAR more than a decade ago. By treating AI-powered document review as functionally equivalent to traditional TAR for purposes of the proportionality analysis, the court provides a roadmap for responding parties considering the use these tools in discovery. 

Organizations are encouraged to take note of the court’s emphasis on transparency, adherence to ESI protocols, and reasonableness – principles that have guided TAR jurisprudence since Da Silva Moore and that are likely to remain relevant as AI-enabled review technologies continue to evolve.

If you have questions about how this decision may affect your organization’s e-discovery practices, please contact the authors of this alert, your usual DLA Piper contact, or any member of DLA Piper’s eDiscovery and Information Management practice.