Court Declines to Give Generative AI Review Special Scrutiny, Treats It as TAR

A magistrate judge in the Northern District of California held that a party’s use of generative artificial intelligence (AI) in document review is subject to the same reasonableness and proportionality standards as any other review methodology, and that courts will not dictate how the technology is deployed absent a specific showing of deficiency. In Schulte v. LinkedIn Corp., LinkedIn’s use of Relativity aiR, a generative AI review tool, was challenged during discovery. The central point is that the court treated that tool as a form of Technology Assisted Review (TAR), not a novel technology requiring its own rules, and that framing guided each of the requests before it. Schulte v. LinkedIn Corp., No. 22-cv-00237-HSG (LB) (N.D. Cal. June 30, 2026).

Three points stand out:

• Generative AI review is being treated as a form of TAR, not as a novel review methodology requiring its own discovery rules. The court applied the same TAR disclosure obligation from the operative electronically stored information (ESI) protocol to the producing party’s use of a generative AI review tool, with no suggestion that generative AI requires a separate or heightened standard.
• Search-term pre-culling ahead of generative AI review is permissible. That has long been true ahead of traditional TAR as well, at least where the requesting party does not show that the search terms are deficient.
• The rule against “discovery on discovery” applies to generative AI review too. The court required a particularized showing of a specific production deficiency before it would compel disclosure of validation metrics, a rule that applies regardless of the review technology at issue.

Background

Schulte is a putative antitrust class action against LinkedIn pending in the Northern District of California. In discovery, pursuant to the ESI protocol governing the case, LinkedIn disclosed that it would use Relativity aiR to make responsiveness determinations, after first applying 25 search strings to narrow the universe of custodial documents provided to the platform. Plaintiffs asked the court to bar the pre-culling and instead compel LinkedIn to run Relativity aiR across all agreed-upon custodial files rather than just the culled subset, and separately to compel disclosure of additional validation metrics, including elusion estimates and reviewer counts. Magistrate Judge Laurel Beeler denied both requests.

Generative AI Review Is Treated as TAR

The court analyzed LinkedIn’s use of Relativity aiR under the same disclosure obligation that governs traditional TAR. The operative ESI protocol required LinkedIn to disclose whether it intended to use TAR to filter non-responsive documents, and the court treated LinkedIn’s notice that it would use Relativity aiR as satisfying that requirement. Nothing in the opinion suggests generative AI review calls for a separate or heightened disclosure standard.

For a producing party, this decision supports treating generative AI review the same as TAR for disclosure purposes under an existing ESI protocol. Whatever the protocol requires for TAR, whether that is more detailed disclosure or a specific protocol, that same requirement would likely apply to generative AI review. The information LinkedIn provided, including whether a seed or training set was used, whether the tool makes final responsiveness calls, and what human quality-control process supports it, was more than enough to satisfy the disclosure required under the protocol.

Search-Term Pre-Culling Ahead of Generative AI Review Is Permissible

The plaintiffs asked the court to bar LinkedIn from pre-culling with search terms and to compel LinkedIn to run Relativity aiR across all agreed-upon custodial files rather than just the culled subset. The court denied the request. It applied the same case law governing pre-culling ahead of traditional TAR review to hold that using search terms to narrow the document population before Relativity aiR review satisfies the reasonableness and proportionality standards of Rules 26(b) and 34(b)(2). Livingston v. City of Chicago, No. 16-cv-10156, 2020 WL 5253848, at *3 (N.D. Ill. Sept. 3, 2020); In re Biomet M2a Magnum Hip Implant Prods. Liab. Litig., No. 3:12-MD-2391, 2013 WL 1729682, at *2 (N.D. Ind. Apr. 18, 2013). To depart from that rule, the plaintiffs needed to show that LinkedIn’s search strings were themselves deficient or too narrow, and they never made that showing, arguing only that pre-culling itself was improper. The court ordered the parties to meet and confer on the search strings. The court also noted that requiring LinkedIn to process its full document set through Relativity aiR across all 19 custodians would be disproportionate, a practical point underscoring why pre-culling matters to the proportionality analysis in the first place: handling that volume would carry significant cost for processing, hosting, and human review.

No Discovery on Discovery Without a Specific Deficiency

The court also declined to compel disclosure of additional validation metrics, including elusion estimates and reviewer counts, for two independent reasons. First, LinkedIn’s disclosures already satisfied the ESI protocol’s TAR obligation, the same disclosure standard discussed above. Second, the court applied the general rule that discovery on discovery is disfavored absent a showing of a specific production deficiency, a rule that applies regardless of the review technology at issue. Taylor v. Google LLC, No. 20-cv-07956-VKD, 2024 WL 4947270, at *2 (N.D. Cal. Dec. 3, 2024). Plaintiffs pointed to the size of the target document population, but the court found that fact alone insufficient.

Across both rulings, the court’s inquiry remained focused on reasonableness and proportionality under Rule 26, not on approving or auditing the AI product itself. In our view, that is the right result. Generative AI is not a new species of technology requiring its own rules; it is another predictive engine that can be substituted into the same TAR workflow long used for traditional machine-learning review tools. The underlying process of scoping the review, training the model, applying it to classify documents, and validating the results does not change based on which predictive engine is doing the work. For a deeper look at how generative AI fits within the established TAR framework, see Tara Emory, Jeremy Pickens & Wilzette Louis, TAR 1 Reference Model: An Established Framework Unifying Traditional and Gen AI Approaches to Technology-Assisted Review, 25 Sedona Conf. J. 109, 117-18 (2024). It is a welcome development that courts are beginning to build case law consistent with that view.

© Arnold & Porter Kaye Scholer LLP 2026 All Rights Reserved. This Blog post is intended to be a general summary of the law and does not constitute legal advice. You should consult with counsel to determine applicable legal requirements in a specific fact situation.