Join our mailing list to receive the latest updates and alerts Flag Subscribe

Key Takeaways

  • Address AI use and AI document review early. Discuss it with your client, e-discovery counsel, and e-discovery vendor in Rule 26(f) discussions and discovery planning, including negotiating both a protective order and an ESI order. Do not wait for a dispute to set the ground rules for AI-assisted review.
  • Generative AI that makes responsiveness calls is Technology Assisted Review (TAR). A federal court has held that a GenAI review platform is judged under the same reasonableness and proportionality standards that govern established TAR. There is no separate, heightened standard for AI.
  • Pre-culling with search terms before GenAI review remains proper. A party may use search strings to narrow the documents sent to a GenAI tool. A requesting party that objects should challenge whether the search terms are adequate, not the use of pre-culling itself.
  • Challenging an AI review method requires specific evidence of a deficiency. The size of the review population alone does not justify auditing the producing party’s AI process.

As generative AI moves from drafting and summarization into the responsiveness determination itself, a question has followed: Does a party that lets AI decide what is responsive owe the other side a special accounting? In Schulte v. LinkedIn Corp., Magistrate Judge Laurel Beeler said no. She treated a GenAI review platform used to make final responsiveness calls as a form of TAR and applied the same disclosure and proportionality principles that govern TAR. Schulte v. LinkedIn Corp., No. 22-cv-00237-HSG (LB), 2026 WL 1905851 (N.D. Cal. July 1, 2026). Thus, courts are handling AI review through the same discovery rules and ESI protocols litigants already use, rather than creating new ones.

The Dispute: GenAI Making Final Responsiveness Calls

Schulte is a putative antitrust class action. The plaintiffs allege that LinkedIn monopolized and attempted to monopolize through its handling of API access and data integration. On May 15, 2026, LinkedIn gave the plaintiffs 25 search strings and told them it would use Relativity aiR, a generative AI review tool, to assist in filtering out nonresponsive documents in its custodial production. When the plaintiffs asked for more information, LinkedIn disclosed that it used no seed or training set, that aiR was making the final responsiveness calls and that human reviewers were conducting quality control on samples taken from each responsiveness type. Schulte, 2026 WL 1905851, at *2.

The plaintiffs objected on two fronts. First, they argued that LinkedIn should be barred from using search strings to “pre-cull” the documents before running aiR, and should instead be required to run aiR across all custodial files. Second, they moved to compel disclosure of several validation metrics: elusion estimates, document error rates and the number of human reviewers validating aiR’s predictions. Schulte, 2026 WL 1905851, at *2. The court denied both requests.

The Court Treats GenAI Review as TAR

The heart of Schulte is its refusal to treat generative AI as a special case. The governing ESI protocol required the producing party only to disclose whether it intended to use TAR to filter out nonresponsive documents. The court held that Relativity aiR is a form of TAR. It found that LinkedIn’s disclosure, along with the additional information it provided on request, more than satisfied the protocol, and that nothing in the protocol required LinkedIn to disclose the metrics the plaintiffs demanded. Schulte, 2026 WL 1905851, at *2-3.

Notably, by placing GenAI responsiveness review within the existing TAR category, the court applied a body of law with settled expectations instead of inventing a new level of scrutiny for AI. It treated the demand for elusion rates, error rates and reviewer counts as a request for “discovery on discovery,” meaning discovery into the other side’s preservation and collection efforts. Courts disfavor that kind of discovery because it is typically not relevant to the merits and is rarely proportional. Schulte, 2026 WL 1905851, at *3. Such discovery is available only on a showing of a “specific deficiency,” and “mere speculation” is not enough. Id. (quoting Taylor v. Google LLC, 2024 WL 4947270, at *2 (N.D. Cal. Dec. 3, 2024)) The only deficiency the plaintiffs identified was the size of the review population, 204,444 documents, which the court found insufficient to justify an “audit” of LinkedIn’s use of aiR. Schulte, 2026 WL 1905851, at *3.

Thus, a producing party using GenAI for responsiveness review carries the same validation and defensibility burden that already applies to TAR: a reasonable, proportional process with human quality control. Absent a protocol provision or a concrete showing of deficiency, it carries no more than that. This is consistent with The Sedona Principles, which the court cited elsewhere in the same order for the point that the responding party is in a better position to decide what is relevant and proportional for its own data. Schulte, 2026 WL 1905851, at *6 n.50 (citing The Sedona Principles, Third Edition, 19 Sedona Conf. J. 1, 119-20 (2018)).

Put simply, demands for AI metrics ask more of GenAI review than parties have traditionally asked of manual review. In a conventional review, a producing party does not usually disclose how many documents each reviewer coded, its reviewers’ error rates or the recall of its review. It is expected to run a reasonable and proportional search for documents relevant to the claims and defenses, and courts generally do not second guess that effort without a specific reason to question it. See Schulte, 2026 WL 1905851, at *3. A demand for elusion estimates, error rates and reviewer counts asks a review that uses AI to meet a transparency standard that has not traditionally applied to human review, on the unstated premise that AI is presumptively suspect. Schulte declines that asymmetry. Thus, absent a protocol provision or a specific deficiency, detailed metrics about a GenAI review are likely not needed to show that the review was reasonable, any more than they would be for a team of human reviewers.

Pre-Culling With Search Terms Survives

The court also rejected the idea that a party must choose between search terms and AI review. Following other courts, it held that using search terms to pre-cull documents before providing them to a technology review platform satisfies the reasonableness and proportionality standards of Federal Rules of Civil Procedure 26(b) and 34(b)(2). Schulte, 2026 WL 1905851, at *2. The plaintiffs had not shown that LinkedIn’s 25 search strings were deficient (for example because they were too narrow) and they did not contend that they raised concerns when the strings were disclosed. Id. at *2-3. If they had concerns about the documents going to aiR, the remedy was to ask for changes to the search strings, not to ban pre-culling.

The court also noted the burden: The files of just two custodians totaled about 800 gigabytes and running aiR across all 19 custodians would have meant feeding multiple terabytes into the tool, with significant processing, hosting and review costs. Rather than prohibit pre-culling, the court ordered the parties to meet and confer within 21 days about whether the search strings should be adjusted. Id. at *3. In practice, that means a requesting party with concerns about search terms should raise them when the terms are disclosed, not after the review is underway.

Conclusion

Schulte signals that courts will fit GenAI review into existing TAR frameworks, and a later decision from the same district confirms that the ESI protocol sets the scope of what must be validated and disclosed. In In re Lyft, Inc. Passenger Sexual Assault Litigation, the court expressly distinguished Schulte and ordered Lyft to add a sample of documents excluded by its search terms to its TAR validation. In re Lyft, Inc. Passenger Sexual Assault Litigation, 2026 WL 2821947, at *2-3 (N.D. Cal. Sept. 20, 2026). The difference was the protocol. The Schulte ESI protocol required only disclosure of an intent to use TAR, while the Lyft protocol, built from provisions the parties negotiated and agreed to, called for validation across the entire collection against which search terms were run. Lyft argued that this language was a “placeholder,” and the court rejected that argument. Id. at *2. Thus, the extra validation in Lyft came from the parties’ bargain, not from any default rule

The takeaway for litigants is practical. A disclosed GenAI review, paired with search-term pre-culling and human quality control, can be reasonable and proportional without heightened, AI-specific disclosures. But what a party ultimately must validate and reveal will turn on the ESI protocol it negotiates. Parties that want broader validation or defined metric disclosure should bargain for it early so they can use these tools efficiently and avoid disputes once production is underway.

For more information or questions, contact Tim O’Shea.

Professionals

Jump to Page

Necessary Cookies

Necessary cookies enable core functionality such as security, network management, and accessibility. You may disable these by changing your browser settings, but this may affect how the website functions.

Analytical Cookies

Analytical cookies help us improve our website by collecting and reporting information on its usage. We access and process information from these cookies at an aggregate level.