When the AI Is the Reviewer: Schulte and the Emerging Rules for Generative AI in Document Review

A recent discovery order from the Northern District of California in Schulte v. Linkedin (Schulte) one of the first judicial decisions concerning the use of generative AI as the reviewer in civil discovery, and it speaks to an emerging issue in discovery: Can generative AI be effectively used in discovery, and what will need to be disclosed about the process?[1]  

The answer, for now, is reassuring for producing parties. The court treated generative AI review like established review technology, applied familiar proportionality principles, and refused to let the opposing party audit the AI processes. But the decision leaves open questions, including about testing, prompts, and the AI’s written output, that will be important to consider as AI becomes a more regularly utilized tool in discovery.

Generative AI and TAR in E-Discovery

For over a decade, litigants have used Technology Assisted Review (TAR or predictive coding) to manage large document populations. Traditional TAR works by having attorneys code a sample set of documents (a training or seed set). The technology then learns from those judgments and prioritizes the rest, and humans review what the software surfaces. Courts have accepted TAR as an established document review protocol, with one court in the Southern District of New York observing that “it is now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it.”[2] Unlike generative AI review, however, TAR presumes an attorney reviewer at the center of the review process: TAR prioritizes the document for the attorney to review and then the attorney makes a judgment call on responsiveness and privilege.

Generative AI can, in certain circumstances, replace the first level review from an attorney. For example, in Schulte, the defendant disclosed that its AI tool (Relativity aiR) was “being used to make final responsiveness calls,” subject to quality control.”[3] In effect, the AI was not simply surfacing the document for an attorney to review, it was making the decision on responsiveness itself. To ensure that the AI was making accurate calls, counsel were then conducting “quality control” through a “human review of samples taken from each responsiveness type.”[4]

Notably, neither the plaintiffs nor the court questioned whether this was permissible. The court simply applied the disclosure rule the parties had negotiated for TAR. While that issue was not the subject of a ruling, the fact that it was not disputed is a sign that courts and litigants may be coming to accept that generative AI can substitute for attorney review of documents when subject to appropriate testing.

What the Court Held

While the plaintiffs did not challenge the defendant’s use of AI, they made two requests regarding its use of AI in the review process. The court denied both requests.

Keyword Filtering. The defendant ran 25 disclosed search terms to narrow the document population before conducting an AI-based review. The plaintiffs asked the court to prohibit that filtering, arguing it “artificially reduce[d]” the review population and would effectively “remov[e] responsive documents from the target population.”[5] The court denied that request, holding that “using search terms to pre-cull documents before providing them to technology-review platforms satisfies the reasonableness and proportionality standards” under the Federal Rules of Civil Procedure.[6] The court observed that prohibiting the defendant from using search terms to narrow the “target document population would impose a substantial burden” because it would require the company to “feed multiple terabytes of documents into Relativity aiR, resulting in significant costs related to processing, hosting, and human review.”[7] The court noted that the plaintiffs had not made any claim that the search terms were “too narrow” and noted that if they had, then the concern about the search terms “might be warranted.”[8]

Disclosure of Metrics. The plaintiffs demanded that the defendant produce various metrics about its use of Relativity aiR, including “elusion estimates” reflecting the rate of responsive documents that the AI did not code as responsive.[9] In effect, the plaintiffs were requesting the output of the company’s quality control review. The court denied this request. The defendant had disclosed to plaintiffs that its “quality control review is being conducted by human review of samples taken from each responsiveness type” and the court held that this disclosure was sufficient.[10] The court held that this disclosure “more than satisf[ies] the demands of the Interim ESI Order” and the plaintiffs had not made a showing that any “audit” of the defendant’s use of Relativity aiR was warranted.[11] In effect, the court put the burden on the plaintiffs to identify something “deficient” about the productions to warrant the disclosure of these metrics, citing the principle that “discovery on discovery is disfavored” and is only warranted if there is a “specific deficiency” that goes beyond “mere speculation.”[12] This is a notable holding because it put the burden on the plaintiff to identify the deficiency in the production in order to gain access to these quality control metrics. However, as noted below, this may hinge on having an agreed-upon ESI protocol in place.

The Key Questions: Metrics/Samples, Prompts, and AI Outputs

While this was a notable ruling, there are several key questions that remain open.

Where Metrics or Samples May Be Required. While the Schulte court declined to require the disclosure of metrics as part of what the court referred to as an “audit,” there are precedents from the TAR context where courts have required more comprehensive disclosures of the process that goes into the TAR review. Notably, on July 2, 2026, a court in the same district required the producing party to produce a sample of the documents that had been marked as non-responsive in the TAR sample set.[13] There, the court held that, in light of the Federal Trade Commission’s objections to the scope of documents being produced, the company should produce a “random sample of the training documents” that the document reviewers had tagged as non-responsive as this would provide insight into how the TAR was trained.[14] Whereas in Schulte the disclosure was limited to the obligations specified in the ESI protocol, here the court permitted discovery into the underlying process used by the producing party. Historically, the TAR cases have been “split” on the scope of these disclosure issues and it is likely that, absent an ESI protocol, courts may reach different conclusions on when discovery on the production process is appropriate.[15]

Prompts. Generative AI review is directed by written instructions, referred to as “prompts,” which counsel draft to instruct the AI on the subject matter of the case and the responsiveness criteria. These prompts can be quite detailed, akin to the review protocol that may be drafted for attorney reviewers. While document review protocols are attorney work product and generally not subject to disclosure, it remains to be seen if courts will apply that same logic to the prompts provided to the AI or may require some degree of disclosure of the prompt.[16]

AI Outputs. Generative AI can produce numerical and written explanations for each document decision. For example, Relativity aiR provides not only a “score” that “indicates how strongly relevant the document is,” but also an “explanation” of the outcome (the rationale) and an output where the AI “is asked to argue against itself, identifying what assumptions it’s making or additional facts that could reverse the prediction” (the consideration).[17] These outputs are not directly drafted by the attorney but are the outcome of the attorney-directed process. It remains to be seen whether courts may mandate the disclosure of these types of outputs in certain circumstances.

Practical Takeaways

AI-based technology is moving into the review process, and Schulte provides a helpful signal that courts will accept it and may, in certain circumstances, push back on demands for “discovery on discovery.” However, the TAR cases show that courts may be willing to permit some discovery that looks into how the technology is being utilized. Litigants will need to carefully evaluate the pros and cons of using AI in the discovery process and consider whether they want to use it to augment or replace human review. Importantly, what the ESI protocol says in a case about parties’ ability to use AI will also govern what other parties are able to do and what the disclosure requirements are for all parties. Being able to effectively utilize AI will rely on taking deliberate steps throughout the discovery process, including:

• Determine how AI fits in the review process. Schulte illustrates how AI can be used as a substitute for an attorney reviewer. It can also be used to assist reviewers in finding responsive documents, similar to more traditional TAR. Counsel should determine at the outset of the litigation if and how AI may be beneficial in the review process.
• Consider the ESI protocol. Schulte underscores that where there is an ESI protocol in place and agreed to by the parties, courts will defer to it. Once a litigant has a view on the use of AI in the litigation, and how much disclosure they want to permit, they should seek agreement on that in the ESI protocol.
• Disclose AI usage early and transparently. Prompt, affirmative disclosure of AI usage, with reasonable voluntary detail, was key to defeating the “audit” demand in Schulte. Courts may defer to the producing party’s choice of technology in the review process if there is early and prompt disclosure of the tool.
• Treat prompts as attorney work product. Prompts should be prepared by counsel and appropriately labeled as attorney work product. Counsel should exercise caution before any disclosure or declaration that provides insight into the prompts as that could open the door to a waiver argument.
• Establish the quality control framework. Even if the disclosure about quality control is general in nature, it is important to establish a quality control framework to review AI output through attorney review of representative samples. That review can be measured through various metrics, including but not limited to (i) the percentage of documents that are coded as non-responsive, but that are in fact responsive (the elusion rate) and (ii) the percentage of documents coded as responsive that were truly responsive (the precision rate). Those types of metrics can establish a strong basis for quality control and, if necessary, counsel can update the prompt to improve its accuracy.

Investigations & White Collar Defense associate William Farrell contributed to this client alert.