Crowder v. LinkedIn Corp., No. 22-cv-00237-HSG (LB) (N.D. Cal. June 30, 2026), full opinion (PDF)
GenAI Review has reached the discovery motion stage. In an antitrust class action against LinkedIn, the plaintiffs asked the court to prohibit keyword culling ahead of LinkedIn's GenAI Review tool, to force that tool to review every custodial file, and to compel disclosure of the tool's validation metrics. Magistrate Judge Laurel Beeler of the Northern District of California denied all three requests, and her reasoning gives producing parties an early map for defending an AI review workflow.
What Happened
The plaintiffs allege that LinkedIn monopolized a market for professional social networking services and overcharged Premium subscribers, in violation of Section 2 of the Sherman Act. On May 15, 2026, LinkedIn gave the plaintiffs twenty-five search strings for its custodial documents and disclosed that it would use a generative AI review tool, "to assist in filtering out non-responsive documents." Asked for more detail, LinkedIn explained that no seed or training set was used, that generative AI makes the final responsiveness calls, and that human reviewers perform quality control on samples from each responsiveness type.
The plaintiffs called the disclosures inadequate and the keyword pre-culling "inappropriate." After an unsuccessful meet and confer, they asked the court to bar search strings ahead of GenAI Review, to compel GenAI Review of all nineteen custodians' files, and to order disclosure of elusion estimates, the document error rate, and the number of human reviewers validating generative AI predictions.
The Court's Analysis
The search-string ruling turned on a gap in the plaintiffs' argument. They asserted that pre-culling "artificially reduce[s]" the population the AI reviews, yet they never argued that LinkedIn's twenty-five search strings were deficient. "Courts have held that using search terms to pre-cull documents before providing them to technology-review platforms satisfies the reasonableness and proportionality standards of Rules 26(b) and 34(b)(2)," the court wrote, and it found the same logic applied here. Burden sealed the outcome. The files of two custodians alone total roughly 800 gigabytes, so running GenAI Review across all nineteen custodians would mean processing, hosting, and human review of multiple terabytes.
The metrics request failed as disfavored discovery on discovery. The court treated generative AI review as "a form of Technology Assisted Review" under the parties' interim ESI order, which obligated LinkedIn to disclose its intent to use TAR to filter out non-responsive documents and required nothing further. LinkedIn's disclosures "more than satisfy the demands of the Interim ESI Order." The only deficiency the plaintiffs identified, a target review population of 204,444 documents, was the kind of "mere speculation" that cannot justify an audit of another party's review process. The parties must still meet and confer on the search strings within twenty-one days, and the plaintiffs may return to the court if adjustments cannot be agreed. The same order also declined to add an in-house attorney as a document custodian and declined to compel production of text messages that the interim ESI order excluded.
Why It Matters
The ESI protocol decided this fight. LinkedIn's protocol required disclosure of TAR use and nothing more, so its early disclosure and voluntary detail left the plaintiffs with no obligation to enforce. The contrast with the same district two days later is instructive. In FTC v. Uber Technologies, Inc., No. 25-cv-03477-JST (TSH) (N.D. Cal. July 2, 2026), no transparency terms had been agreed, the FTC had doubted Uber's ten percent responsiveness rate from the beginning, and Magistrate Judge Thomas Hixson ordered Uber to produce a random sample of 300 non-responsive documents from the seed set used to train its TAR model. Uber's work product objection got no sympathy.
Read together, the two orders reward parties who settle AI review terms at the protocol stage. A producing party that disclosed its workflow early kept its validation metrics to itself. A producing party operating without agreed terms is now giving its opponent a window into its reviewers' coding calls. Crowder still leaves room for challengers, because a requesting party who can show a specific deficiency in the production, whether missing attachments, gaps against known documents, or search terms shown to be too narrow, retains a path to relief. GenAI review is earning deference the way TAR did a decade ago, one careful disclosure at a time.
The full opinion is available as a PDF.
