A new ruling out of the Northern District of California did not generate headlines about a court approving the use of AI for legal data review. In fact, that debate was never really up for discussion. Notably, that is exactly what makes it worth your attention.
In Schulte v. LinkedIn Corporation (N.D. Cal. 2026), Magistrate Judge Laurel Beeler issued a discovery order addressing three separate disputes, one of which directly concerned LinkedIn’s use of Relativity aiR for Review for their production.
The plaintiffs raised procedural objections to how LinkedIn was using aiR. They challenged LinkedIn’s decision to apply search strings to cull documents before running the data set through aiR, and they moved to compel additional disclosures about aiR’s performance metrics. The court denied both motions.
What the plaintiffs did not do – and what neither party nor the court treated as remotely controversial – was challenge LinkedIn's use of aiR to make final responsiveness calls.
The Dispute, In Brief
LinkedIn disclosed its use of aiR to the plaintiffs on May 15, 2026, alongside 25 search strings it intended to apply before feeding documents into the platform. When the plaintiffs pushed back, LinkedIn provided additional information: no seed or training set was used, aiR was making final responsiveness determinations, and quality control was being handled through human review of samples drawn from each responsiveness category. LinkedIn maintained it had no obligation to say more.
The plaintiffs disagreed on two fronts. First, they argued that culling with search strings “artificially reduces” the document population available to aiR, potentially excluding responsive material. Second, they sought disclosure of specific performance metrics – elusion estimates, document error rates, and the number of human reviewers validating aiR’s predictions.
Judge Beeler denied both requests as neither reasonable nor proportional to the needs of the case.
On Early Culling: The Court Backs a Familiar Workflow
The court’s reasoning on the search-string question is straightforward and grounded in established case law. Using keyword search to create a target population before applying a technology-assisted review workflow is not a new practice, and courts have consistently found it satisfies the reasonableness and proportionality standards of Rules 26(b) and 34(b)(2). Judge Beeler applied the same analysis here.
Critically, the court noted that the plaintiffs never actually challenged the adequacy of LinkedIn’s 25 search strings when they were first disclosed. Their concern about early culling was procedural, not substantive – and without a showing that the strings were too narrow or otherwise deficient, the argument didn’t have legs. The court pointed out the appropriate remedy: if the plaintiffs believed the search strings were leaving responsive documents out of aiR’s reach, they could address that concern by negotiating adjustments to the strings themselves.
The practical lesson extends well beyond this case. The producing party’s workflow – search strings to narrow the field, aiR to make responsiveness determinations, human sampling for quality control – is a recognizable, defensible architecture that follows an accepted TAR workflow. It maps cleanly onto the kind of validated, proportionate approach that courts and practitioners have been building toward for years. When that workflow is documented and disclosed in good faith, it holds up.
In fact, the court described aiR as “a form of technology-assisted review” within the ruling –accepting that such generative AI tools are simply a natural progression from early-generation TAR tools that courts have approved for use over the last decade.
On Metrics: "Discovery on Discovery" Remains a High Bar
The plaintiffs’ request for aiR performance metrics ran into an equally firm wall, though for somewhat different reasons.
Judge Beeler framed the request under the well-established “discovery on discovery” doctrine – the principle that probing a producing party’s evidence collection and review methodology is generally disfavored unless there’s a specific, demonstrated deficiency in the production itself (see: Taylor v. Google LLC, No. 20-CV-07956-VKD, 2024 WL 4947270, at *2 (N.D. Cal. Dec. 3, 2024)). Suspicion alone isn’t enough. The only concrete concern the plaintiffs identified was the size of LinkedIn’s target document population: 204,444 documents. The court found that number, standing alone, insufficient to justify a deeper audit of aiR's performance.
LinkedIn had already disclosed the platform it was using, confirmed that no training set was required, explained that aiR was making final calls, and described its human QC process. That level of transparency exceeded what the Interim ESI Order required: Paragraph 5(a) of the Order requires the producing party to “disclose to the receiving party if they intend to use Technology Assisted Review (“TAR”) to filter out non-responsive documents.” Without evidence that something had gone wrong in the production, the court wasn’t going to require LinkedIn to open the hood further.
This holding doesn’t mean that metrics like elusion rates and error rates are irrelevant to defensible AI-assisted review – they are very relevant, and practitioners building robust validation workflows know their value well. What it means is that opposing counsel can’t compel those disclosures on the basis of speculation, just like with traditional TAR. The burden remains on the producing party to validate their work as they see fit to meet their discovery obligations, and on the requesting party to show a real problem, if they feel there is one.
Courts Are Asking Different Questions Now
In a case where sophisticated plaintiffs’ counsel chose not to challenge the use of aiR for Review as a final decision-maker, and a federal magistrate judge accepted that use without comment, we see evidence that some baseline assumptions around AI for discovery have changed.
Generative AI for legal data review is actively in use, and the simple fact of that use is uncontested. The questions courts are being asked to resolve are operational:
- How was the document population constructed?
- What did you disclose?
- Was the workflow proportionate?
Those are the same questions practitioners have been answering about TAR for more than a decade.
For legal teams making strategic decisions about their review workflows, the conversation has moved from whether the use of AI is defensible altogether and onto how to apply modern review protocols to generative AI review.
What This Means for Your Practice
A few practical takeaways worth carrying into your next review project:
- Disclose pertinent information proactively and document thoroughly. LinkedIn’s disclosures – platform, methodology, QC approach – exceeded what the Interim ESI Order required and gave the court a clear picture of a reasonable process. That documentation became the foundation of the defense.
- Build QC into the workflow by design, not as an afterthought. The human review component of LinkedIn’s process – sampling from each responsiveness category – reflects the meaningful human oversight that courts and regulators are increasingly looking for in AI review workflows.
- Know your validation story before you need it. The plaintiffs didn’t get the metrics they sought here, but that won’t always be the case. Being able to articulate your elusion rates, precision, and recall – and how you evaluated them – puts you in a much stronger position in any dispute.
Schulte v. LinkedIn offers a window into how courts are treating legal data review workflows using AI as a routine matter of proportionality and process. How you build and document that process is what will matter most going forward – whether you’re defending a production or planning one.
Graphics for this article were created by Kael Rose.
Recommended for you
