Federal Court Accepts Generative AI for eDiscovery: What the LinkedIn Decision Means for Legal Teams

The most consequential fact about the discovery order Judge Laurel Beeler signed on June 30 is what it does not contain. There is no new framework for artificial intelligence in litigation, no special hearing on model reliability, and no separate disclosure regime for generative AI. A party proposed to let a generative AI system make the final first pass responsiveness determinations on roughly two hundred thousand documents in a federal antitrust case, the opposing side objected, and the court resolved the dispute with the discovery doctrine it already had: reasonableness, proportionality, and the settled reluctance to allow discovery about discovery.

For years the public story of AI in law has been dominated by hallucinated case citations and sanctioned filings. Schulte v. LinkedIn is the other story, the one that matters to businesses: enterprise AI, deployed inside a structured workflow, disclosed to the other side, challenged, and defended successfully under ordinary rules. Commentators including retired Magistrate Judge Andrew Peck, whose 2012 opinion in Da Silva Moore v. Publicis Groupe first approved Technology Assisted Review, have described Schulte as the first federal decision to accept generative AI making the final responsiveness call in discovery.

What LinkedIn Actually Did

The workflow matters, because it went further than AI helping lawyers read faster. On May 15, 2026, LinkedIn gave the plaintiffs the 25 search strings it intended to run against its custodians' documents, and disclosed that it would use Relativity aiR, a generative AI document review tool, to filter nonresponsive documents from the results. When the plaintiffs asked for more detail, LinkedIn disclosed three specifics that define the case. No seed or training set was being used. Relativity aiR would make the final responsiveness calls. Human quality control would consist of reviewing samples drawn from each responsiveness category.

In other words, the AI was not a suggestion engine with a lawyer approving every determination. It was incorporated directly into the first pass decision process, with people auditing its output by sampling rather than re-reviewing it. After the search strings were applied, the population presented to aiR for review came to approximately 204,444 documents.

The scale explains the design. The files of just two of LinkedIn's 19 designated custodians totaled roughly 800 gigabytes; running generative AI review across every custodial file would have meant processing multiple terabytes, with the associated processing, hosting, and human review costs. That burden argument became central to the ruling.

What the Plaintiffs Wanted

The plaintiffs asked the court for three things: an order preventing LinkedIn from using search terms to cull the document population before AI review, a requirement that aiR be run across all agreed custodial files, and disclosure of additional metrics about aiR's use and validation, including elusion estimates, document error rates, and the number of human reviewers validating the system's predictions. Their theory on culling was that filtering by search string could artificially shrink the population the AI ever saw, eliminating responsive documents before review began. On the size point, they argued that a 204,444 document population justified closer scrutiny of the methodology.

Notably, as several analyses of the order have observed, the plaintiffs did not challenge the 25 search strings themselves as deficient or too narrow, and they did not challenge the underlying proposition that aiR could make final responsiveness determinations at all. The fight was about process around the AI, not permission for it.

What the Court Held

The court denied all three requests. On pre-culling, Judge Beeler noted that courts have permitted search terms to narrow document populations before Technology Assisted Review, citing Livingston v. City of Chicago and In re Biomet M2a Magnum Hip Implant Products Liability Litigation, and applied the reasonableness and proportionality standards of Federal Rules of Civil Procedure 26(b) and 34(b)(2). Because the plaintiffs had not shown the search strings were deficient, culling before AI review presented no responsiveness problem, and requiring aiR to process the entire multi terabyte custodial universe would impose a substantial and disproportionate burden.

On the disclosure requirement, the operative Interim ESI Order obligated a producing party to disclose any intent to use Technology Assisted Review to filter nonresponsive documents. The court expressly characterized Relativity aiR as "a form of technology-assisted review" and held that LinkedIn's disclosures satisfied the order. That characterization is the heart of the decision. TAR has existed for over a decade through predictive coding and machine learning review, with its own accepted case law. Rather than treating generative AI as a novel category demanding new doctrine, the court treated it as the next generation of an established methodology, governed by the same concepts of reasonableness, proportionality, disclosure, validation, and production quality.

On the requested metrics, the court applied the established rule that discovery on discovery is disfavored absent evidence of a specific deficiency in the producing party's collection or production. Speculation is not enough, and the size of the review population did not by itself establish that anything was wrong with LinkedIn's production. The court did not hold that AI validation metrics are never discoverable. It held that using generative AI does not automatically entitle the other side to an audit of the methodology.

Nor did LinkedIn receive unlimited freedom. The court ordered the parties to meet and confer within 21 days about the search strings and whether adjustments were warranted, leaving the plaintiffs a path back to court if they identified actual problems the parties could not resolve. Search methodology remains fully challengeable on evidence.

The Technology, Briefly and Neutrally

According to Relativity's documentation, aiR for Review uses large language models and natural language processing to analyze document text against user developed prompt criteria, classifying documents as relevant, not relevant, or borderline, and providing rationale with citations to the passages supporting each conclusion. Relativity markets the product for responsiveness review, issues review, key document identification, and related analyses. Those are vendor descriptions, not findings. The court did not independently validate aiR's accuracy, and nothing in the order should be read as judicial certification of the product. What the court evaluated was whether LinkedIn's methodology, and the plaintiffs' demands for more, were reasonable and proportional on this record.

Why This Matters to Law Firms

For firms weighing AI assisted review, the practical significance is defensibility. There is now a federal decision demonstrating that a generative AI review workflow, properly structured and disclosed, can be defended under existing discovery principles rather than fought over as a novelty. The structure is the lesson: defined search criteria disclosed in advance, an ESI protocol addressing TAR, documented AI responsiveness determinations, human quality control by sampling, and formal disclosure to the opposing party. Every element of that design remained essential to the outcome, and counsel's responsibility for a defensible process is undiminished. A firm that treats this ruling as permission to skip validation and documentation has read it backwards.

Why This Matters to Corporate Legal Departments

For the businesses that pay for litigation, this is an efficiency story about the single most labor intensive line item in large cases. Document review in major corporate litigation routinely spans millions of emails, chat messages, and files, and traditional eyes on review of that volume is enormously expensive. LinkedIn's numbers illustrate the stakes: 800 gigabytes for two custodians, 19 custodians in total, terabytes in the full universe. AI review aimed at reducing the population requiring human eyes is being evaluated precisely where the money is. The order does not say what LinkedIn saved, and no responsible reading of it produces a savings figure. What it does say is that a court allowed the economics of proportionality to matter, declining to force the more expensive process without evidence the cheaper one was deficient.

What the Decision Does Not Say

Restraint is warranted, because the order is narrow. It does not establish that generative AI review systems are reliable in general, that Relativity aiR is accurate, that search term pre-culling is always appropriate, or that validation metrics can never be discovered. It does not hold that human review is unnecessary, that AI has replaced attorneys in discovery, or that courts everywhere have accepted generative AI review. It is a discovery order from a magistrate judge in one district, binding no other court, and it is fact specific: had the plaintiffs shown the search strings were too narrow or identified a concrete deficiency in the production, the analysis could have run the other way. A producing party with a genuinely deficient AI process cannot hide behind this ruling, and the meet and confer requirement shows the court keeping that door open.

The Bigger Picture for Business AI

Read from this publication's vantage point, Schulte is a case study in what mature enterprise AI adoption looks like: defined inputs, an established protocol, machine determinations bounded by human quality control, documentation, disclosure, and review by an authority applying existing rules. That is the same supervised pattern this publication has described across business operations and legal practice, here surviving contact with an adversary and a federal judge. The decision does not establish that generative AI can be trusted blindly in litigation. It demonstrates something more useful: that at least one federal court was willing to evaluate a generative AI review process under the same principles that already govern technology assisted discovery, which removes one real element of uncertainty from the adoption decision. The technology still requires controls, validation, and defensible procedure. But in legal discovery, generative AI has begun the unglamorous passage from experiment to infrastructure.