TL;DR
On June 30, 2026, a federal court approved generative AI making final document-review decisions for the first time, without writing new law to get there. Judge Beeler denied all three plaintiff challenges by applying a single clause in the parties’ existing ESI protocol requiring disclosure of Technology Assisted Review, and found LinkedIn’s disclosure satisfied it in full. The real lesson centers on what your ESI protocol says about AI review tools before a dispute ever starts.
The strongest position in a discovery fight over AI review is a clause you negotiated months before the fight started.
A federal court approved generative AI for making final discovery decisions by applying language the parties had already agreed to. On June 30, 2026, Magistrate Judge Laurel Beeler of the Northern District of California signed a twelve-page order in Schulte v. LinkedIn Corp. resolving three discovery disputes at once. The opinion reads like a judge applying an existing clause to a new tool.
What Happened in Schulte v. LinkedIn?
Schulte is a putative antitrust class action alleging that LinkedIn monopolized and attempted to monopolize the market in ways that enabled it to overcharge Premium subscribers. On May 15, 2026, LinkedIn gave plaintiffs twenty-five search strings for its custodial documents and disclosed that it would use Relativity aiR, a generative AI review tool, to filter out non-responsive material. Pressed for detail, LinkedIn explained that no seed or training set was used, that aiR would make final responsiveness calls, and that human reviewers would sample from each responsiveness category for quality control.
Plaintiffs raised three challenges. They asked the court to bar the search-string culling and compel LinkedIn to run aiR across the full custodial population. They asked for additional validation metrics, including elusion estimates and reviewer counts. They asked the court to apply heightened scrutiny to an AI tool serving as final decision-maker on responsiveness. Judge Beeler denied all three.
Why Didn’t the Court Need to Write New Law?
The court applied Paragraph 5(a) of the parties’ Interim ESI Order, which required a producing party to disclose whether it intended to use Technology Assisted Review to filter out nonresponsive documents. Judge Beeler characterized aiR as a form of technology assisted review and held that LinkedIn’s disclosures more than satisfied the order. That single clause carried the entire disclosure analysis. No new framework, no special category for generative AI, no heightened standard.
Retired Magistrate Judge Andrew Peck, who authored Da Silva Moore and opened the TAR era more than a decade ago, described Schulte, with DLA Piper colleague Reema Holz, as the first federal decision to accept generative AI in a final responsiveness role. Their reading matches the order. Review technology keeps moving. The reasonableness and proportionality standard courts apply to it has stayed stable, and Schulte shows those two speeds connecting when the protocol language does its job. The same pattern appeared in the protective order rulings from earlier this year, where courts governed AI use through existing case-management tools.
What Made LinkedIn’s Disclosure Enough?
The disclosure went further than Paragraph 5(a) technically required, and that extra specificity appears to be what made the ruling straightforward. Four elements did the work: the platform name, confirmation that no seed or training set was used, a clear statement that aiR made final responsiveness calls, and a description of the human sampling behind it. A party disclosing only that it plans to use AI-assisted review leaves a court guessing about scope, training data, and human oversight. That gap is where objections like the ones plaintiffs raised here tend to succeed.
Why Did the Plaintiffs’ Other Two Arguments Fail?
The metrics request failed during discovery. Judge Beeler applied the standard from Taylor v. Google LLC, under which such discovery is disfavored and requires a showing of specific deficiency before a court will order it. The only deficiency plaintiffs identified was the size of the review population after the search strings ran, 204,444 documents. Volume alone amounted to speculation, and speculation does not justify auditing another party’s review process.
Scale drove the pre-culling ruling. Two custodians alone held more than 800 gigabytes of data, and running aiR across all nineteen designated custodians would have meant feeding multiple terabytes through the platform, with processing, hosting, and human review costs to match. Absent evidence that the existing methodology was producing deficient results, the court found that the exercise was disproportionate. The parties were directed to meet and confer on the search strings within twenty-one days.
Something the plaintiffs did not challenge deserves attention. Nobody questioned LinkedIn’s use of aiR to make final responsiveness calls. Neither party treated it as controversial, and the court did not either. Every argument in the case concerned workflow and disclosure. For attorneys tracking whether generative AI review has become defensible, that silence carries more weight than any holding in the order.
What Should Your ESI Protocol Say About AI Review?
Four elements from LinkedIn’s approach are worth building into every ESI protocol going forward:
- Write the clause broadly enough to cover any AI tool. Require disclosure of any AI-assisted review tool and the role it plays: triage, prioritization, or a final responsiveness call.
- Define what disclosure means in practice. Require the platform name, whether a seed or training set is used, the tool’s role in the final call, and the human quality-control process behind it.
- Decide on validation metrics before a dispute starts. This ruling confirms courts will decline to compel elusion rates or reviewer counts absent a specific problem. Negotiate that visibility into the protocol itself if it matters to your case.
- Revisit protocols drafted before generative AI tools existed. Confirm that the existing TAR language covers newer AI-assisted workflows and leaves no open questions for a judge to answer later.
One caution on scope. The order is nonprecedential and resolves one protocol dispute on a single record. It binds no other court. What it offers is a map of how a thoughtful judge reasoned through the question, and a citation worth having when opposing counsel demands an audit of your review process.
How Should Attorneys Prepare for the Next Dispute Like This One?
The next dispute over AI-assisted review will likely turn on whose ESI protocol has already answered the question before the court. LinkedIn had that answer ready before plaintiffs raised a single objection. That is the position worth building toward now, ahead of a similar dispute in one of your own matters, and it pairs with knowing which AI tools your own client is using.
Complete Legal helps clients build that language into their protocols before a live dispute tests it. A second look at your current ESI protocol, focused on whether it covers how your review technology works today, is worth the time before the next production deadline forces the question. Waiting until opposing counsel raises the objection puts you in a reactive posture for the rest of the case. Getting the language right at the outset keeps that argument from happening at all.
Want a second read on your ESI protocol?
📧 info@completelegal.us | 📞  888-885-9909Â
Key Takeaways
- Judge Beeler denied all three plaintiff challenges by applying Paragraph 5(a) of the parties’ existing ESI protocol, which required disclosure of any intent to use TAR to filter nonresponsive documents, and found LinkedIn’s disclosure of its Relativity aiR workflow satisfied that clause in full.
- Courts will decline to compel AI validation metrics like elusion rates and reviewer counts absent a specific production deficiency. A review population of 204,444 documents was not enough on its own, and scale cut the other way on pre-culling, where 19 custodians would have meant multiple terabytes through the platform.
- Nobody challenged the use of generative AI to make final responsiveness calls. The order is nonprecedential, and specific, tool-agnostic disclosure language in an ESI protocol is what let new review technology get absorbed by existing rules.
