A Northern District of California court just denied a bid to force disclosure of how LinkedIn used generative AI to review documents — the contested mirror image of last week’s stipulated AI protocol in James v. Cerebras. The lesson for litigators: the transparency you want from an opponent’s AI review is something you negotiate into the ESI protocol up front, not something you win on a motion after the review is done. Reading time approximately 11 minutes.
Case: Schulte v. LinkedIn Corp., No. 22-cv-00237 Court: Northern District of California | Decision: Discovery Order on ECF Nos. 186, 188, 190 — all three motions denied | Judge: Magistrate Judge Laurel Beeler | Signed: June 30, 2026; Filed: July 1, 2026
Read the full order on Minerva26 →
By Kelly Twigger
Podcast | Transcript
The Week After the Milestone
Last week I broke down James v. Cerebras Systems — the first stipulated ESI protocol I have seen that governs generative AI as its own category of document review, where two sophisticated parties simply agreed to hand over their prompts, their validation math, and their elusion rates. The question that ran through that whole episode was: why would anyone voluntarily take on obligations the rules never put on them?
This week gives us the other half of the picture, and it is every bit as important — because this week, a party did not agree. In Schulte v. LinkedIn Corporation, one side asked a federal court to compel essentially the same AI transparency the Cerebras parties gave away by stipulation. And Magistrate Judge Laurel Beeler said no.
Hold onto the throughline for the entire post: what you can negotiate for in an ESI protocol and what a court will order over an opponent’s objection are two very different things, and the gap between them is enormous. If you want visibility into how the other side is using AI to review documents, you build it into the protocol at the outset. You will not get it later by motion. This order is the proof.
Case of the Week
What the case is about
The underlying dispute is an antitrust class action. LinkedIn Premium subscribers allege that LinkedIn monopolized and attempted to monopolize in violation of Section 2 of the Sherman Act, overcharging Premium subscribers through two practices: giving potential rivals access to LinkedIn’s private user data through APIs on the condition that they not compete, and integrating LinkedIn’s user data with parent company Microsoft’s Azure cloud, tying up scarce hardware and driving up prices.
You do not need the antitrust theory to use this order. But keep one fact in view: the defendant is a Microsoft-owned data business sitting on an enormous volume of ESI — and that volume does real work in the Court’s proportionality analysis across all three rulings.
Three discovery letter briefs were before the Court. The Court denied all three. The requesting party — the plaintiffs — went zero for three, on (1) LinkedIn’s use of the generative AI tool Relativity aiR and the search terms it used to cull the population before that review; (2) a motion to add an in-house lawyer as a custodian; and (3) a motion to compel text messages from nineteen custodians.
Be precise about the technology
Relativity aiR is Relativity’s generative AI review product — it uses large language models to make responsiveness and privilege calls and produces a rationale for each. That is not traditional technology assisted review. Traditional TAR, which we have well over a decade of case law on going back to Da Silva Moore in 2012, trains a classifier on a human-coded seed set and ranks documents by likely responsiveness. Generative AI review does not necessarily use a seed set at all.
The tell is in this record. LinkedIn disclosed that it used no seed or training set, that aiR was making the final responsiveness calls, and that human review was limited to quality-control sampling of each responsiveness category. So this is genuine generative AI review — making final calls, with humans checking samples rather than every document. Remember that, because the Court is about to treat it as ordinary TAR, and that move is the whole ballgame
Ruling one: pre-culling is proper, and you can’t audit the AI on speculation (ECF No. 186)
There are two fights bundled here, and they need to stay apart.
The pre-cull. LinkedIn gave the plaintiffs twenty-five search strings, said it would run them first and feed the survivors into Relativity aiR. The plaintiffs asked the Court to prohibit that pre-cull and to compel LinkedIn to run aiR across all custodial files, arguing the search strings “artificially reduce” the population aiR ever sees, so responsive documents without a search term get stripped out before the AI can find them.
That is the layering problem we covered in Cerebras — stacking two culling methods so a document must clear both, with each method independently missing some responsive material. In Cerebras, the parties handled it by agreeing to disclose the intent to layer, meet and confer, compare hit counts with and without layering, and let the requesting party sample the excluded set.
Here is the contested answer. The Court denied the request, and the reasoning is the practical core of the order. The plaintiffs argued in the abstract that pre-culling is improper — but never showed the twenty-five search strings were actually deficient. If they had shown the strings were too narrow, the Court said, their concern “might be warranted.” They did not, and they had not even raised concerns when the strings were disclosed on May 15. The Court then held flatly that using search terms to pre-cull before handing documents to a review platform satisfies the reasonableness and proportionality standards of Rules 26(b) and 34(b)(2), citing Livingston v. City of Chicago and In re Biomet M2a Magnum Hip Implant Products Liability Litigation. Pre-culling is normal, not improper.
Then the proportionality hammer: two custodians alone total roughly 800 gigabytes; across nineteen custodians, running aiR on everything with no pre-cull would mean feeding multiple terabytes into the tool. The Court denied the request and ordered the parties to meet and confer within twenty-one days about the search strings, with leave to return if specific adjustments cannot be agreed.
The lesson in miniature: challenge the search strings, with a specific showing, when they are disclosed — not the concept of pre-culling, in the abstract, months later.
The demand for aiR’s metrics. Separately, the plaintiffs moved to compel LinkedIn to disclose elusion estimates, the document error rate, and the number of human reviewers validating the AI. Look at that list: it is nearly the exact set of disclosures the Cerebras parties agreed to hand over in Appendix 4. The Schulte plaintiffs asked a court to order what the Cerebras parties volunteered.
The Court said no, using the doctrine you need to know cold: “discovery on discovery” — discovery into how the other side ran its search and collection, rather than of the underlying facts. Under Taylor v. Google LLC, it is disfavored — typically not relevant to the merits, rarely proportional — and available only on a showing of a specific deficiency, not “mere speculation.” The Court denied the request for two reasons. First, LinkedIn had satisfied the parties’ Interim ESI Order, whose Paragraph 5(a) required only that a producing party disclose if it intends to use Technology Assisted Review to filter out non-responsive documents. LinkedIn disclosed it was using aiR — “a form of Technology Assisted Review” — and answered follow-up questions; that “more than” satisfied the order. Second, the only “deficiency” the plaintiffs identified was that the target population came to 204,444 documents, and a raw count, without more, is not a specific showing.
Sit with the move the Court made, because it runs in the opposite direction from Cerebras. LinkedIn used a generative AI tool making final calls with no seed set, and the Court folded it under the TAR umbrella for the ESI order’s TAR-disclosure clause. One week after a Northern District of California magistrate judge deliberately broke AI out of the TAR category and gave it a separate rulebook, another Northern District of California magistrate judge folded genAI review right back in and applied a decade-old TAR standard. Same district, same technology, opposite instincts, one week apart. Is generative AI review just TAR by another name, or a different animal that needs its own rules? That question is now live and unresolved — and you will see it fought in your own cases.
Ruling two: adding an in-house lawyer as a custodian (ECF No. 188)
The plaintiffs moved to add LinkedIn in-house attorney Jane Slater as a custodian, arguing she was one of three lawyers who negotiated the private API agreements and the only person who could fill the gap left by a business executive, Rohan Verma, who left in 2021 and whose files were not preserved.
The standard is the one to write down: an added custodian must have “uniquely relevant information that is not available from the sources already designated,” shown through a “particularized showing” of the “specific gaps” only that custodian can fill — from the Court’s prior order and In re Facebook, Inc. Consumer Privacy User Profile Litigation.
The plaintiffs failed it. LinkedIn was already producing from five current and former business-development leaders on the API relationships, two of whom were Verma’s direct supervisors, so the gap was largely covered. And Slater is a lawyer whose role was to give legal advice on the API terms, so her internal communications are likely overwhelmingly privileged. The Court found the gap “speculative,” resting on two unproven assumptions — that Slater and Verma deliberated internally with no other custodian present, and that any such deliberations were non-privileged “business-negotiation” communications. The burden of collecting, reviewing, and logging a lawyer’s files to find a handful of arguably non-privileged messages outweighed the benefit. Denied — though LinkedIn had already agreed to produce Slater’s external communications with the API partners, where the non-privileged substance lives.
The practice point: to add a custodian, especially a lawyer, make a particularized showing of specific, non-privileged, non-duplicative material only that person holds. “Involved and might have something unique” loses.
Ruling three: the text-message motion (ECF No. 190)
The plaintiffs moved to compel text messages — iMessage, SMS, WhatsApp, Signal — from all nineteen custodians on personal and work devices, arguing executives “routinely” text about business.
But the parties’ Interim ESI Order, at Paragraph 4(b), already carved text messages out absent a showing of good cause by the requesting party. This motion was not on a blank slate; the plaintiffs had contracted into a good-cause burden. They could not carry it. Their evidence was mostly emails referring to texts, which in context looked logistical or cumulative: “text if you need anything urgent” while traveling; an “overview of our progress” that was already in the produced email. Their strongest example — a produced screenshot of texts about topics for an upcoming quarterly business review (“QBR”) — failed because LinkedIn had already produced the final QBR prep documents, so related texts were likely duplicative. The plaintiffs could not point to a single substantive business discussion occurring by text that was not captured elsewhere.
Then proportionality again: nineteen custodians, multiple platforms, personal and work devices, and the “disruption to the private lives” of each custodian. And a good-faith beat worth stealing: LinkedIn had identified three repositories most likely to hold responsive, non-duplicative material — M365 mailboxes, Google Drive, and Slack — and voluntarily included Slack, itself a category of text messaging excluded under the order. That helped the court credit LinkedIn’s search as reasonable, and it closed by invoking the Sedona Principles, Third Edition — that the responding party is ordinarily best positioned to decide what is relevant and proportional in its own data.
The lesson is the same one that decided ruling one: the ESI order did the work. Paragraph 4(b) carved out texts and shifted the burden; Paragraph 5(a) required only disclosure of TAR use. Two provisions, negotiated at the front of the case, decided two of three motions before the court ever reached the merits.
Where this sits in the AI-in-discovery line
Put it on the map. In Warner v. Gilbarco and Morgan v. V2X, the AI user was a litigant and the question was whether that litigant’s own prompts were protected work product — Warner gave us “tool, not a person,” and both protected the prompts. In United States v. Heppner, the exception: use the tool outside the lawyer relationship and lose the protection. In Conservation Law Foundation v. Shell, expert prompts were discoverable. Last week, James v. Cerebras — counsel and their vendors using AI as the review engine, with the parties agreeing to a sweeping transparency protocol.
Schulte is the contested counterpart to Cerebras. Same posture — AI as the review engine — but where the Cerebras parties handed over the elusion rate, the error rate, the reviewer counts, and the prompts by stipulation, the Schulte plaintiffs asked a court to compel that same list and hit the discovery-on-discovery wall. What one set of parties gave away voluntarily, another could not pry loose by motion. Because the parties negotiated their ESI protocol in 2023 before ChatGPT hit the market and we were thinking about AI review. The protocol didn’t cover it, and the Judge lumped AI review with TAR, which was in the protocol, and said that information is not required. The parties didn’t try to amend the ESI protocol.
The sharp question to carry into your next matter: if the other side is using generative AI to decide what you get, how will you know whether it worked — and where does that right come from? After Schulte, the answer is not “I’ll move to compel the metrics.” It is “I negotiated the disclosure I need into the ESI protocol, at the front of the case, when I still had the leverage.”
What to do this week
If you want visibility into an opponent’s AI or TAR review — elusion rates, error rates, validation, reviewer counts, prompts — negotiate it into the ESI protocol at the outset. Schulte says the discovery-on-discovery doctrine will not hand it to you later; Cerebras shows exactly what those provisions can look like. Use one to build the other.
If you are challenging a pre-cull or search terms, do not argue that pre-culling is improper in the abstract. Make a particularized showing that the specific search strings are deficient or too narrow, and raise it when the strings are disclosed — not months later.
If you are the responding party, know what your ESI order actually requires. Under a standard TAR-disclosure clause, disclosing that you used TAR or generative AI to filter non-responsive documents is generally enough; you are not automatically obligated to produce your elusion estimate, error rate, or reviewer headcount. The protection comes from the doctrine plus a clean disclosure — not from stonewalling.
If you are moving to add a custodian, especially an in-house lawyer, make a particularized showing of specific, non-privileged, non-duplicative material that only that person holds. A speculative gap plus a likely-privileged file set loses.
On text messages, if your ESI order carves them out absent good cause, “executives text about business” is not good cause. You need a specific, substantive business discussion that happened by text and is not captured anywhere else. And for producing parties: voluntarily including a source like Slack helped LinkedIn establish good faith. Reasonableness is a record you build.
And the one underneath all of it: draft the Interim ESI Order like it will decide your case, because it will. Two provisions here resolved two of three motions on their own terms. The leverage is at the drafting table, not the motion-to-compel stage.
What to watch next
Watch whether the Northern District of California — and other courts — treat generative AI review as “a form of Technology Assisted Review,” as Schulte did, or as its own category needing its own rulebook, as Cerebras did. That unresolved split, playing out one week apart in the same district, is going to define what disclosure obligations attach when a party uses an LLM to decide what gets produced. And watch the parties’ twenty-one-day meet and confer on LinkedIn’s search strings — the court left the door open for a specific, timely challenge, which is exactly the kind of record that could change the analysis.
Listen to the full episode
This week’s Case of the Week segment of the Meet and Confer podcast walks through all three rulings in Schulte v. LinkedIn — the Relativity aiR and pre-culling fight, the discovery-on-discovery limit on auditing an opponent’s AI, the custodian standard, and the text-message carve-out — and how they connect to last week’s stipulated AI protocol in James v. Cerebras. Listen on Meet and Confer →
See Minerva26 in action
Minerva26 is the discovery intelligence platform that connects case law, rules, and real-world workflows. We curate and tag close to 50,000 discovery decisions by issue— including the ESI protocols and orders themselves — so when you are negotiating a protocol, preparing for a meet-and-confer, or deciding how to handle AI in review, the rulings and drafting precedents you need are already organized for you. Book a 30-minute demo
Related on Minerva26: James v. Cerebras Systems · Warner v. Gilbarco · U.S. v. Heppner · Morgan v. V2X · Conservation Law Foundation v. Shell · Taylor v. Google · In re Facebook Consumer Privacy
This decision is available on the Minerva26 platform with full issue tagging. If you’re a litigator navigating discovery strategy and want to stay ahead of decisions like this one, visit Minerva26.com to learn more or schedule a demo. Every decision covered on Case of the Week is searchable by issue, jurisdiction, and judge.
If it’s about the discovery of ESI, it’s covered in Minerva26, your discovery strategy platform.
Kelly Twigger is CEO and founder of Minerva26 and Principal at ESI Attorneys. She has been a discovery strategist and practicing attorney for nearly 30 years. Case of the Week is a segment of the Meet and Confer podcast, breaking down one recent ESI discovery decision each week into practical strategy you can use.
- 📩 Get the newsletter
- 🎧 Listen to Meet and Confer
- 🔁 Share this post on LinkedIn
