This past February, at Consero’s inaugural eDiscovery Leadership Forum, we put a question to a room of legal, technology, and litigation leaders: if review costs drop dramatically, do “unduly burdensome” and “proportional” still mean the same thing? Does faster review expand what is reasonable, or does it just introduce new risks? 

We closed that session with more questions than answers. At the end of April, we came back to the discussion in our webinar, Reality Check: Rethinking Proportionality in the GenAI Age, and asked: are generative AI and technology-assisted review apples and oranges? And does generative AI need a landmark decision – its own Da Silva Moore –  to make it safe to use? 

Now, the court has weighed in. A recent ruling in Schulte v. LinkedIn answers three of our questions, while leaving more to be discussed. 

What happened 

In Schulte v. LinkedIn Corporation, 2026 WL 1905851 Case No. 22-cv-00237-HSG (N.D. Cal. 2026), Magistrate Judge Laurel Beeler resolved three discovery disputes in a putative antitrust class action. One concerned LinkedIn’s use of Relativity aiR for document review. 

LinkedIn disclosed in May that it would apply twenty-five search strings to its custodial documents and then use aiR to make final responsiveness calls, with quality control by human review of samples drawn from each responsiveness category. It said no seed or training set was used. 

The plaintiffs asked the court for three things: bar the search-string pre-culling, compel LinkedIn to run aiR across all custodial files, and compel disclosure of validation metrics including elusion estimates, error rates, and reviewer counts. The court denied all three as neither reasonable nor proportional, and ordered the parties to meet and confer on the search strings within twenty-one days. 

The legal analysis has been well covered elsewhere, and the coverage has largely landed in the same place: old rules, new tools. Fundamentally true. But our focus has been more specific to questions of proportionality, and this ruling speaks directly to three of them. 

Question #1: Does cheaper and faster review expand what is reasonable? 

No. The court’s reasoning shows why. 

The plaintiffs’ argument was not that AI is dangerous. It was, in effect, that AI is capable, so point it at everything. They wanted the pre-culling stopped and the tool run across the full custodial set. 

The court said no. Two custodians alone hold roughly 800 gigabytes. Across nineteen designated custodians, running everything through aiR would mean processing multiple terabytes, with significant cost for processing, hosting, and human review. That burden was disproportionate to the needs of the case. 

The court did the proportionality math with AI factored in, and the tool’s capabilities did not expand the scope of what a party can be compelled to do. 

That is a direct answer to February. Proportionality has not changed as a legal concept, but the assumptions behind it were built on human review economics. When the economics shift, the arguments need to be reexamined. Schulte is that reexamination happening in a live matter. 

Question #2: Are generative AI and TAR apples and oranges? 

Not to this court. 

The operative ESI order required a producing party to disclose if it intended to use technology-assisted review (TAR) to filter out non-responsive documents. When LinkedIn made that disclosure in May, the court recorded that it would use Relativity aiR, which the order describes as “a form of technology assisted review,” to filter out non-responsive documents. The court found those disclosures more than satisfied the ESI order. 

That single clause does a lot of work. The court did not build a new framework for generative AI; instead, it applied an existing TAR provision to a generative AI tool and equated the two in the process.  

Question #3: Does Gen AI need its own Da Silva Moore? 

We do not think so – because it already has one. 

Judge Andrew Peck’s trilogy still governs the ground here. Da Silva Moore v. Publicis Groupe (2012) was the first judicial approval of TAR. Rio Tinto v. Vale (2015) made it, in his words, “black letter law,” holding that a producing party does not need permission to use TAR and that TAR should not face a higher standard of scrutiny than manual review or keyword searching. Hyles v. City of New York (2016) drew the other boundary: courts will not force an unwilling party to use TAR, because the producing party decides how to search its own data. 

That last one matters most here, and it is the part the coverage has missed. Hyles said a court would not compel a producing party to adopt TAR. Schulte says a court will not compel a party to extend generative AI beyond its planned use, or to abandon keyword searching in favor of it. Ten years apart, the same principle, now applied to a new tool. 

So, it may be hyperbolic to say a court has “approved” Relativity aiR. That is not what happened. What happened is more useful: this order strengthens the connection between generative AI tools and Peck’s TAR decisions, which means fourteen years of developed case law comes with it. 

What nobody argued about 

It’s worth noting what was not in dispute. No party challenged whether a generative AI tool could make final responsiveness determinations at all. Sophisticated plaintiffs’ counsel did not raise it, and the court did not pause on it. 

That is remarkable, and it deserves a closer look because it cuts against how this technology usually gets discussed. In theory, generative AI review is always framed with a human in the loop, and no document leaves the door without an attorney seeing it. Under that model, aiR is a culling tool. A million documents go in, three hundred thousand come back as responsive, counsel reviews those, and samples the remainder to confirm nothing was missed. 

LinkedIn did something different. According to the published decision, it let the AI determinations stand, with sampling as quality control rather than as a safety net under full human review. And the plaintiffs did not object. 

There are reasons for that. Requesting parties often want more information and want it sooner, and a language model’s responsiveness call may look faster and more objective to them than a contract reviewer’s judgment. Every case is different. This may be a dispute between technologically sophisticated counsel who understand the technology. Perhaps counsel is following Sedona’s cooperation proclamation. LinkedIn may have determined that the time and cost savings from Relativity aiR outweighs the production risk. That combination is not every case; it’s not common, but we see it.  

It is worth remembering how far that risk assessment has moved with AI. At an ILTA presentation in the early 2010s, a well-known eDiscovery attorney, Browning Marean, opined to the audience that he personally considered producing documents without looking at them ‘malpractice.’ While that view was mainstream, and may still be, in Schulte, the opposing party did not raise it.  

We are not recommending that approach; that is a strategic decision for counsel, involving factors too numerous to enumerate. We are simply noting that the ground moved, and that it moved without an argument. 

What we are seeing in our own matters 

Case law tells you what one court will tolerate. It does not tell you what the workflow costs or what validation looks like when you run it. Here is what we can add, having worked with aiR since it became available. 

Adoption of generative AI is accelerating faster than TAR ever did. Since February, the volume of documents CDS runs through aiR each month has roughly doubled. Clients who tried Gen AI on one matter are using it on the next. Early performance results have been consistent: we see recall in the 90% range, and precision percentages range from the low 80s to high 90s. 

Those are generalizations; results vary by data set, matter type, and how carefully the workflow is built. But the pattern shows something the court’s order only implies: no one argued about whether the tool works because, at this point, the market has largely settled that question. 

What is still unsettled 

Three areas are still moving. 

Data security, which Schulte does not address. This line runs through ESI protocols rather than through this order. A growing set of protocols, including the ones entered in three recent cases: 

• Jeffries v. Harcros Chems. Inc., No. 25-2352 (D. Kan. Mar. 25, 2026)  
• Rudasill v. Swiss Re Am. Holding Corp., No. 1:25-cv-01403 (S.D.N.Y. May 13, 2026) 
• U.S.A. v. Mora, 2026 WL 2058416 (S.D.N.Y. Jul. 16, 2026) 

The Jeffries, Rudasill, and Mora matters permit generative AI on conditions: that confidential data will not be used to train the model, that prompts and analyzed data are deleted, and that the platform meets recognized cybersecurity standards such as SOC 2 Type 2 or ISO certification. The practical effect is a line between closed, enterprise-grade systems and publicly available consumer tools. 

The privilege cases sharpen that line. In United States v. Heppner (S.D.N.Y. 2026), a defendant used a consumer chatbot to work through defense strategy before retaining counsel and then tried to shield those exchanges as privileged work product. The court held the logs discoverable, reasoning that an AI tool is not an attorney and that consumer terms of service defeat any reasonable expectation of confidentiality. Three additional privilege decisions reach different conclusions but focus on the critical issue of data security: Warner v. Gilbarco, Morgan v. V2X, and Tate Group Automotive v. Legacy Auto Capital.    

Transparency, and how much is enough. Schulte puts the burden on the requesting party to identify a specific gap in the production before a court looks inside the process. The plaintiffs here pointed only to the size of the target population – 204,444 documents – and the court held that alone did not justify discovery on discovery. That is a workable standard, but it has been tested exactly once. Schulte adds its analysis to two other cases with different approaches to transparency: EEOC v Tesla and James v. Cerebras Sys. Inc. In other words, the debate over transparency continues. 

Whether counsel is expected to read what goes out the door. This is the open question that matters most, and no holding yet resolves it. If LinkedIn reversed course tomorrow and told the court it needed more time because it had decided to review everything, no one would be surprised and few would object. The reasonableness of an attorney reviewing documents before production is still widely assumed, even as practice moves away from it. 

What to do about it 

A few things follow for anyone planning a review right now. 

The leverage is early. The plaintiffs never argued that LinkedIn’s twenty-five search strings were deficient, and the court noted that a showing that the strings were too narrow might have changed the pre-culling analysis. The place to contest the document population is the ESI protocol and the search strings, not a later audit. 

Disclose deliberately, and document as you go. LinkedIn’s disclosures went beyond what the protocol required, and that record is what the court had in front of it. 

Build quality control into the design. Sampling across responsiveness categories was part of the workflow, not something added when questioned. 

Know your validation story before anyone asks. The metrics were not compelled here. That will not always be true. 

Read your ESI protocol for AI terms, not just TAR terms. The security conditions above are increasingly where real constraints live. 

Where this leaves us 

One order is not a doctrine. Schulte is a discovery ruling from a magistrate judge in one district, in a case with its own facts and its own temperature, and it says nothing about matters that look different. 

But it answers the three questions we have been asking all year. Faster and cheaper review does not automatically widen what is reasonable. Generative AI and TAR are not apples and oranges. And generative AI did not need a new landmark, because the old one still applies. 

What comes next is harder to see. The proportionality math will shift as this technology becomes ordinary, but it will not shift because a judge multiplies a document count by a speed factor. It will happen gradually, as parties recognize they can get more done and start expecting to. That change will arrive in a case, not in a rollout, and probably before anyone is ready to argue it. 

We will keep watching, and we will keep asking. 

If you are negotiating an ESI protocol with AI terms in it or building a review workflow you may have to defend later, let’s talk. For a recap of the webinar, you can view it on-demand here, and be sure to catch us at Relativity Fest in Chicago, September 29 to October 1. 

This article offers a reference point, not legal guidance. CDS recommends consulting with counsel regarding the relevance of this decision to your specific matter and jurisdiction.