What precedent made technology-assisted review defensible, and what did those opinions actually hold?
The foundational opinion is Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012), in which Magistrate Judge Andrew J. Peck recognized computer-assisted review as an acceptable way to search for relevant electronically stored information in appropriate cases. The case was an employment discrimination action against Publicis Groupe and MSL Group — gender and pregnancy discrimination claims under Title VII along with Equal Pay Act, Fair Labor Standards Act and New York State and City claims — in which more than three million emails from agreed-upon custodians were at issue. The holding is narrower than its reputation: the court approved the method, and conditioned that approval on process, writing at 193 that “counsel must design an appropriate process, including use of available technology, with appropriate quality control testing, to review and produce relevant ESI while adhering to Rule 1 and Rule 26(b)(2)(C) proportionality.” The opinion then limited itself in terms worth quoting in full, because they are routinely dropped: “That does not mean computer-assisted review must be used in all cases, or that the exact ESI protocol approved here will be appropriate in all future cases that utilize computer-assisted review. Nor does this Opinion endorse any vendor.”
Three years later, in Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125, 127 (S.D.N.Y. 2015), Judge Peck wrote that “the case law has developed to the point that it is now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it.” Read the sentence closely. It establishes that a producing party may use the technology. It says nothing about a producing party being obliged to, and nothing about what the requesting party gets to see.
A footnote in the same opinion also addressed how much a requesting party is entitled to police the process, noting that requesting parties can ensure training and review were done appropriately by other means — statistical estimation of recall at the conclusion of the review, whether there are gaps in the production, and quality control review of samples from the documents categorized as non-responsive. That is the architecture of the whole area: the producing party chooses the method, and the requesting party tests the output.
Can a court force a party to use TAR, or generative AI, instead of keyword searching?
A court refused to, in an opinion that still governs how these fights are framed. Hyles v. New York City, No. 10 Civ. 3119 (AT)(AJP) (S.D.N.Y. 1 August 2016) (Peck, M.J.), put the question directly: whether the responding party can be forced to use TAR when it prefers keyword searching. The opinion answered, in its own words, “a decisive ‘NO.’”
What makes the opinion useful is that the judge plainly wanted the opposite result. Judge Peck wrote that “for most cases today, TAR is the best and most efficient search tool,” that the court “would have liked the City to use TAR in this case,” and then that the court “cannot, and will not, force the City to do so.” The reason was Sedona Principle 6: responding parties are best situated to evaluate the procedures, methodologies and technologies appropriate for preserving and producing their own electronically stored information. He also quoted the Tax Court in Dynamo Holdings Ltd. Partnership v. Commissioner, 143 T.C. 9 (2014), which observed that a court would not ordinarily dictate whether a paper review be done by a paralegal, a junior attorney or a senior attorney.
The sentence most often quoted from Hyles is the one that dates itself on purpose: “There may come a time when TAR is so widely used that it might be unreasonable for a party to decline to use TAR. We are not there yet.” That was August 2016. No opinion found for this article has declared that time to have arrived.
Has a court approved generative AI for responsiveness calls in discovery?
Yes. In Schulte v. LinkedIn Corp., No. 22-cv-00237-HSG (LB) (N.D. Cal.), Magistrate Judge Laurel Beeler signed a discovery order on 30 June 2026, entered on the docket 1 July 2026, resolving three letter briefs in a putative class action alleging monopolization under section 2 of the Sherman Act. One of them concerned LinkedIn’s use of Relativity aiR, a generative AI product, for document review.
The sequence in the order is worth having in front of you. On 15 May 2026 LinkedIn gave the plaintiffs twenty-five search strings to be applied against its custodial documents and told them it would use Relativity aiR “to assist in filtering out non-responsive documents.” On 23 May 2026, in response to the plaintiffs’ request for more information, LinkedIn disclosed “(1) there was no seed or training set used; (2) Relativity aiR is being used to make final responsiveness calls; and (3) quality control review is being conducted by human review of samples taken from each responsiveness type.”
The plaintiffs asked the court to prohibit the search-string pre-culling, to compel LinkedIn to run Relativity aiR across all custodial files, and to compel disclosure of further metrics. All three requests were denied as not “reasonable or proportional to the needs of the case.” The court did order the parties to meet and confer within twenty-one days about the search strings, leaving the plaintiffs a route back if adjustments were warranted. The Institute’s Document review and Litigation workflow pages cover how firms structure the review itself.
How much do you have to disclose about the review protocol and its validation?
In Schulte v. LinkedIn Corp. the answer was: what the governing ESI order required, plus what was volunteered on request, and no more. Paragraph 5(a) of the parties’ interim ESI order required a producing party to “disclose to the receiving party if they intend to use Technology Assisted Review (‘TAR’) to filter out non-responsive documents.” The court held that LinkedIn had disclosed on 15 May 2026 that it would use Relativity aiR — which the order described as “a form of Technology Assisted Review” — to filter out non-responsive documents, and that its disclosures “more than satisfy the demands of the Interim ESI Order.”
The plaintiffs had moved to compel elusion estimates, the document error rate, and the number of human reviewers validating the tool’s predictions. The court analyzed that as discovery on discovery, quoting Taylor v. Google LLC, No. 20-cv-07956-VKD, 2024 WL 4947270, at *2 (N.D. Cal. 3 December 2024): such discovery is “disfavored, as such discovery is typically not relevant to the merits of a claim or defense, and is rarely proportional to the needs of a case,” and is warranted only where the requesting party demonstrates a specific deficiency in the production rather than “mere speculation.” The only deficiency the plaintiffs identified was that the target population came to 204,444 documents, which the court held did not, standing alone, justify an audit.
Two things follow for a producing party, and they run in opposite directions. The ESI protocol is doing more work than the case law is — LinkedIn’s obligation came from a paragraph the parties negotiated, not from a rule about AI. And the metrics a firm never has to produce are still the metrics it would need if a specific deficiency were ever shown, which is an argument for generating them and not for skipping them.
What changes when the review tool is generative rather than a trained classifier?
The training step disappears, and with it the thing requesting parties spent a decade arguing about. A classifier-based TAR system learns from documents that human reviewers coded, which is why seed sets, training rounds and continuous active learning were the vocabulary of the old disputes. LinkedIn’s disclosure in Schulte v. LinkedIn Corp. — “there was no seed or training set used” — is the concise statement of the difference. What the tool is working from is an instruction, not a coded sample, and there is no seed set for anyone to demand.
The failure modes move too, and two are specific to this architecture. First, generative systems are non-deterministic: the same input can produce different output on different runs. Thinking Machines Lab, in a 10 September 2025 technical analysis titled “Defeating Nondeterminism in LLM Inference,” traced this to how inference servers batch concurrent requests together — a model can return different results for an identical prompt because the batch it happened to run in changed, not because anything about the prompt did, and this holds even when a setting meant to force deterministic output is turned on. Second, these systems run on a fixed context window. Anthropic’s own developer documentation for the Claude API states that a request whose input already exceeds the model’s context window is rejected outright rather than partially processed, which means a document too long to fit in one window has to be broken apart and handled across separate calls. Neither failure has an analogue in a classifier that assigns the same score to the same document every time.
What actually decides these disputes when they are litigated?
Proportionality arithmetic and who failed to make a showing, rather than any view about artificial intelligence. In Schulte v. LinkedIn Corp. the court rejected the attack on pre-culling because the plaintiffs “do not argue or otherwise show that LinkedIn’s twenty-five search strings are deficient,” adding that if they had shown the strings were too narrow “their concern about pre-culling the document population might be warranted.” The motion failed on what was not argued.
The burden side was concrete. Two LinkedIn custodians alone held roughly 800 gigabytes; across nineteen designated custodians, running everything through the platform would have meant feeding multiple terabytes into it, with the attendant processing, hosting and human review costs. The court cited Livingston v. City of Chicago, No. 16-cv-10156, 2020 WL 5253848, at *3 (N.D. Ill. 3 September 2020) and In re Biomet M2a Magnum Hip Implant Products Liability Litigation, No. 3:12-MD-2391, 2013 WL 1729682, at *2 (N.D. Ind. 18 April 2013) for the proposition that using search terms to pre-cull before handing documents to a technology-review platform satisfies the reasonableness and proportionality standards of Rules 26(b) and 34(b)(2).
Cooperation shows up as timing. The court noted the plaintiffs did not contend they had raised any concerns about the search strings when those were disclosed on 15 May 2026, and its remedy was to send the parties to meet and confer about the strings within twenty-one days. Adjacent negotiations are running on a parallel track: in Jeffries v. Harcros Chemicals, Inc., 2026 WL 820218 (D. Kan. 25 March 2026), Magistrate Judge Angel D. Mitchell granted an amended protective order barring the parties from uploading materials produced in discovery — including documents not marked confidential — into public or open-loop generative AI tools, while leaving closed or secure tools available. None of this is legal advice and nothing here predicts how a particular court will rule; what the record shows is that the disputes are being decided on the ordinary discovery standards, and are being won and lost on showings. The Institute’s Approved tools page and which AI tier actually protects client data cover the terms-of-service side of running an adversary’s documents through a tool.