This article explains how technology-assisted review functions under existing procedural rules and case law. It is information for legal-operations and compliance planning, not legal advice, and does not substitute for jurisdiction-specific counsel on a live discovery dispute.
What counts as technology-assisted review, and how is it different from keyword search?
Technology-assisted review (TAR), also called predictive coding, uses a machine-learning model trained on human-coded example documents to rank or score an entire review population by likely relevance, rather than filtering documents through a fixed list of search terms chosen in advance.
The workflow typically starts with a sample set of documents that one or more attorneys code as responsive or non-responsive. The model learns from that coding, scores the remainder of the collection, and goes through iterative training rounds until scoring stabilizes. Reviewers then validate the result statistically, usually by sampling the "null set" — the documents the model predicted as non-responsive — to estimate how much relevant material was missed. Federal courts have been evaluating this workflow since 2012, when a Southern District of New York opinion in Da Silva Moore v. Publicis Groupe became the first to hold that computer-assisted review is an acceptable method for searching electronically stored information in appropriate cases (SRC-03). A related and recurring dispute is sequencing: whether keyword culling should happen before TAR is applied to the surviving set, an approach one Eastern District of Michigan court called the "preferred method" in 2017 litigation, or whether TAR should run against the full collection first (SRC-03).
Can a court force a party to use TAR instead of keyword search or manual review?
No. Courts have consistently held that the responding party — not the requesting party or the judge — decides how to search for and produce its own electronically stored information, so long as the result is reasonable and defensible.
A Southern District of New York decision in 2016 invoked Sedona Conference Principle 6, which states that responding parties are best situated to evaluate the procedures appropriate for preserving and producing their own electronically stored information, and held that courts cannot compel a party to adopt TAR over its objection (SRC-03). A Northern District of California ruling the same year reached the same conclusion (SRC-03). That deference has a limit, however: a 2020 District of New Jersey opinion in emissions-related litigation declined to order TAR but warned that a party who refuses to use it may find its later burden-and-proportionality arguments less persuasive if the volume of documents becomes an issue (SRC-03). In practice, the choice to decline TAR is a strategic one that can resurface later in a cost dispute.
How does Rule 26 proportionality shape what a TAR review actually costs?
Federal Rule of Civil Procedure 26(b)(1) limits discovery to material that is proportional to the needs of the case, weighing the importance of the issues at stake, the amount in controversy, the parties' relative access to information and resources, and whether the burden or expense of the discovery outweighs its likely benefit (SRC-01).
That proportionality language is what makes TAR attractive in high-volume matters and what makes review validation results relevant to cost disputes. When a review population turns out to be mostly non-responsive, proportionality can cut against the requesting party: a 2020 District of Kansas order shifted TAR-related review costs to the party pursuing discovery after the dataset returned a responsiveness rate of roughly 3.3%, on the reasoning that continuing the review had become disproportionate to its likely yield (SRC-03). Rule 26(f) also requires the parties to address ESI form of production, and any privilege-related timing issues, at the mandatory discovery-planning conference — the point at which a TAR protocol, if the parties intend to negotiate one jointly, should be raised (SRC-01).
What protects privilege when a review tool has to process documents that turn out to be privileged?
Federal Rule of Evidence 502 is the backstop. Subsection (b) provides that an inadvertent disclosure of privileged material in a federal proceeding does not waive privilege if the disclosure was inadvertent, the holder took reasonable steps to prevent it, and the holder promptly took reasonable steps to correct it (SRC-02).
Subsection (d) goes further: a federal court may enter an order providing that disclosure connected with the pending litigation does not waive privilege, and that order is enforceable against other federal and state proceedings — the basis for the "clawback" orders that make large-scale automated review workable without treating every model misclassification as a waiver event (SRC-02). Counsel negotiating a TAR protocol typically pair it with a proposed Rule 502(d) order precisely because a predictive model, like any human reviewer, will misclassify some documents, and the clawback order is what keeps that error from becoming a privilege waiver.
| Dimension | Keyword search | Technology-assisted review |
|---|---|---|
| How relevance is determined | Fixed terms negotiated or unilaterally chosen in advance | Model trained on attorney-coded example documents, then applied to the full set |
| Who typically controls the method | Responding party, subject to negotiation over term lists | Responding party, per case law on Sedona Principle 6 (SRC-03) |
| How courts have evaluated it | Long-accepted baseline; not immune from proportionality disputes | Accepted since 2012 and "not held to a higher standard" than keyword search or manual review, per Southern District of New York case law (SRC-03) |
| Typical validation step | Hit-count and sampling review of search-term results | Statistical sampling of the predicted-non-responsive "null set" to estimate recall (SRC-03) |
What belongs in a TAR protocol before review starts?
A workable protocol states the training methodology, the validation sampling approach, and how the parties will resolve disagreements about the results — decided before review begins, not after production is challenged.
Case law shows what happens when that groundwork is skipped. A Northern District of Illinois multidistrict litigation protocol from 2018 built in transparency requirements, disclosure of culling technologies, and validation and recall-estimation procedures up front (SRC-03). By contrast, a District of Columbia antitrust matter saw discovery deadlines extended after a TAR production came back roughly 83% non-responsive, illustrating what insufficient testing and validation can cost in delay (SRC-03). Courts have also penalized parties who changed review methodology mid-stream without disclosure: a 2014 District of Nevada order barred an unannounced shift to TAR after the parties had already agreed on a different protocol, while a Middle District of Tennessee court the same year allowed a mid-discovery switch to TAR because the process stayed transparent and cooperative (SRC-03). The distinguishing factor across these rulings is not which technology was used, but whether the process was disclosed and defensible.
FAQ
For a related compliance perspective, read FinCEN's SAR Rule Explained: The 30-Day Deadline, the $5,000 Threshold, and What Changed in 2025.
For more context, read Trace a viral photo's origin with reverse image search.
For more context, read Trace any viral photo to its real source.
For more context, read What age does each platform say you have to be?.