Reviewing modern data volumes document-by-document is neither affordable nor, in many cases, possible. Technology-assisted review (TAR) has become a standard tool for exactly this reason, and courts have accepted it for over a decade. Judicial approval dates to Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012), the first decision to expressly approve the use of predictive coding, and by Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125 (S.D.N.Y. 2015), the same court could observe that it was "now black letter law" that TAR is an acceptable way to search for relevant ESI where the producing party chooses to use it. But the recurring lesson from these cases is that defensibility comes from how you use the technology — your process, validation, and transparency — not from the technology itself.

The vocabulary, briefly

TAR generally

Technology-assisted review is an umbrella term for using machine-learning to prioritize or classify documents for relevance based on human reviewers' decisions on a subset. The human teaches the system; the system extends those judgments across the larger population.

TAR 1.0 vs. continuous active learning

  • TAR 1.0 typically trains a model on a control set and a series of training rounds, then applies the trained model to the rest of the population. Training is largely front-loaded.
  • Continuous active learning (CAL), sometimes called TAR 2.0, learns continuously: the system keeps surfacing the documents it believes are most likely relevant, reviewers code them, and the model updates throughout the review. Learning never really stops.

Generative AI

Newer approaches apply large language models to classification, summarization, and issue-coding. These tools are promising and developing quickly, but they raise the same fundamental question every prior technology did: can you explain and validate what the system did? Novelty does not change the defensibility standard — it raises the bar on documentation.

The principle that endures

Courts have repeatedly signaled that parties are best served by being transparent and cooperative about review methodology. The defensibility of TAR has never turned on choosing the "right" algorithm — it turns on a sound, documented, validated process.

What makes TAR defensible

Whatever method you choose, the same elements separate a defensible review from a fragile one:

  • A documented protocol. Decide and write down how the tool will be used, who makes relevance calls, and how disagreements are resolved — before the review starts.
  • Quality control. Consistent coding decisions are what the model learns from; inconsistent reviewers produce an unreliable result no matter how good the technology.
  • Validation. Use sampling and recall/precision measures to demonstrate that the review found what it should have. The point is to be able to show the result was reasonable, not to claim perfection.
  • Proportionality. Rule 26(b)(1) supports using efficient methods on large populations; TAR is often the proportionate choice precisely because linear review would be disproportionate.
  • Transparency. Methodology that you are willing to disclose and discuss is methodology a court is far more likely to accept.

Address it at the Rule 26(f) conference

The best time to establish a defensible review is before it starts — at the meet-and-confer. Raising your intended methodology, validation approach, and the degree of transparency you will offer puts the issue on the record and reduces the risk of a later challenge. Parties who spring a methodology dispute mid-review tend to lose time and credibility.

Govern AI use deliberately

As generative AI moves into review, organizations should treat it the way they treat any other review technology: with a documented protocol, validation, human oversight, and attention to confidentiality and data-handling. The legal standard has not changed — reasonable, good-faith, proportional effort that you can explain. The tools that satisfy it are simply getting more powerful.

If your playbook does not yet document how technology-assisted review is selected, validated, and disclosed, that is a gap worth closing before your next large matter — because the time to design a defensible review process is never in the middle of one.