court-accepted since 2012

Technology-Assisted Review (TAR) is a machine learning-based approach to document review in e-discovery that uses algorithms to classify documents as relevant or non-relevant based on attorney training decisions. It's the legal industry's term for what the broader tech world calls supervised machine learning applied to text classification.

TAR has been court-accepted since 2012 and is now the default methodology for large-scale document reviews. The two dominant versions — TAR 1.0 (seed-set training) and TAR 2.0 (continuous active learning) — represent fundamentally different approaches, and understanding the distinction matters for both cost control and court defensibility.

TAR 1.0: The Seed-Set Approach

TAR 1.0 follows a batch training model. A subject matter expert (typically a senior associate or partner) reviews a statistically random sample of documents — the "seed set" — and codes each as relevant or non-relevant. The algorithm learns from this seed set and scores the entire document collection. Documents scoring above a cutoff threshold go to human review; those below are designated non-responsive. The process includes validation rounds where attorneys review a sample of the algorithm's decisions to measure recall and precision. If accuracy is insufficient, additional training rounds occur. TAR 1.0 is straightforward and well-understood by courts, but it has a significant limitation: the model is static. It doesn't improve after the initial training, even if reviewers encounter new document types or issues during review.

TAR 2.0: Continuous Active Learning (CAL)

TAR 2.0 — also called Continuous Active Learning (CAL) — is a fundamentally better approach. Instead of training on a fixed seed set, TAR 2.0 continuously learns from every coding decision a reviewer makes throughout the entire review. The algorithm prioritizes documents for review based on where it's most uncertain, meaning human reviewers spend their time on the most informative documents rather than the most random ones. Key advantages over TAR 1.0: higher recall (finds more relevant documents), faster convergence (reaches reliable results with fewer human reviews), adaptive (adjusts as new issues emerge during review), and no seed-set bias (doesn't depend on the quality of an initial random sample). Studies from the TREC Legal Track show TAR 2.0 achieves 80-90% recall compared to TAR 1.0's 70-80% — a meaningful difference when millions of dollars in litigation outcomes depend on finding the right documents.

When Courts Require TAR

Courts don't universally mandate TAR, but the trajectory is clear. Compelled TAR: In *In re Broiler Chicken Antitrust Litigation* (2018), the court ordered a party to use TAR despite objections, ruling that manual review of the document volume would be disproportionate. Expected TAR: Multiple federal judges have stated that for collections exceeding 500,000 documents, parties should demonstrate why TAR wasn't used if they opted for manual review. Disclosed TAR: *Livingston v. City of Chicago* (2022) required parties to disclose their TAR methodology, seed set composition, and validation metrics. The practical standard in 2026: if your collection exceeds 100,000 documents and you're not using TAR, you need a defensible reason. "We've always done it this way" isn't one.

TAR Protocols and Defensibility

A defensible TAR protocol includes five elements. Transparency: document the workflow, including which TAR version, what platform, and who served as subject matter expert. Validation: measure recall using statistical sampling — courts generally expect 70%+ recall as a minimum, with 80%+ preferred. Quality control: conduct ongoing checks, including elusion testing (sampling documents below the cutoff to estimate how many relevant documents were missed). Documentation: maintain records of all training decisions, algorithm settings, and validation results. Proportionality: demonstrate that the TAR approach is proportional to the case value and document volume. Platforms like Relativity, Everlaw, and Reveal include built-in TAR workflows with audit trails that produce court-ready documentation.

TAR Costs and ROI for Managing Partners

The economics are straightforward. Manual review: $1.50-3.00 per document using contract reviewers, plus quality control layers. A 2 million document collection costs $3-6 million in review spend. TAR-assisted review: $0.10-0.50 per document for the TAR platform, plus senior attorney time for training and validation. The same 2 million documents cost $200K-1M — a 70-85% cost reduction. But the ROI isn't just cost savings. TAR produces more accurate results (75-85% recall vs. 55-65% for manual), faster timelines (weeks instead of months), and better defensibility (statistical validation metrics that courts accept). For a managing partner evaluating whether to invest in TAR capabilities: every large litigation matter your firm handles without TAR is leaving money on the table and delivering inferior results.

The Bottom Line: TAR is machine learning applied to e-discovery document review. TAR 2.0 (Continuous Active Learning) is the current standard, achieving 80-90% recall while cutting review costs by 70-85%. Courts increasingly expect TAR for large collections, and some have ordered its use over party objections. If your firm handles litigation with 100K+ document reviews, TAR isn't optional — it's the defensible standard.