TL;DR: Technology-assisted review (TAR), also called predictive coding, is the use of machine learning to prioritise and classify documents by relevance in e-discovery, so reviewers focus on the documents most likely to matter. A model trained on reviewer decisions ranks a large set; people verify samples. Courts have accepted TAR since Da Silva Moore v. Publicis Groupe (2012). It is classification, not generative AI.
Modern litigation can turn on a few crucial documents buried in millions of irrelevant ones. Reviewing that volume by hand is slow, expensive, and - counterintuitively - often less accurate than a well-run machine process. Technology-assisted review emerged to solve exactly this problem, and over the past decade it has moved from a contested novelty to an accepted, court-approved discipline. This article explains what TAR and predictive coding are, how they work, how their accuracy is measured, and how they differ from the generative AI now dominating headlines.
What is TAR (technology-assisted review) and predictive coding?
Technology-assisted review (TAR), also called predictive coding, is the use of machine learning to prioritise and classify documents by relevance in e-discovery, so reviewers focus on the documents most likely to matter. Instead of reading every file in order, a team trains a model on a sample of documents that lawyers have marked relevant or not; the model then ranks the entire collection by likely relevance, pushing the important material to the top and the noise to the bottom.
The terms are largely synonymous. "Predictive coding" was the original phrase for the technique; "technology-assisted review" is the broader, now-standard term used by courts and practitioners. Both describe a supervised classification task: the system is learning a boundary between responsive and non-responsive documents from human examples, not generating anything. That places TAR firmly in the classic machine-learning camp rather than the generative one, a distinction we draw out in machine learning vs generative AI in legal work.
How does TAR work?
TAR works by turning a handful of human review decisions into a model that can score an entire document population. Lawyers review and code a set of documents; the system learns the features that distinguish relevant from irrelevant; and it applies that learning to rank everything else. The workflow then loops or finishes depending on which generation of TAR you use. Two ideas - the training set and the learning style - explain most of how it behaves.
Seed sets and training
A seed set is the initial batch of documents that subject-matter experts code to teach the model what relevance looks like in this matter. The quality of that seed set largely determines the quality of the result: a representative, carefully coded set produces a model that generalises well, while a biased or careless one teaches the system the wrong boundary. From there, supervised learning takes over - the model predicts, reviewers correct, and the predictions sharpen with each round of feedback.
TAR 1.0 vs TAR 2.0: continuous active learning
TAR 1.0 uses a one-time training phase: experts code a seed set, the model is trained and validated, and it then classifies the rest of the collection in a single pass. TAR 2.0, known as continuous active learning, instead learns continuously - the model keeps proposing the next most likely relevant documents, reviewers code them, and it updates after every batch until few relevant documents remain. Continuous active learning has largely become the modern default because it adapts as understanding of the matter evolves and does not depend on getting one seed set perfect at the outset.
How is TAR accuracy measured?
TAR accuracy is measured with two complementary metrics: recall and precision. Recall is the share of the truly relevant documents that the process actually found - the measure that matters most for defensibility, because it answers whether you missed anything important. Precision is the share of the documents the model flagged as relevant that really are relevant - the measure of how much reviewer time is wasted on false positives. The two trade off against each other, and a well-run TAR project targets a recall level the parties agree is reasonable, validated by sampling.
This statistical grounding is what makes TAR defensible. Because you can draw a random sample of the documents the model set aside and estimate how many relevant ones it missed, you can put a number on the process's completeness - something manual review, surprisingly, rarely does. Studies have repeatedly found that well-run TAR matches or beats exhaustive human review on recall while reviewing a fraction of the documents. That measurability is the deep difference between classification AI and generative AI, which cannot be audited the same way.
How do courts treat TAR and predictive coding?
Courts have treated TAR as an accepted, even encouraged, e-discovery method for over a decade. The landmark is Da Silva Moore v. Publicis Groupe, decided in the Southern District of New York in 2012, in which Magistrate Judge Andrew Peck issued the first federal opinion expressly approving the use of predictive coding. The opinion did not mandate TAR, but it confirmed that using it was reasonable and defensible where proportionate - a turning point that opened the door to widespread adoption. You can look up the docket and related filings on CourtListener.
Since then, a body of guidance has matured. The Sedona Conference, an influential think tank on e-discovery, publishes widely cited commentary and principles on the defensible use of TAR; the Sedona Conference materials are a standard reference for cooperation, proportionality, and validation protocols. The practical consensus today is not whether TAR is permissible but how to run it defensibly: transparent protocols, agreed recall targets, and sampling-based validation. Courts care about the process being reasonable, not about the brand of software.
How is TAR different from generative AI?
TAR is a classification technology, not a generative one. It does not write summaries, draft arguments, or answer questions in prose - it ranks and labels documents by relevance using supervised machine learning, and its output is a prioritised, scored set you can validate statistically. Generative AI, by contrast, produces new text and carries the risk of fluent fabrication that TAR simply does not have, because TAR never generates a claim that could be false; it points you at documents that already exist.
This is why TAR's defensibility rests on measurable recall and precision while generative AI's rests on citation and human verification. The two can complement each other: a TAR process narrows millions of documents to the relevant few thousand, and a generative tool can then summarise or extract from that smaller, validated set - provided every generated claim is cited and checked. Confusing the two leads to applying the wrong safeguards, which is the core argument of our explainer on machine learning vs generative AI.
TAR vs manual review: what changes?
The clearest case for TAR is a direct comparison with linear, document-by-document human review. TAR does not remove lawyers - it changes what they spend their time on, moving them from reading everything to training, validating, and judging the documents that matter. The table contrasts the two approaches.
| Dimension | Manual linear review | TAR / predictive coding |
|---|---|---|
| How documents are ordered | Arbitrary or chronological; all read | Ranked by predicted relevance; least likely set aside |
| Reviewer effort | Read every document in the set | Code a training sample; review high-probability documents |
| Speed and cost | Slow and expensive at scale | Far faster and cheaper on large collections |
| Measuring completeness | Rarely quantified | Recall and precision estimated by sampling |
| Consistency | Varies by reviewer and fatigue | Uniform model applied across the set |
| Human role | Read and code everything | Train, validate, and decide on the important documents |
The pattern is familiar from every good use of legal AI: the machine handles scale, the lawyer keeps judgment. TAR reviews more consistently and lets you prove completeness, but a person still defines relevance, settles close calls, and stands behind the production.
Where does Judicio fit?
TAR proper is a specialised e-discovery discipline, and Judicio is not a dedicated e-discovery review platform. What Judicio shares with TAR is the underlying philosophy: apply AI to large volumes of documents, keep a human in the loop, and make every result traceable to its source. Where TAR ranks documents by relevance, Judicio's Document Review and Review Matrix read and extract from multiple files in a single run - and up to 25 questions in a matrix - with every finding and answer cited to the exact page and quoted passage.
That makes Judicio a strong fit for the analysis that surrounds and follows a relevance review: due-diligence document sets, contract portfolios, and investigation files where you need structured answers, not just a relevance rank. One upload into the File Library feeds every tool, citations are deterministic rather than AI-generated, and you can export an evidence pack. Judicio does not train on your data and hosts on Google Cloud Platform. For a related workflow at deal scale, see our guide to AI due diligence in mergers and acquisitions and the principles behind AI contract review.
How do you get started with TAR?
Getting started with TAR is mostly a matter of process, not software. Agree the protocol with the other side early, choose a continuous-active-learning workflow unless there is a reason not to, invest in a well-coded training sample, and validate completeness by sampling against an agreed recall target. The defensibility comes from running the process transparently and documenting it - the tool is secondary to the protocol.
If your need is the broader document analysis around a review - extracting terms, answering questions across a file set, building a timeline - Judicio lets you try that on your own matters with a 7-day free trial of 500 credits and no credit card. Professional plans are $200 per month for 5,000 credits, and you can contact us for a walkthrough. Whether you run formal TAR or AI-assisted review, the principle holds: let the machine handle scale, measure your results, and keep a lawyer in charge of relevance and judgment.
Judicio's outputs are review and research aids, not legal advice; a qualified lawyer remains responsible for every production and decision.