What's the difference between recall and precision in TAR validation, and which one matters more for eDiscovery?

The Direct Answer: Recall and Precision Measure Different Failures

In technology-assisted review (TAR) validation, recall and precision answer two different questions about the quality of a document review. Recall measures completeness: of all the responsive documents that actually exist in the corpus, what percentage did the review process find? Precision measures accuracy: of the documents the review flagged as responsive, what percentage truly are responsive? A TAR workflow that achieves 90% recall has missed 10% of the responsive documents; a workflow that achieves 80% precision is producing a production in which one out of every five documents should not have been included.

Also worth reading: How Should Legal Teams Measure eDiscovery Validation Performance in 2026? · How Do You Build an AI eDiscovery Validation Checklist for Court-Defensible Review? · What Are the Definitive AI eDiscovery Validation Protocols for 2027?

Neither metric can be evaluated in isolation, because they trade against each other. You can trivially achieve 100% recall by marking every document responsive, just as you can achieve high precision by marking only the most obviously responsive documents. Validation is the discipline of measuring where your system sits on that trade-off curve and deciding whether that position is defensible for the case at hand. Courts, regulators, and opposing counsel increasingly expect teams to report both numbers, along with confidence intervals, rather than a single vague claim that the review was 'accurate.'

The distinction matters more now than it did five years ago because generative AI review tools have entered the market alongside classic TAR. Recent industry coverage — including Law.com reporting on a major study finding generative AI performed 'nearly expert' in complex document review, and Legal IT Insider's characterization of Gen AI TAR as 'nearly expert, entirely binary' — suggests these systems behave differently from traditional statistical classifiers, which changes how recall and precision should be measured and interpreted during validation.

Why Both Metrics Exist: The Statistical Foundation

Recall and precision come from the confusion matrix, the four-cell table that classifies every document review decision as a true positive (correctly marked responsive), false positive (incorrectly marked responsive), true negative (correctly marked non-responsive), or false negative (incorrectly discarded). Recall equals true positives divided by all actual positives (true positives plus false negatives). Precision equals true positives divided by everything called positive (true positives plus false positives).

In eDiscovery, the two errors carry very different consequences. A false negative is a document that should have been produced but wasn't — this is the error that creates sanctions risk, spoliation arguments, and motions to compel. Rule 26(g) of the Federal Rules of Civil Procedure requires certifications that productions are complete and correct, so low recall directly threatens the certifying attorney. A false positive, by contrast, produces an over-inclusive set: privileged-by-mistake documents, irrelevant material shipped to opposing counsel, wasted hosting and review costs downstream. Over-production also erodes confidentiality and can waive privilege protections if privileged documents slip through in an over-broad batch.

This asymmetry explains why recall has historically been treated as the primary validation target in litigation review. Most published TAR guidance targets recall in the range of 75% to 95% depending on the matter's risk profile, while precision is monitored but rarely held to a hard threshold. In internal investigations, however, the calculus shifts: regulators reviewing an investigation care about whether the company found the documents, but they also scrutinize whether the methodology was disciplined enough to be credible, which pushes precision up the priority list.

How TAR Validation Actually Works: Control Sets and Elusion

Validation in TAR rests on sampling. The standard approach builds a control set (sometimes called a training or seed set) of documents reviewed by subject-matter experts, then uses statistical inference to estimate how the full population would score. Two sample types dominate practice:

First, a random sample drawn from the entire corpus estimates prevalence — the baseline percentage of responsive documents — and provides the denominator needed to compute estimated recall. If a random sample of 1,000 documents from a 500,000-document corpus finds 50 responsive documents (5% prevalence), statisticians can calculate with known confidence levels how many responsive documents likely exist in total, and therefore what a given number of missed documents implies for recall.

Second, an elusion sample is drawn exclusively from documents the system classified as non-responsive. This is the critical test for recall: if experts find responsive documents inside the discard pile at a meaningful rate, the elusion rate is too high and the model needs more training or a lower cutoff. Many practitioners target an elusion rate below roughly 2-5%, though the acceptable figure depends on prevalence — a 3% elusion rate when prevalence is 5% means the system may be discarding a large fraction of all responsive documents, whereas the same elusion rate at 40% prevalence is far less alarming.

Confidence intervals are not optional decoration. A recall estimate of 85% ± 4% at a 95% confidence level tells a court something usable; a bare '85%' does not. Sample sizes typically run from several hundred to a few thousand documents per validation round, and each round of model retraining requires fresh validation sampling before the production is certified.

Comparing TAR Validation Approaches: Classic TAR vs. Generative AI Review

The arrival of large language model-based review has complicated the validation picture. Traditional TAR — whether TAR 1.0 (simple active learning on a fixed training set), TAR 2.0 (continuous active learning, or CAL), or classifier-based workflows — produces probabilistic scores that map naturally onto recall/precision curves. Generative AI tools often produce binary judgments ('responsive' or 'not') with explanations, which is why Legal IT Insider described Gen AI TAR as 'nearly expert, entirely binary.' Binary outputs still yield confusion matrices, but the underlying behavior differs in ways validators must account for.

Reporting on the major Law.com study suggested generative AI reached performance close to experienced reviewers on complex review tasks, which is impressive but not the same as being validated. 'Nearly expert' performance on a benchmark study does not transfer automatically to your corpus, your definition of responsiveness, or your judge's expectations. The Pryor Cashman analysis of using AI in regulator-facing internal investigations makes a related point: regulators will credit AI-driven review only when the testing methodology resembles the statistical validation they already understand from the TAR era. Old tests, new machines.

A practical consequence: teams running generative AI review should still build control sets and run elusion samples exactly as they would for CAL. The metrics don't change; only the system being measured does. Teams that skip validation because the vendor demo looked strong are repeating the mistake early TAR adopters made in assuming the algorithm would simply work.

Practical Steps: Running a Defensible Validation

A defensible validation sequence looks like this. First, finalize the responsiveness definitions and negotiate any scope questions with counsel before sampling begins, because ambiguous definitions poison both the training data and the validation statistics. Second, draw a random sample of the full corpus to establish prevalence with a stated confidence level — commonly 95% confidence with margins of error in the 2-5% range, which usually means samples of roughly 400 to 1,000+ documents depending on corpus size and observed prevalence.

Third, train the model (or configure the AI prompts) using expert-reviewed seed sets, iterating until the model stabilizes. Fourth, run the classification across the corpus and select a cutoff. Fifth, draw an elusion sample from the discard pile, sized to detect whether the elusion rate exceeds your tolerance. Sixth, compute estimated recall and precision with confidence intervals and document everything: sample sizes, reviewer agreement rates, cutoff decisions, and retraining rounds. Seventh, if recall falls short, pull additional training documents (often from the region near the cutoff score, where errors concentrate), retrain, and re-validate.

Two refinements deserve attention. Inter-reviewer agreement should be measured on the control set itself — if two experts agree on only 70% of documents, no algorithm trained on their labels can exceed that ceiling, and the problem is the definitions, not the machine. And stratified sampling near the decision boundary is more efficient than pure random sampling for estimating precision, because that's where misclassifications cluster.

Common Mistakes That Undermine Validation

The most frequent error is validating once and declaring victory. Prevalence drifts as the review progresses and as new custodians or date ranges enter scope; a model validated at week three may degrade by week eight. Continuous active learning was designed partly to address this, but even CAL workflows need periodic elusion checks.

Second, teams sometimes chase recall to absurd extremes and ignore precision entirely, shipping productions stuffed with non-responsive documents. This inflates hosting costs, buries the truly important documents under noise, and creates privilege exposure. Third, some practitioners report precision and recall without confidence intervals or sample sizes, which renders the numbers unfalsifiable and invites challenge. Fourth, conflating elusion rate with recall is a persistent confusion: elusion is the contamination rate of the discard pile, while recall is the share of all responsive documents captured. They're related through prevalence but not interchangeable, and quoting one while implying the other has burned more than one team in meet-and-confer disputes.

Fifth, with generative AI specifically, teams sometimes accept the tool's self-reported confidence or its written explanation as evidence of correctness. An LLM can produce a fluent, plausible-sounding rationale for a wrong call. Only independent expert sampling reveals the truth. Finally, skipping documentation is its own failure mode: a validation you cannot reconstruct in front of a judge functions as if it never happened.

When to Act and What It Costs

Validation should begin at project kickoff, not after review completes. Budget time for control set creation (typically days to a couple of weeks depending on definitional complexity), iterative training rounds, and final elusion testing before certification. In fast-moving matters, plan validation checkpoints weekly; in stable matters, at each major scope change and before production.

Cost-wise, the incremental expense of proper validation is modest relative to total review spend. Statistical sampling and expert review of a few thousand validation documents might add low tens of thousands of dollars on a mid-sized matter — small against multi-hundred-thousand-dollar review budgets, and cheap insurance against a motion to compel or a redo. Generative AI platforms price variously by token consumption, query volume, or per-document fees, and those usage-based costs make over-inclusive (low-precision) outputs directly expensive: every false positive is a document you paid the model to classify and may pay again to host or re-review. High precision isn't just a legal-quality metric; it's a cost-control metric in per-use pricing models.

Act immediately if an elusion sample returns responsive documents at a rate suggesting recall below your tolerance, if prevalence estimates shift materially mid-review, or if a regulator or opposing counsel requests methodology details — in those moments, having current validation statistics ready is the difference between a routine response and a crisis.

The Bottom Line for Legal Teams

Recall protects you from the sanction risk of missing documents; precision protects you from the cost and confidentiality risk of producing junk. TAR validation is the statistical machinery that lets you measure both honestly, with confidence intervals a court can rely on. Generative AI review tools have raised the performance ceiling — studies reported through 2026 describe results approaching expert-level on complex reviews — but they have not changed the validation fundamentals. Control sets, elusion sampling, documented cutoffs, and honest confidence intervals remain the standard, whether the engine behind the classifications is a 2015-era classifier or a frontier language model. Teams that treat AI output as presumptively correct, and teams that fixate on a single metric, are both exposed. Measure both numbers, document the method, and re-test whenever the ground shifts.