Q: How do active learning and traditional predictive coding differ in practice?

With traditional predictive coding, attorneys build and code a seed set first, then the model predicts across the full population. With active learning, the model starts predicting immediately and improves as reviewers code, eliminating the separate seed set phase. Active learning is generally more efficient, particularly on large, heterogeneous document populations.

Q: How do I know when active learning review is done?

Completion is assessed through recall validation sampling — reviewing a random sample of model-low-scored documents and measuring actual relevance. When the estimated relevant population in the unreviewed documents falls below an agreed threshold, review is complete. The appropriate threshold is a legal judgment based on case needs, not a fixed technical standard.

Q: Does active learning work on small document sets?

Active learning provides the most efficiency gain on large document sets (100,000+ documents). On smaller sets, the efficiency advantage over linear review or simple keyword filtering is less pronounced. For very small document sets, linear review may be more straightforward. --- *Last reviewed: 2026-05-19 by LawyerAI Editorial Team.*

Active Learning (eDiscovery)

An iterative ML approach in eDiscovery where the model continuously updates relevance predictions as reviewers code documents, prioritizing the most uncertain documents for review.

Last reviewed: 2026/05/19

Definition

Why It Matters for Lawyers

How AI Tools Handle It

Frequently Asked Questions

Q: How do active learning and traditional predictive coding differ in practice?: With traditional predictive coding, attorneys build and code a seed set first, then the model predicts across the full population. With active learning, the model starts predicting immediately and improves as reviewers code, eliminating the separate seed set phase. Active learning is generally more efficient, particularly on large, heterogeneous document populations.
Q: How do I know when active learning review is done?: Completion is assessed through recall validation sampling — reviewing a random sample of model-low-scored documents and measuring actual relevance. When the estimated relevant population in the unreviewed documents falls below an agreed threshold, review is complete. The appropriate threshold is a legal judgment based on case needs, not a fixed technical standard.
Q: Does active learning work on small document sets?: Active learning provides the most efficiency gain on large document sets (100,000+ documents). On smaller sets, the efficiency advantage over linear review or simple keyword filtering is less pronounced. For very small document sets, linear review may be more straightforward. --- *Last reviewed: 2026-05-19 by LawyerAI Editorial Team.*

Related Tools

DISCO
AI-native legal technology platform for eDiscovery, case building, and legal holds used by Am Law 200 firms.
Logikcull
Self-service eDiscovery platform designed for instant setup, used by solo firms through Fortune 500 legal teams.

AI Tools by Practice Context

Move from this definition to role-based legal AI shortlists and the selection criteria that matter for each type of legal team.

AI Tools for Big Law
Am Law 200 and global firm workflows: accuracy at scale, security compliance, and matter-level auditability.

Browse all legal AI solutions by role

Related Reading

Last reviewed: 2026/05/19. Definitions are written by the LawyerAI Editorial team. Commercial relationships are disclosed and do not determine editorial scores or conclusions. See our Sponsorship & Affiliate Disclosure.

← All glossary terms

Active learning is an iterative machine learning methodology used in eDiscovery document review in which the predictive model continuously updates its relevance classifications as reviewers code documents, rather than being trained once on a fixed seed set. The system prioritizes presenting reviewers with the documents on which the model is most uncertain, incorporating each coding decision to improve subsequent predictions. This continuous feedback loop makes active learning more efficient than traditional predictive coding — the model improves throughout production review rather than only during an initial training phase. Also referred to as continuous active learning (CAL).

The efficiency advantage of active learning over traditional predictive coding is substantial in large document sets. By prioritizing uncertain documents — those the model cannot confidently classify — active learning directs reviewer attention to the documents where human judgment adds the most value, while confidently classifying the low-uncertainty population without review.

For litigation teams, active learning means that the review workflow itself is the training process. Reviewers do not need to complete a separate seed set training phase before production review begins. Review starts immediately, and the model improves in real time.

The operational implication is that review quality at the beginning of an active learning project affects model quality throughout. Inconsistent or incorrect coding early in the review — before the model has trained sufficiently — has downstream effects. Review teams using active learning should invest in reviewer training and quality control, particularly in the early stages of review.

Active learning also requires thoughtful decisions about when review is complete. Unlike linear review with a defined endpoint, active learning review concludes when a validation sampling process confirms adequate recall — a legal judgment call, not a system-generated endpoint.