What Is Active Learning in eDiscovery?

In machine learning, active learning describes a model that participates in its own training. Instead of learning from a fixed, pre-labeled dataset, the algorithm selects the examples a human should label next. It picks the ones that will teach it the most, so it needs far less labeled data to reach the same accuracy.

In eDiscovery, active learning sits inside the review stage as the engine behind modern technology assisted review (TAR). Reviewers code documents, the model learns from every decision, and the queue of documents reorders around what the model has learned.

How Does Active Learning Work?

Active learning runs as a loop between reviewer and model, repeated until the review is done.

• Reviewers code a batch of documents as relevant or not relevant.
• The model retrains on those decisions and re-scores the collection.
• It selects the next batch and serves it to reviewers.
• The loop repeats, and each pass sharpens the model.

The selection step is what makes the method active. Systems typically choose documents in one of two ways. Relevance feedback serves the documents the model currently ranks most likely relevant. Uncertainty sampling serves the documents the model is least sure about, since those teach it the most.

Either way, no separate training phase is required. The review itself is the training.

What Is Continuous Active Learning (CAL)?

Continuous active learning, or CAL, is the protocol most large reviews now use, and it is what practitioners mean by TAR 2.0.

Earlier TAR protocols separated training from review. Experts coded a seed set, the model stabilized, and it then scored the collection in a single pass. Simple active learning (SAL) kept that training phase but let the model pick its own training documents, usually through uncertainty sampling.

CAL removes the boundary entirely. The model learns from every coding decision, keeps re-ranking the remaining documents, and never stops training. New documents can join a rolling collection mid-review, and the model folds them in. Reviews typically continue until batches contain few relevant documents, which signals the collection is close to exhausted.

Why Does Active Learning Matter?

Document review remains the most expensive stage of eDiscovery, and active learning changed its economics. Because the model prioritizes continuously, teams often review only a fraction of a collection before relevant documents stop appearing.

It also removed the most awkward requirements of earlier TAR. There is no seed set to stabilize, no fixed training phase, and no penalty when collections arrive on a rolling basis. Court acceptance follows the same line as TAR generally, provided the process is transparent, documented, and validated with metrics such as recall.

‍

Frequently Asked Questions

What is active learning in eDiscovery?

Active learning is a machine learning method where the model selects which documents reviewers code next and learns from every decision. It powers modern technology assisted review by continuously ranking a collection by likely relevance.

Is active learning the same as TAR?

Not exactly. TAR is the broader practice of using machine learning in document review. Active learning is the training method behind TAR 2.0, where the model learns continuously instead of from a fixed seed set.

What is continuous active learning (CAL)?

CAL is the protocol where the model never stops training. It learns from every coded document, re-ranks the rest in real time, and absorbs new documents added mid-review. It has become the dominant approach in large reviews.

Is this the same as active learning in education?

No. In education, active learning refers to teaching techniques that engage students directly. In eDiscovery and machine learning, it refers to a model that chooses its own training examples.

‍