What is technology-assisted review (TAR)?

TAR emerged as a response to review volumes that made linear, document-by-document review impractical. Rather than reading everything, a team trains a model on a set of human coding decisions and uses it to rank or classify the remaining population.

Two broad approaches are commonly distinguished.

TAR 1.0 trains a model on a control set coded by a senior reviewer, applies it to the full population once, and then reviews the documents it ranks as likely responsive. Training happens up front and then stops.

TAR 2.0, or continuous active learning, retrains as reviewers work. Each new coding decision refines the ranking, so the most likely responsive documents keep rising to the top. This is now the more common approach.

TAR has a substantial history of judicial acceptance in the United States and elsewhere, generally on the basis that the process used was reasonable and documented, rather than because any particular tool was approved. Courts have tended to focus on validation: what recall was achieved, how it was measured, and whether the protocol was disclosed to the other side.

The main limitation is that TAR learns from example rather than from instruction. It needs enough coded documents before it becomes useful, and if the definition of responsiveness shifts mid-review, the model has to be retrained.

Newer language-model approaches take the opposite path: the criteria are stated in plain language and applied directly to each document, with a citation to the passage supporting each call. That removes the training-set requirement and makes individual decisions inspectable.

Related reading: Sedona Canada and the New Reality of Technology-Assisted Review in 2026.

Claira works this way. You can see the difference on your own data.