TAR and Predictive Coding in 2026: What Litigators Need to Know

More than a decade after Da Silva Moore blessed predictive coding, technology-assisted review is no longer novel — it is expected. But the fights have shifted from whether you can use TAR to how you validate it and what you must disclose. Here is the 2026 state of play for litigators.

TAR is now routine on large matters, and courts increasingly view proportionality through the lens of whether a party used available technology efficiently. The defensibility question has moved downstream: not 'may we use predictive coding,' but 'can you prove the result was adequate.' That puts validation methodology at the center of the dispute.

TAR 1.0 vs. TAR 2.0 — and why it matters

The distinction still drives strategy. TAR 1.0 (simple or simple-active learning) trains a model on a control set, then classifies the population in a single pass — efficient when the collection is stable and the issues are well defined. TAR 2.0 (continuous active learning, or CAL) keeps feeding reviewer decisions back into the model, prioritizing the most likely responsive documents until returns diminish. CAL has become the default on rolling collections because it tolerates evolving issues and late-arriving data.

Practical consequence

With CAL, the meaningful metric is not a one-time training accuracy but the point at which continued review stops surfacing responsive material. Litigators should expect their expert to define a defensible stopping criterion in advance rather than reverse-engineering one to justify where review happened to end.

Validation is the whole ballgame

A TAR result is defensible only if it can be measured. The core metrics remain:

• Recall — the share of truly responsive documents the process actually found; this is the number opposing counsel will attack.
• Precision — the share of documents tagged responsive that really are, which speaks to efficiency and over-collection.
• Elusion testing — sampling the discard pile to estimate how many responsive documents were left behind, the most persuasive evidence of adequacy.

Each estimate needs a stated confidence level and margin of error from a properly drawn random sample. An expert who reports a recall figure without the underlying sample design and error bars is offering a number that will not survive cross.

The disclosure fight

Parties continue to spar over how much of the TAR protocol must be shared — seed sets, training decisions, validation results. The pragmatic trend favors transparency about the validation outcome and the protocol while protecting the mental impressions embedded in individual coding calls. Negotiating these terms in the ESI protocol up front avoids a mid-production motion when the stakes are highest.

Where generative models fit

Large language models are increasingly layered onto review for issue coding, privilege triage, and summarization. They do not displace the validation discipline above — if anything, they raise the bar, because a model that classifies in ways a human cannot fully audit demands even more rigorous sampling to prove the result. Treat these tools as additions to a measured workflow, not replacements for one.

Retain validation expertise early

The most expensive TAR mistakes are baked in before the first document is reviewed — in protocol terms, sample design, and stopping criteria. An expert who can validate the workflow and defend the numbers under cross is worth far more at the planning stage than as a clean-up after a challenge. If TAR is central to your matter, reach out through our home page to scope the engagement.