Predictive coding is the use of machine learning to prioritise or classify documents for review: attorneys code a sample, the system learns from those decisions, and it ranks the remaining population by likely relevance. Courts have accepted it for well over a decade where the methodology is disclosed and validated.
How it actually works
A reviewer codes documents as responsive or not. The system builds a model from those decisions and applies it to the rest of the collection, producing a relevance ranking. Review then proceeds in ranked order, and the model is retrained as more decisions come in.
Two broad designs exist. In the older approach, a control set is coded first, the model is trained on it, and the ranking is then applied to the population. In continuous active learning, there is no separate training phase — the model updates constantly as review proceeds, feeding reviewers the highest-ranked unreviewed documents next. Continuous learning has largely won on both cost and simplicity, and it removes the seed-set disputes that generated much of the early case law.
Why it is cheaper without being worse
The economics come from the ranking, not from the machine reading documents instead of people. In a typical collection, responsive documents are a small fraction of the whole. Linear review reaches them at a constant rate; ranked review front-loads them, so review can stop once the recall achieved is defensible and the remaining unreviewed population is demonstrably low-yield.
The research point that surprises people is that this is not a trade of accuracy for cost. Human linear review is measurably inconsistent — two reviewers agree less often than intuition suggests, and the same reviewer is not perfectly consistent with themselves. A validated model applied uniformly frequently outperforms it.
Validation is the part that matters legally
What makes a predictive coding process defensible is not the algorithm. It is the measurement: a statistically sound sample of the documents the process excluded, reviewed by humans, producing a recall estimate with a stated confidence interval.
Without that, a party is asserting the process worked. With it, they can show it. Any dispute about methodology is far more productively fought over the validation protocol than over the choice of tool.
What to settle before starting
- Disclosure. Whether the use of TAR is disclosed, and in what detail. Most protocols now require disclosure; fewer require exposing training decisions, and demanding them is often a fight not worth having.
- The validation standard. Target recall, sample size, and who reviews the sample.
- What is excluded from the population. Spreadsheets, images, foreign-language material and short chat messages behave poorly in text-classification models and are usually handled separately.
- The stopping rule. Agreed in advance, or it becomes a dispute at exactly the moment one party wants to stop.
Where it does not fit
Small collections do not justify the setup. Populations that are mostly non-text — engineering drawings, media files, structured data — give the model little to learn from. And a matter turning on a handful of known documents does not need a ranking; it needs targeted searching. Reaching for TAR because it is modern rather than because the population warrants it adds cost and complexity to a problem that did not have either.
From our work
Dealing with predictive coding in a live matter?
Our examiners and testifying experts work these questions for a living. Tell us what you're facing.
Reviewed by Law & Forensics. See our editorial standards.