Traditional TAR needs a seed set: attorneys code examples, and the model ranks or predicts the rest. Generative AI review does not. It applies a written protocol or prompt to the corpus, so first-pass results can start on upload.
Courts are nonetheless starting to treat GenAI as a form of TAR for disclosure and reasonableness — see Schulte v. LinkedIn, 2026 WL 1905851 (N.D. Cal. June 30, 2026). “No seed set” is a technical difference, not a free pass on validation.
The mechanical difference
TAR 1.0 trains on a coded control set, then predicts. TAR 2.0 / CAL keeps learning as reviewers work and is iterative. Both need attorney judgments as examples and struggle on rare or unique documents that never appeared in the training pool.
A generative classifier does not wait for those examples. It needs a prompt that reads like a review memo, and it can usually explain a coding call in prose. It is typically not iterative once the prompt has been run. That moves validation earlier: test the prompt on a sample, measure precision and recall, revise, then scale.
What you still owe the court
Human review of borderline calls, a record of the tool and settings, and metrics you can defend. Rule 26(g) still sits on the lawyer’s signature. DecoverAI’s document review does not require a seed set; attorneys still override every classification before production.
When you would still pick classic TAR
Huge, homogeneous collections with a stable responsiveness definition and a review team already in CAL can still be a good TAR 2.0 matter. GenAI is the better default when you need privilege, confidentiality, and responsiveness in one pass, when you cannot wait to build a seed set, or when you need a sentence of reasoning on a coding call.
Hybrid workflows exist: keywords or TAR to cull, then GenAI on the remainder — the pattern in Schulte. Whatever you pick, write down why it fits this collection and this budget. That paragraph is the proportionality record, and it is lawyer work, not a vendor feature.
The cost story follows the same split. Classic TAR’s expensive step is the attorney time that builds and refreshes the seed set. GenAI’s expensive step is prompt design and the sample you score before you scale. Neither one is “set and forget,” and neither one is priced honestly if you hide QC.