The AI Learning Hub Journal

E-Discovery and the TAR Precedent

Technology-assisted review, and where its defensibility comes fromhumans set the standard, the classifier scales it, and sampling shows what it costSeed setA sample of documents isread and coded by peoplewho know the matter andthe issues in dispute.Human judgement, recordedClassifier trainedA classifier learns fromthose coded decisions —what relevant means inthis matter, not in general.It learns your standardCorpus rankedEvery remaining documentis scored and ordered,most likely relevantfirst, least likely last.Effort goes where it paysSampling and validationIndependent samples fromacross the ranking givean estimate of what theprocess would leave out.Measured, not assumedWHAT MAKES IT DEFENSIBLEThe protocol was written down before it was runWho did what, and when, is on the recordRecall was measured rather than assertedThe sampling was independent of the rankingThe method can be explained to the other sideWHAT DOES NOT MAKE IT DEFENSIBLEThe tool is sophisticatedThe vendor says it performs wellIt surfaced a great many documentsNobody checked what it left behindThe method was never written down anywhereDefensibility is a property of the process, not of the softwareWhat can be shown, sampled and explained is what survives a challengeThe question put to the method is never was it clever — it is can you show what it missed
Assisted review survives challenge on the strength of its documentation and its sampling, not on the sophistication of the tool

Algorithmic Review Is Not New Here

Discovery is the one area of legal practice with a long, established history of courts accepting machine-assisted review. Technology-assisted review — predictive coding — has been used for well over a decade to prioritise and classify documents for responsiveness, with courts in several jurisdictions accepting it as a reasonable method. That acceptance did not arrive because the technology was impressive. It arrived because practitioners built a defensibility apparatus around it: documented protocols, agreed seed sets, statistical sampling to estimate recall, validation against control sets, and disclosure to opposing parties. The technology was permitted because the method was measurable and could be explained. That is the transferable lesson, and it is a methodological one rather than a technological one.

  • Courts have accepted predictive coding in discovery for well over a decade in several jurisdictions
  • In US federal practice the landmark is Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012), the first judicial approval of computer-assisted review
  • Acceptance followed from measurable, documented, disclosable method — not from technical impressiveness
  • Seed sets, control sets, sampling, and recall estimation are the apparatus that made it defensible
  • The transferable lesson is the methodology, not the specific algorithm

What TAR Teaches About Generative Tools

Predictive coding and a generative model are different technologies, and the differences matter. TAR is a classifier trained on human decisions for a specific matter, producing a score whose accuracy can be estimated statistically against a control set. A general-purpose language model applied to review is not trained on your matter, produces no calibrated confidence, and can be prompted into different behaviour by the content of the documents themselves. So the TAR precedent does not automatically extend. What extends is the expectation: if you use an algorithmic method, be prepared to describe it, measure its performance, and defend the choice. Generative tools currently make that harder to satisfy, not easier, and that gap is the honest state of play.

  • TAR is a matter-specific classifier with statistically estimable recall; a general model is neither
  • Generative output has no calibrated confidence score to sample against
  • The transferable expectation is defensibility: describe the method, measure it, justify it
  • Do not cite the TAR precedent as blanket judicial approval of generative review — it is not

Privilege Review Is the Hard Case

Of all discovery tasks, privilege review carries the least tolerance for error, because the consequence of a false negative is potential waiver rather than wasted effort. Privilege depends on context a document often does not contain — who a participant is, whether legal advice was actually being sought, whether a communication was later shared in a way that broke confidentiality. Models are poor at exactly this kind of contextual inference and will confidently classify on surface features such as the presence of a lawyer in a recipient list. Machine assistance can reasonably narrow the field for human review, but a privilege call should be made by a person, and clawback arrangements should be in place regardless of how the review was conducted.

  • Privilege errors risk waiver — the asymmetry is far sharper than in responsiveness review
  • Privilege depends on context outside the document; models classify on surface features
  • Use machine assistance to narrow the field, but keep the actual privilege call with a person
  • Have clawback and inadvertent-disclosure protections in place whatever the review method
  • In US federal practice, a Rule 502(d) order under the Federal Rules of Evidence is the standard protection against waiver on inadvertent disclosure

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.