There is a particular irony in how the legal profession has approached e-discovery automation over the last decade: the industry spent years building a careful, court-tested record establishing that algorithmic document review — technology-assisted review, commonly called TAR or predictive coding — could be trusted, defensible, and even superior to exhaustive manual review, only to now face a new generation of generative AI tools that threaten to reopen exactly the defensibility questions that earlier record was built to settle. Understanding what made TAR defensible in the first place is the clearest guide to what generative AI tools need to demonstrate before they earn the same trust.
1. How TAR Earned Judicial Acceptance
The turning point for predictive coding's judicial acceptance is widely traced to Da Silva Moore v. Publicis Groupe, decided in the Southern District of New York in 2012, in which Magistrate Judge Andrew Peck approved the use of predictive coding for the first time in a reported federal decision, reasoning that the technology, properly validated, was at least as accurate as exhaustive manual review and considerably more proportionate to the scale of the document population involved. What made that acceptance possible was not the underlying machine-learning technique in the abstract — it was a specific, documented validation methodology: a statistically sound sample, a measured recall and precision rate, and a transparent process the opposing party could scrutinize and, if warranted, challenge.
Federal Rule of Civil Procedure 26(b)(1)'s proportionality standard, and Rule 26(g)'s certification requirement that a reasonable inquiry supports a discovery response, together provided the legal framework TAR needed to satisfy, and the technology's defenders spent years building the validation practices that satisfied it. That validation discipline — measurable accuracy, a documented methodology, transparency sufficient for adversarial scrutiny — is precisely what a generative AI tool applied to document review or privilege logging needs to reproduce, and it is precisely what many current-generation tools have not yet built a comparable record for.
2. Where Generative AI Genuinely Differs From TAR
Classic predictive coding is fundamentally a classification task: relevant or not relevant, responsive or not responsive, a binary or small-category decision that lends itself naturally to the kind of recall-and-precision measurement Da Silva Moore's framework depends on. Generative AI applied to e-discovery increasingly does something structurally different — summarizing documents, drafting privilege log entries, synthesizing a narrative across a document set — tasks that do not reduce as cleanly to a measurable accuracy statistic, because "is this summary a fair and complete representation of the underlying document" is a qualitatively different question from "is this document responsive," and it resists the same statistical validation methodology that made TAR defensible.
This does not mean generative AI cannot be used defensibly in e-discovery — it means the defensibility argument has to be built differently, around traceability and human verification rather than around a single aggregate accuracy statistic. A generative summary that can be traced back, sentence by sentence, to the specific underlying document language it was derived from, and that was reviewed by a human before being relied upon, has a defensibility story available to it that a summary produced and used without that traceability does not.
3. The Sedona Conference's Continuing Relevance
The Sedona Conference's principles on e-discovery and, more recently, on the use of generative AI in the legal process, provide the closest thing the field has to an emerging consensus on what defensible use looks like, emphasizing proportionality, documented methodology, and appropriate human oversight calibrated to the stakes of the specific task. Firms and legal departments building AI-assisted e-discovery workflows should treat Sedona's principles as the current best articulation of what a court is likely to expect, in the absence of a settled body of case law specific to generative AI comparable to what Da Silva Moore and its progeny built for predictive coding.
This matters practically because opposing counsel, increasingly sophisticated about AI's limitations, are already probing exactly this gap in deposition and in discovery disputes — asking not merely "did you use AI" but "what was your validation methodology, and can you produce it," questions a party using generative AI without Sedona-style documentation may struggle to answer as confidently as a party that built its predictive-coding validation record a decade ago.
4. Architecture as the Practical Answer
A masked secure-cloud deployment, stripping privileged and personally identifying content before AI processing occurs, remains appropriate for large-scale, lower-sensitivity first-pass review across a document population where speed and scale matter most. A fully local deployment — keeping privileged material entirely inside the firm's own infrastructure — is the more defensible posture for the specific subset of work where privilege risk is highest: privilege log drafting, work-product analysis, and any AI-assisted task whose output will itself become part of the record a court or opposing party may scrutinize.
The deployment decision here reinforces, rather than substitutes for, the traceability and human-review discipline discussed above — an architecture that protects the data does not by itself make a generative AI tool's substantive output defensible, but it removes one entire category of risk from a defensibility challenge that already has enough novel ground to cover.
5. Building the Validation Record Before You Need It
The practical lesson from TAR's decade-long path to acceptance is that the validation record has to be built proactively, before a court forces the question, not reconstructed defensively after opposing counsel challenges a specific AI-assisted work product. Firms adopting generative AI for e-discovery tasks should establish, in writing, before deployment: what tasks the tool is used for, what human review step follows every AI-generated output before it is relied upon, and what sample-based quality checks are run periodically to confirm the tool continues to perform as expected as document populations and case types vary.
This is not meaningfully different in spirit from what TAR's early adopters did a decade ago — it is the same discipline, applied to a technology whose specific failure modes are different enough that the old validation methodology cannot simply be copied wholesale, but whose underlying principle, defensibility through documented, transparent process, has not changed at all.
6. Conclusion
Generative AI in e-discovery inherits the defensibility bar that predictive coding spent a decade establishing, without automatically inheriting the specific validation methodology that met it, because the underlying tasks are often structurally different in ways that resist the same statistical proof. An architecture built around masked or fully local processing for privilege-sensitive work, combined with genuine traceability and documented human review, is what allows a firm to use generative AI in discovery without surrendering the defensibility ground the profession worked a decade to win.