Embracing new technology, even in the current age of rapid technological advancement, does not happen overnight. When technology assisted review based on simple active learning (“TAR 1.0”) entered the eDiscovery scene, many litigants were wary of pioneering through the unknown. Except for a handful of trailblazers, most parties were content with old-fashioned, manual eDiscovery reviews, either not sufficiently trusting the technology, having reservations about sharing what is arguably attorney work product, or simply fearing setting potentially bad precedent. Along with the rapid advancement of artificial intelligence, we now have the latest iteration of TAR - TAR 2.0 a/k/a continuous active learning (“TAR 2.0”).
What’s The Difference Between TAR 1.0 and TAR 2.0?
While both TAR 1.0 and TAR 2.0 involve training the review platform to replicate coding by attorney reviewers, the training does not stop early on in TAR 2.0 but rather is a continuous and ongoing act of refined learning, until validation. The recent case of In re Insulin Pricing Litig. provides a good of example of the level of transparency to expect by the producing party and court when undertaking a TAR 2.0.[1] In Insulin Pricing, while the parties agreed on many aspects of the TAR 2.0 process, there were three main areas of disagreement – training the TAR model, the stopping criterion, and the validation process.
How Should You Train TAR 2.0?
The parties’ first point of contention was the training phase. Plaintiffs wanted Express Scripts to use “all relevance-based coding decisions” but Express Scripts proposed to use “all reviewer determinations at the outset” but limit training input towards the end of the review to only documents that were QC’d.[2] The Court ruled that Express Scripts was permitted to use all reviewer determinations for the majority of the review, and to limit training to its use of only QC’d documents towards the end, provided that it gave plaintiffs advance notice of doing so to allow the parties to meet and confer and give plaintiff’s the opportunity to consent. If the parties reached an impasse, they could address it with the Court.[3]
When Do You Validate TAR 2.0?
The next point of contention was the stopping point before Express Scripts could proceed to validation. While the parties agreed that a stopping point before validation was necessary, plaintiffs proposed a stopping point based on a pre-determined criterion, but Express Scripts proposed a stopping point based on “the actual in-process workflow and its progress.”[4] Siding with Express Scripts, the Court ruled that “…the better course is to determine the stopping point in real time as the process proceeds.”[5] In so doing, the Court recognized that the stopping point itself is not a validation test and that it should be determined by “reasonableness and proportionality” based on quantitative as well as qualitative factors.[6]
What Does TAR 2.0 Validation Require?
Further, though the parties agreed that validation was required, they disagreed as to what it would entail. Express Scripts proposed to validate the documents that had not been QC’d, using an elusion set.[7] However, plaintiffs argued that validation should be based on a sampling taken from the entire universe of documents including those that were unreviewed.[8] Concluding that the approach proposed by Express Scripts was not “sufficiently reasonable,” the Court adopted the approach advanced by plaintiffs that had been undertaken by the Northern District of Illinois in
In re Broiler Chicken Antitrust Litig., No. 1:16-CV-08637, 2018 WL 1146371, at *4 (N.D. Ill. Jan. 3, 2018).[9] The validation sample in In re Broiler Chicken was based documents from three different categories:
(1) documents marked responsive,
(2) documents marked non-responsive by human reviewers, and
(3) documents excluded by TAR from manual review.[10]
The results of that validation review were then to be presented in a table listing all of the documents reviewed and identifying the corresponding Bates number, sample from which the document came (or if it was excluded by TAR from review), responsiveness coding by the human reviewer, and privilege coding by the human reviewer.[11] In addition, Express Scripts was to produce all responsive, non-privileged documents not previously produced.[12] Next, the parties were required to meet and confer “to determine whether or not the Parties agree that the recall estimate, and the quantity and nature of the responsive documents identified through the sampling process, indicate that the review is substantially complete.”[13] If the recall estimate and validation review of documents in categories 2 and 3 still contained a “substantial number of non-marginal, non-duplicative responsive documents” compared to those from category 1, the review would continue and be repeated.[14] If the parties were then unable to agree on whether the review was “substantially complete,” the issue would be resolved by the Court.[15] The Court declined plaintiffs’ request to require Express Scripts to disclose the category from which each responsive document was drawn.[16]
What Recall is Reasonable in TAR 2.0?
Finally, the parties disagreed as to the validation recall target. Plaintiffs wanted a recall rate “between 70% to 80%, or higher” while Express Script’s proposed a “reasonable and proportional recall rate of 70% or higher.”[17] The Court sided with Express Scripts, finding its approach “sufficiently reasonable at this stage of the process” because recall was only one factor to be considered in assessing the adequacy of the review.[18] Indeed, the Court reasoned that the adequacy of a TAR is based on both “quantitative and qualitative [factors].”[19]
Conclusion
As attorneys, we are instinctively – and understandably – extremely protective of our work product. When it comes to eDiscovery, however, it is important to recognize that when undertaking a TAR 2.0 review, flexibility, willingness to be fully transparent, and close ongoing collaboration with opposing counsel throughout the process are critical. In this case, although the parties disagreed as to how the TAR 2.0 would proceed, each party proposed reasonable and sound approaches resulting in the Court effectively splitting the baby, siding with plaintiffs on certain issues, and siding with Express Scripts on others. All the Court’s rulings were driven primarily by reasonableness, transparency, and fairness. When you consider undertaking TAR 2.0, it is important for you to be mindful of these considerations. The ability to be transparent and flexible with your ESI processes, and cooperative with opposing counsel are paramount.
[1] No. 23-MD-3080 (BRM) (RLS), 2025 WL 1112837, at *1 (D.N.J. Apr. 11, 2025).
[2] In re Insulin Pricing Litig., 2025 WL 1112837, at *3.
[3] Id. at *4.
[4] Id.
[5] Id. at *5.
[6] Id. at *4-5.
[7] Id. at *5. Elusion testing is based on a random sampling of documents from the null set (e.g., documents marked not responsive by TAR) to determine how many should have been marked responsive.
[8] Id.
[9] Id. at *7.
[10] Id. at *4-5.
[11] Id. at *5.
[12] Id. at *6.
[13] In re Broiler Chicken Antitrust Litig., 2018 WL 1146371 at *6; Recall measures the completeness of the review, e.g., the percentage of documents marked Responsive out of the entire universe of Responsive documents.
[14] Id.
[15] Id.
[16] In re Insulin Pricing Litig., 2025 WL 1112837, at *7.
[17] Id.
[18] Id.
[19] Id (internal citations omitted).
