Technology-assisted review (TAR) has more than a decade under its belt, so while it’s not cutting-edge in today's AI-driven legal landscape, its impact on eDiscovery is undeniable. TAR paved the way for the more advanced solutions we see today and understanding this journey informs us how eDiscovery evolved along with it.
TAR 1.0
Prior to TAR’s development, document review meant reviewing each document individually. It was less like searching for a needle in a haystack and more like searching for a specific needle in a pile of needles. Then, TAR emerged, a software sidekick programmed to identify potentially relevant documents.
The turning point for TAR came in 2012. Magistrate Judge Andrew Peck issued a ruling in Da Silva Moore v. Publicis Groupe 287 F.R.D. 182 (S.D.N.Y. 2012), endorsing TAR while emphasizing lawyer oversight and proper methodology. This paved the way for wider adoption and spurred investment in additional eDiscovery capabilities such as early case assessment, concept searching, advanced analytics, and others.
TAR was a substantial step forward - a 2013 Vanderbilt paper stated TAR improved attorney review speed by 70% - but software marketers got a bit over their skis.
Although TAR could deliver substantial time and cost savings, it was no “easy button.” It usually required multiple training rounds wherein counsel QC’ed a subset of the document corpus reviewed by associates or a managed review provider to achieve certain recall thresholds – only then could review proceed for documents over the recall threshold. It was not uncommon for counsel to complain that TAR took too many iterations and too much time.
TAR 2.0 (aka, Continuous Active Learning)
Enter TAR 2.0, aka Continuous Active Learning, or CAL.
This development was more flexible because the model continuously learns rather than applying fixed decisions from counsel on the initial seed set. It can also more easily account for new documents that are promoted to the workspace.
Additionally, depending on which process you use, counsel can opt for coverage review, wherein the software feeds reviewers random documents to determine a cutoff between what’s responsive and not responsive, and prioritized review, wherein the software feeds reviewers documents similar to those they’ve coded responsive from an automated queue.
The Dawning Generative AI Era
In 2023, software providers including Relativity, Everlaw, DISCO, Reveal, and others, started dabbling with Generative AI features.
Most of this technology is in a beta or limited release phase at the time of this writing. Anecdotally, the industry seems to be coalescing around a fact finding/question-and-answer workflow (e.g. “Did Custodian X know about fraud at Company Y?”) and a document review workflow (e.g. code all responsive documents).
My crystal ball is cloudy, so I don’t know which software will gain the most favor in the marketplace, which workflows will be seen as most efficient, or how soon a Court decision will endorse the use of Generative AI in eDiscovery, but I do know that technology will continue to advance. Finding defensible, ethical uses that increase the just, speedy, and affordable access to justice is a worthy pursuit.
Conclusion
So, is TAR a stepping-stone or an ancient relic? Neither, as it turns out. It’s still in use. While overshadowed by successors, TAR remains a useful tool. This is especially true in the context of reviews for the Department of Justice, which, as recently as 2022, had very specific requirements that lend themselves to a TAR 1.0 review workflow.
eDiscovery providers continue to innovate so that counsel can handle the ever-growing volume and variety of discoverable data, and TAR paved the way for advanced AI-powered solutions, forever changing the eDiscovery landscape.
Want to read more? Check out Proteus' Read my "Section for the Skeptics" that addresses the top 4 concerns about CAL and Generative AI in document review. Or download a full copy of the eBook here.
