If GenAI a replacement for TAR? We’ll see. But this Sedona Conference Journal article discusses unifying traditional and GenAI approaches to TAR.

The article (TAR 1 Reference Model: An Established Framework Unifying Traditional and GenAI Approaches to Technology-Assisted Review, available here) was written by Tara Emory, Jeremy Pickens & Wilzette Louis of Redgrave Data and released yesterday.

For well over a decade (and court approved for over twelve years), Technology-Assisted Review (TAR) has been used for document review in discovery, but now the emergence of GenAI has sparked discussions about its potential as an alternative. There has been a lot of speculation on whether Generative AI (GenAI) will eventually be a replacement for TAR, or if it will be a complement to TAR.

From the perspective of the authors in this article, GenAI and TAR are complementary, not conflicting. The article introduces GenAI as a new prediction algorithm that can and for maximal effectiveness should be applied according to established TAR workflows (specifically, the workflow known as “TAR 1”). The article presents algorithm-agnostic steps of this workflow through a TAR 1 Reference Model and contains diagrams of the specific tasks for traditional (which the authors refer to as “discriminative” because they separate, or discriminate, between positive and negative labels) and GenAI algorithms.

To summarize at an extremely high level, the traditional workflow for TAR 1 includes the following steps: Scope, Label Control Set, Iterate Model, Classify and Validate (optional). As the name implies, the Iterate Model contains four steps of its own – Evolve, Encode, Apply and Evaluate – which are repeated until the assessment of quality & improvement is considered complete.

One thing notable about the comparison between the traditional Discriminative TAR 1 approach and the GenAI TAR 1 approach is that they have the same steps and sub-steps, so they are very similar. The key difference is in the Iterate Model steps of Evolve, Encode, and Apply, where the creation, application, and iteration of a GenAI prompt replaces traditional steps of selection and review of additional training documents, as follows:

The end result is unifying traditional and GenAI approaches to TAR with the GenAI TAR 1 approach. This discussion of the workflows and their (somewhat subtle) differences is the main portion of the article.

However, the article does go on to discuss several very important considerations for GenAI versus Discriminative algorithms in TAR 1, including: 1) precision and recall, 2) risk of sensitive information disclosure, 3) knowledge gain and accomplishment of related tasks, 4) total project cost, 5) total project time, 6) ease of use, and 7) whether different algorithms may best serve different needs. These are important considerations for any automated approach, with or without TAR and/or GenAI. The authors also discuss hybrid workflows and that GenAI may be integrated with other eDiscovery tools to yield even more possibilities for TAR 1 and other workflow improvements.

The authors note in the conclusion that “Especially as GenAI capabilities increase and costs and time it requires go down, GenAI has potential to become a preferred approach to TAR 1”. But, in the meantime, they’ve provided an approach that still follows the established steps of TAR 1, involving sampling and statistics, that “will help promote successful outcomes on first-pass review projects for practitioners using GenAI — as it has helped those same practitioners when using discriminative approaches.”

Again, you can check out the 29-page article here, which provides a much more detailed discussion of what I just summarized. It will be interesting to see how eDiscovery professionals respond to this proposed approach.

So, what do you think? Is unifying traditional and GenAI approaches to TAR a good idea? If so, is this a solid approach to doing so? Please share any comments you might have or if you’d like to know more about a particular topic.

Image created using GPT-4’s Image Creator Powered by DALL-E, using the term “robot holding a chocolate and vanilla swirl ice cream cone”.

Disclaimer: The views represented herein are exclusively the views of the author, and do not necessarily represent the views held by my employer, my partners or my clients. eDiscovery Today is made available solely for educational purposes to provide general information about general eDiscovery principles and not to provide specific legal advice applicable to any particular circumstance. eDiscovery Today should not be used as a substitute for competent legal advice from a lawyer you have retained and who has agreed to represent you.