In the rapidly evolving digital landscape, data management has become increasingly complex. Traditional e-discovery processes are grappling with the formidable challenges posed by the rapid surge in data volume, velocity, and diversity.
E-Discovery’s New Chapter
During the recent ILTACON 2023 session titled, “Revolutionizing e-Discovery Again?: Harnessing the Power of Portable Active Learning Models”, experts delved into alternatives to familiar technology-assisted review (TAR) and continuous active learning (CAL) methods. Central to their discussion was the promising evolution of portable artificial intelligence (AI) models, touted as the future for e-discovery professionals navigating the overwhelming ocean of data.
Decoding Portable Models
The essence of portable models is their adaptability and transferability. Imagine utilizing the TAR or CAL process for one matter and seamlessly transitioning its application to another. Once trained in a particular scenario, these models can be conveniently saved, packaged, and repurposed in varied contexts.
Krysta Slavik, an authority at K&L Gates, remarked on this development: “As datasets expand, the demand for tools that rapidly sift through them intensifies. Portable models are among those pivotal tools.”
The flexibility of these models is evident in their ability to be custom-built, from intricate details to broader strokes. Their prowess lies in identifying common themes that persist across different matters. George Socha of Reveal provided an intriguing analogy, “Think of them as Lego blocks; individualized yet capable of forming cohesive structures. Instead of a singular overarching model, professionals can now fragment it into distinct components. This allows for a focused, strength-based approach.”
The Ideal Context for Portable Models
These models are particularly advantageous for professionals dealing with recurrent issues across varied matters. James Park from DISCO elucidated, “For professionals overseeing employment matters, from harassment complaints to performance issues, these recurring themes make a compelling case for the use of portable models.”
But their application isn’t just singular. George Socha highlighted another perspective, “These models aren’t merely for locating relevant content; they can be inverted to identify content that isn’t a priority.”
Their versatile nature makes them invaluable for legal professionals racing against tight deadlines, aiding in areas from cyber incident responses to keyword selections. Highlighting another use case, Scott Milner from Morgan, Lewis & Bockius mentioned, “Compliance departments often grapple with tracking text message usage. We’ve collaborated with clients to craft portable models aimed at identifying offline communication channels.”
Potential Pitfalls and Safeguards
As with most advancements, portable models aren’t without their challenges. Analogous to TAR or CAL, the accuracy and efficiency of these models rest heavily on the input data. Krysta Slavik warned, “The age-old concept of ‘garbage-in/garbage-out’ is relevant here. For these models to be efficacious, validation of results is paramount.”
Concerns surrounding portable models mirror those associated with other AI systems. Questions arise: Has client data been illicitly utilized to construct these models? Do they infringe upon copyright norms? Is there a risk of client data leakage?
However, a critical distinction was underscored by the panelists. Unlike other Longitudinal Latent Models (LLMs), data employed in training portable models isn’t shared or retained similarly. Krysta Slavik advised a proactive approach, “Before applying a model to a new matter, examine it. Familiarize yourself with its architecture. Inform your clients. In isolation, these models are abstract; their relevance emerges only upon application.”
