Generative Artificial Intelligence and Document Review: The Future is Now

Written by attorneys JP Harris and Lizzy Manning

Published: NH Bar News (p24)

Artificial intelligence is changing the way we practice law, and one task that can be completed with far greater efficiency using generative artificial intelligence (“GenAI”) is the review of electronically stored information (“ESI”) for discovery. GenAI presents opportunities to dramatically change traditional workflows to deliver better outcomes for clients at a lower cost.  Some clients already expect that GenAI will be used due to the cost savings and other benefits it provides.

For over a decade, ESI software platforms have been an essential tool for lawyers facing large volumes of e-discovery. ESI platforms traditionally create indices of all the words contained in all the documents that are loaded into the system. These systems do not understand the meaning of the words, however; they only recognize the frequency with which words appear in the same document. They know that “apple” frequently appears with “tree” and “pie” and “phone,” so when a user searches for “apple” they suggest for consideration those additional related words or documents containing those words. This same concept powers “technology assisted review,” where the ESI platform “learns” from coding decisions made by human reviewers and applies that learning to documents that humans would not review (typically low priority documents in the collection). If a document with “apple” was coded as relevant by a human, the system guesses that a document with “tree” or “pie” or “phone” is also relevant because of the frequency with which those words appear together. After several rounds of training, a user can direct the system to code all the remaining documents for responsiveness all at once.

Bolstered by GenAI, ESI systems now use Large Language Models (“LLMs”) to understand the meaning of words, and not rely solely on the frequency with which they appear together. Today’s platforms actually understand that there are many varieties of apples, that apples grow on trees in particular climates, and the economy of apple harvesting. Instead of searching for “apple,” litigators now summarize the concepts of their case to the ESI platform through a series of prompts, similar to how they would describe the matter to a colleague over a cup of coffee. The user describes the key parties, the relevant timeframe, the important issues, and key legal concepts. Just like a colleague, the system will ask questions for clarification when it does not understand, or make suggestions on how to improve the summary. These questions enable AI to suggest the ultimate prompt to use, but they also provide the user insight on how the system is processing the information. When a user enters a prompt, the system reaches out to the LLM for context and meaning and then deploys that context or meaning to search and score the documents in the data set.

This can be particularly useful when responding to requests for productions (“RFPs”), as an example. After summarizing the matter and the key concepts, the user can enter the text of the RFPs as prompts. The system will reach out to the LLM to understand the context and meaning of the RFPs and then it will provide a sample of documents for the user to code for relevance. From this human input, the system learns what is relevant and suggests ways to fine-tune the prompt, resulting in a refined analysis. The user will then review several rounds of sample documents to train the system. Then, the trained system will predict the relevance of each document in the data set (on a scale of 1-100, for example). Importantly, the system provides an explanation of each document’s score to provide insight into how the system interpreted the meaning and context of the prompt as applied to the document and any judgment calls the system made. The user can then prioritize which documents to review based on their scores. For example, a user may want to start to review documents with a score of 90 or higher. After the user is satisfied with their review, the system can automatically code the remaining documents based on each document’s predictive score.

GenAI powered ESI platforms offer numerous timesaving benefits to lawyers. However, lawyers using these platforms must be aware of the potential ethical issues. Importantly, for client confidentiality, the LLMs in ESI platforms must operate as a closed system to maintain confidentiality. Lawyers should obtain consent from their client prior to using GenAI, due to the potential of GenAI created errors. Use of GenAI may also force law firms to examine their billing practices for GenAI assisted tasks. With GenAI powered ESI platforms, it is possible for an individual to conduct a first-pass review of 200,000 documents in 30 minutes. Without GenAI, a human reviewer would spend hundreds of hours reviewing the same number of documents (costing the client thousands of dollars). Firms may need to shift to charging clients for discovery tasks on a project basis, as opposed to a time basis. Firms may consider charging a flat rate for a first pass review based on the value of that task to the client’s matter, even though the number of attorney-hours actually devoted to the task are greatly lessened through the use of GenAI. Lawyers may also want to consider negotiating ESI stipulations with opposing counsel to address whether GenAI will be used in the case and how (i.e. do the parties agree that a relevance score of 85 is sufficient?).

An ESI system that incorporates GenAI is a powerful tool with cost and time saving benefits, particularly for e-discovery. Harnessing those benefits requires changing workflows to incorporate them into the discovery process. In the near future, these systems will be commonplace in the industry and something clients expect, so now is the time for lawyers to keep up with GenAI and incorporate it into their practices.