by Joseph Rinaldi
The Age of Artificial Intelligence (AI) is upon us, impacting nearly every sector of society: healthcare, education, finance, and the legal industry. There are multiple reasons why the legal industry has quickly adopted AI, and these reasons shine a light on the current state of the legal industry at large. First, while the legal services industry today is in a healthy place, legal demand is no longer at the unsustainable growth rate that existed in 2021. Many firms added lawyers to address the previous growth rate shown in 2021, and as a result, productivity is down while legal costs remain high. Secondly, hourly rates have been rising over the past two years, leading corporate legal departments to push back against rate hikes by keeping more work in-house to cut costs. This has been especially true for corporate and transactional needs that do not involve large-scale mergers and acquisitions, capital markets, or securities work. The final reason, which lends credence to the previous two, is that efficiency is being prioritized like never before, mainly in the form of workflow automation and integrated solutions. Identifying some of the underlying reasons the legal industry has adopted AI technology provides context for how AI specifically has been integrated into the legal profession.
An understanding of how AI has been integrated into the legal profession requires an examination of two key kinds of AI: extractive and generative. Extractive AI contains technology that pulls relevant data points from the information that it has been trained on previously. Alternatively, generative AI creates new outputs based on the data it has been exposed to, is connected to, or is engineered to answer. Unlike extractive AI systems, which are designed to recognize patterns, extract pre-existing data and make predictions, generative AI generates new content in the form of text, images, and more. Thankfully, one does not need to be a software engineer to understand generative AI and its underlying technology. However, understanding a bit about Large Language Models (LLMs) and how they operate is helpful.
Different LLMs have varying strengths and weaknesses. There are five core differences between LLMs that are used to power various legal AI technologies in the market. The first difference deals with the architecture of the LLM. LLMs have different underlying “neural network architectures” that impact their capabilities. For example, some are better at certain tasks such as translation or summarization. The second core difference is the size of the LLM. LLMs can range from millions to trillions of parameters. Larger models are generally more capable, but smaller models can sometimes be more efficient. The third difference centers on the training data. The data used to train LLMs affects their knowledge and performance. Models trained on legal data will have different strengths than those trained on general purpose text. The fourth core difference revolves around fine-tuning. LLMs can be fine-tuned on niche datasets to improve their domain-specific capabilities. The fifth and final difference is whether the LLM is public or proprietary. Open source LLMs allow transparency while proprietary models offer a deeper understanding of the user’s intent to deliver higher-quality responses.
It is important to note that some generative AI utilizes a multi-model approach that draws from more than one LLM in the creation of a new tool. By combining the outputs of different models, the overall predictions and performance surpasses that of any one model, which allows users to benefit from the unique capabilities of each LLM while balancing out each one’s weaknesses. Upon choosing to evaluate different legal generative AI tools, be sure to keep the aforementioned core differences between LLMs in mind and be sure to ask the provider questions about their generative AI model. It is vital for anyone interested in adopting a legal generative AI tool to gather as much information as possible relating to the product’s underlying LLM.
In addition to finding out information regarding a generative AI tool’s LLM, one should also seek out data surrounding what information has been used to train the generative AI tool. It is important to know that any legal research solution is only as good as the breadth and depth of the information repository from which it draws its answers to your search queries. One should choose a generative AI tool that is powered by a global database containing accurate legal content and utilizes semantic search technology to understand a question’s intent and pick up on related terms and contextually relevant documents to surface the most comprehensive, accurate set of results possible relative to a query.
There are three key measures that one can use to understand the inner workings of the generative AI tool and help assess the accuracy and quality of its answers. The first measure is the comprehensiveness of results. LLMs require massive data sets, so the generative AI tool should be able to draw from a large repository of accurate and up-to-date legal content that serves as the grounding data for the model to deliver comprehensive results grounded in authoritative content. The second measure revolves around the semantic search. Semantic search can understand the underlying meaning of a search query, reading between the lines of the words that are typed to grasp intent, and then match the query to related concepts. This is distinct from keyword search, which simply retrieves answers that match the text entered in the search box. Semantic search is a superior model for legal generative AI because it increases the precision of results, delivering answers that are more relevant and saving valuable time required to wade through extraneous information. The third measure involves citation validation and grounding. Legal practitioners are by now very familiar with the fact that early adoption of generative AI by law firms was not without its troubles, most notably the risk posed by open web generative AI tools that infamously “hallucinated” various case citations that did not exist in real life. Traditional generative AI models struggle with legal use cases because the underlying content feeding the models may be dated, may lack citation authority, and may be prone to factual and conceptual hallucinations.
With these three measures in mind, there are some questions that one should consider when evaluating a legal generative AI tool. First, what is the size of the primary and secondary law database that the tool accesses to surface authoritative legal content in response to search queries? Another question would be whether the generative AI tool requires a separate subscription to access those primary and secondary sources or is it all integrated under one product experience? Third, does the generative AI output provide in-line citations and link back to the original source material used for creating its answers? Finally, what steps have been taken, if any, to minimize hallucination risks? Upon asking these questions, one should be able to discern whether the generative AI tool draws its answers from a broad and deep enough information database to help conduct legal tasks both accurately and efficiently.
After sufficiently gathering enough information about a generative AI tool’s underlying LLM and the database of information upon which it has been trained, one should next learn about the specific tasks that a generative AI tool offers. To illustrate the types of tasks that generative AI tools can offer to lawyers to assist them on a day-to-day basis, I will describe the available tasks contained in LexisNexis’ third iteration of Lexis+AI, Protégé. Protégé personalizes generative AI to an attorney’s workflow and style of practice. Such personalization allows the attorney to save time by receiving tailored responses to queries specific to the attorney’s role, case preferences, jurisdiction, and practice area(s).
Protégé delivers four key legal tasks to assist lawyers. The first task is Ask, which entails a multi-turn conversational search wherein the user interacts with the generative AI tool. This task simplifies the complex and time-consuming legal research journey by providing a search experience for diverse legal questions with citations and facilitates lawyers’ ability to complete research effectively and efficiently. The conversational search allows attorneys to interact with the generative AI tool like they would a trusted colleague who intelligently and conversationally responds to their requests. Prompt Assistance is also now offered on Protégé. At the end of each interaction with the generative AI tool, three tailored suggestions related to the previous query and specific to the attorney’s workflow will populate. These customized suggestions can help a user continue their conversation directly from the interface and can help combat writer’s block by suggesting where to research next.
The second task that is available on Protégé is Draft (document drafting). This task guides attorneys throughout the legal drafting process and can generate a first draft of a legal document while allowing users to subsequently change the language and tone of said draft from a single prompt. Upon entering a prompt, the generative AI tool will ask the user to provide some information such as the case fact pattern, practice area, cause of action, and jurisdiction so that a draft can be generated based upon the customized and specific input. While the generative AI tool works best with narrow, direct questions, it can encompass all the unique facts and issues for an attorney’s specific situation. The Draft task can also generate client emails, letters, legal memos, arguments, and contract clauses in an easy and quick manner that allows lawyers to start from an edit rather than from scratch. Once a draft is generated, the user can open the draft in Drafting Mode, which contains various drafting tools. These drafting tools allow attorneys to simplify language in the draft, condense and expand portions of the draft, or redraft entirely.
The third task offered on Protégé is Summarize. The Summarize task provides a custom summary of any case law document to speed up and guide insightful analysis of a case. An attorney can receive accurate and complete legal summaries without clicking into a single search result.
The fourth and final task that is available on Protégé is Documents. The task permits attorneys to upload existing document(s) to summarize and semantically search within the document(s) through the generative AI tool. By using this task, lawyers can find facts related to a given legal issue, find cases and statutes that have been mentioned in the document(s), and interpret terms of art. Users can upload up to ten documents and ask the generative AI tool to analyze the uploaded litigation or transactional documents. In doing so, attorneys will receive an analysis report that provides an overview of the key provisions, provides the risk level of provisions, and identifies issues that might be a problem. It also provides a market comparison, which informs the lawyer how it aligns with current industry standards and provides suggestions to improve the document. Users can also ask the generative AI tool to draft both transactional and discovery documents based upon the uploaded documents. For example, an attorney can upload a complaint and ask the system to draft interrogatories based upon the uploaded complaint. A lawyer similarly could upload a motion and ask the generative AI tool to draft counterarguments or a response to that motion, which would provide the legal argument needed to draft the full document.
Generative AI is a new category of technology that has the potential to transform the way that law is practiced. The way that an attorney utilizes generative AI can vastly improve one’s ability to serve clients more efficiently and effectively. It is important that lawyers select a legal generative AI tool that has been trained specifically for the legal profession, and this requires the ability to identify the key features and functionalities to look for in a legal generative AI tool. Furthermore, when selecting a legal generative AI tool, one should choose a tool that produces outputs that are grounded in authoritative legal content so that one can trust the accuracy and quality of the answers received. One should also focus on selecting a tool that has proven its speed and performance in the day-to-day workflow of practicing attorneys.
The future of legal research seems to include generative AI tools. Lawyers who adopt generative AI tools into their everyday workflow will be able to more efficiently serve their clients. Generative AI presents an invaluable tool for legal professionals moving forward, but these tools require oversight and evaluation by practicing lawyers. Despite advances in technology, attorneys themselves will undoubtedly remain the most important element of effective legal representation both now and in the future to come.
Joseph Rinaldi is a Legal Research Consultant at LexisNexis. He provides support and training to law schools in Orange County and large law firms in Orange County, Los Angeles, and San Diego. He can be reached at joseph.rinaldi@lexisnexis.com.