While people can interact with GPT-3.5 and GPT-4 directly, what will eventually be more common is interaction with GPT teamed up with other software. Lexis, Westlaw, and Bloomberg are all developing generative AI tools incorporating GPT while also adding layers both before and after GPT to adjust for accuracy and verification. These legal research generative AI products are designed for accuracy versus ChatGPT which is designed for language fluency. With these added layers, prompt engineering is less important, but knowing generally how AI works, as well as its strengths and limitations, will still benefit the user.
Potential use cases (some of these perform better than others but since these products are still very young it's hard to say which):
More resources
Mohar Chatterjee, How lawyers use AI, Politico (May 2023)
Legal research vendors have worked aggressively to build products that limit hallucinations and increase accuracy. First, most have developed specialized models trained on narrower, domain-specific datasets. The idea is that "good data," and only good data, is allowed in the system. Secondly, most vendors are using retrieval-augmented generation (RAG) which takes the user’s question and passes it through a database that then adds to the user's question as “context” that is then sent through the model. Third, some products may also use vector embedding as a way to identify concepts, by way of assigning phrases or even entire documents, as numerical vectors. Coupled with RAG, this increases precision and relevancy. Lastly but certainly not least, almost all vendors incorporate human feedback on responses.
It is true that these methods have reduced the risk of fake cases as seen in the Avianca case. However these systems may still hallucinate with false sources or produce flawed outputs due to human user error, prompt misinterpretation, or missed issues.
A 2024 benchmarking study indicated that legal generative AI models do reduce errors compared to general-purpose AI models like GPT-4, but that these models still hallucinate at an alarming rate. In fact, the Lexis+ AI system produced incorrect information more than 17% of the time, while Westlaw’s AI-Assisted Research and Ask Practical Law AI hallucinated more than 33% of the time.
Access: Lexis+AI is currently live and available to all U.S. firms and corporate partners. In the academic space it is available at a small group of law schools with legal technology programs. USD faculty, administrators, 1L, 2L, and 3L students all currently have access to Lexis+AI. Lexis consultants will be offer on campus training for faculty and students. Students should expect to see it in their summer associate positions though use terms and guidance will vary depending on the firm/institutions.
Main features for the academic packages include:
Document modifying (e.g. summarize this, make this more aggressive, make it shorter) and document upload are available in some law firm packages. Pricing varies.
Lexis+AI utilizes three LLMs (OpenAi, Anthropic (AWS), and Google Bert). It is trained on both primary law (statutes, cases, constitutions, state and federal rules, and select administrative agency decisions and regulations) as well as a limited amount of secondary sources owned by Lexis (Matthew Bender, certain treatises, and Practical Guidance).
Lexis has stated that Lexis+AI will not generate false sources. Lexis+AI follows the RELX Responsible AI Principles to ensure accountability and prevent the creation or reinforcement of unfair bias. Both Lexis and Westlaw have stated that confidentiality is paramount and no user data will not be included in training data.
Lexis' Generative Artificial Intelligence (AI) Resource Kit an overview of current practical guidance on generative artificial intelligence organized by practice area: IP, Data Security & Privacy, Civil Litigation, Commercial Transactions, Labor & Employment, Healthcare, Employee Benefits & Executive Compensation, Tax, and Federal Government.
Become a Lexis+AI insider to stay abreast of current developments.
Products and Access: As of February 2024, Westlaw is offering three separate products:
While Ask Practical Law AI has a fairly narrow focus, we expect Westlaw's AI assisted research and CoCounsel Core to be more similar to the robust Lexis+AI product. CoCounsel Core comes from Casetext's generative AI product, CoCounsel. In June 2023, Thomson Reuters announced their intention to acquire Casetext for $635 million, finalizing the merger in August 2023.
Main features:
Both Lexis and Westlaw have stated that confidentiality is paramount and no user data will not be included in training data.
Stay up to date with AI @ Thomson Reuters.
Bloomberg Law Answers, announced in July 2024, enhances the research experience by generating brief answers to searches using gen AI technology and LLMs in combination with primary and select secondary sources.
Legal research behemoths Lexis, Westlaw, and Bloomberg are in an arms race to bring generative AI tools to their platforms. For each of these companies the primary concern is "good data in, good data out."
More resources on how vendors are fine-tuning their generative AI models:
The PDF below shows all the legal tech companies currently working on generative AI tools as of June 2024.
