Ever since OpenAI released its generative AI application ChatGPT, there has been a lot of talk about the technology’s impact on the legal community. And speculation only increased when GPT-4, the world’s most advanced large-scale language model (LLM) (powering subscription service ChatGPT Plus), was announced in his March.
As explained in the first post of this series, it is premature to rely on ChatGPT or GPT-4 in legal practice because lawyers cannot hallucinate or access up-to-date and accurate legal data. And many people think. Own. OpenAI itself warns users that he will rely on GPT-4 output, especially if the risk is high.
But that doesn’t mean lawyers can’t rely on generative AI in their legal practice.It’s true they can’t trust generative AI 1 person— It’s a crucial difference.it is possible to build a product Integrate GPT-4 that meets professional standards. This is exactly what we did with CoCounsel. But how?
Pour the power of GPT-4 into a trusted legal AI platform
It’s only been the last few months that LLM has been in the daily news, but it’s been around for years and our engineers have been working with LLM since 2018 on products like Parallel Search. has been developed. However, the superiority of GPT-4’s inference capabilities marked a turning point in its release. Never before has a model performed legal reasoning so well. So why isn’t GPT-4 enough? The main problem is memory.
GPT-4 may give inaccurate answers to questions or falsify information (hallucinations). This is because GPT-4’s only source of information is his own memories, and that information is limited to publicly available information until September 2021. And that public data contains a lot of unreliable information. “Falsehood, hate speech and other garbage.”
However, GPT-4 is part of an ecosystem of “brain power” that consumes, analyzes, and synthesizes information, not just public information, but also memory, including domain-specific databases. , it is possible to produce reliable output.
That’s why OpenAI chose Casetext to use GPT-4 when building a product fit for legal professionals. As a leader in legal AI innovation since 2013, we have a good “memory” to anchor GPT-4 inferences, and the tools to capture the right parts of that memory. What are the results of this integration? Co-advisors.
Law-based AI
In building CoCounsel, Casetext’s product and engineering teams integrated that information with GPT-4. won’t Our legal database is a comprehensive corpus of accurate and up-to-date legislation including state and federal case law, statutes, regulations, codes and rules.
This means that all CoCounsel output comes from a thorough compilation of legal information. Our engineers “tell” the platform to answer based on the actual sentences contained in the database or not answer at all, leaving no chance for CoCounsel to hallucinate.
In addition to the ‘brain’ (GPT-4) and ‘memory’ (database), the third element of the CoCounsel ecosystem is the ‘appendages’, the proprietary tools Parallel Search and AllSearch. They guide the GPT-4 to retrieve the appropriate data from memory to answer the user’s legal questions and provide quick answers.
4,000+ hours of expert rapid engineering
Legislating GPT-4 is just the first step. Next came Prompt Engineering. This is another term like generative AI, which is now commonly used in mainstream media as well. An LLM “prompt” is essentially a question or query that is asked. Effective prompts prevent hallucinations and ensure accurate and complete answers. To create these, “we need to provide clear and unambiguous language, context and background information, deconstruct complex questions, experiment with different phrasing, and monitor the accuracy and bias of the generated content.” there is.”
Rapid Engineering enlists the Trust Team, a group of expert AI engineers and experienced litigators and transaction attorneys, to come up with the “clear and sufficient context” essential for the model to generate useful answers. It started with the establishment They selected and designed thousands of prompts and entered them into CoCounsel.
The team reviewed the prompt output, made minor changes to the prompt content and wording to improve the quality and accuracy of the output, and then retyped it. After doing this several times per prompt, she filtered and ranked the responses, chose the best, and fed that information back to her CoCounsel. This oversight and refinement is essential to maximizing the value of generative AI. Every aspect of this feedback improves results.
After over 4,000 hours of work based on nearly 30,000 legal questions entered into CoCounsel from October 2022 to March 2023, our team is confident that for the first time our products are safe for professional use and Decided it was ready for launch.
Continuous improvement and expansion of CoCounsel skills
Since launching CoCounsel, we’ve continued to test its seven skills daily, and launched our eighth skill by entering and checking thousands of queries. We also have a backend alerting process built into the product. During this process, CoCounsel screens for potential inaccuracies and flags them for review, preventing inaccuracies from surfacing to end users.
Perhaps most importantly, it is logged and read by Customer Success, Product, and Engineering teams. Every single comment or suggestion Information provided by CoCounsel users. We use this information to make improvements, develop additional skills, and determine which changes and additions to prioritize based on what our customers want and need most.
The next post in this series will discuss the importance of customer and client data privacy and security in AI solutions suitable for professional legal use.
