Elefant and the AI Revolution
We did not see this coming; who did?
Long ago, the tech sector was at peace. But everything changed when ChatGPT attacked!
No, seriously though, before ChatGPT, chatbots were hopeless. We were both working full time in innovation and transformation roles in 2022 when ChatGPT launched. As far as we know, no one saw it coming, despite the previous waves of AI.
Chatbots were pretty much thinly-veiled decision trees, whose branches were painstakingly crafted by hand. Alex, for example, had experimented with Google’s Dialogflow to try and create a chatbot that would help self-represented courts users to access basic legal information. The problem with such chatbots is that they require a clear starting point and a clear progression of the conversation - they don’t really work if users can come to the conversation from multiple angles or in multiple, unexpected ways.
But ChatGPT was just the beginning. It was the underlying advances in Large Language Models that really unlocked the potential in Elefant. The ability to read information and then render it as structured data is really, really helpful for law.
A lot of the things we had wanted to do for years were suddenly doable. For example, many years ago, Ian had experimented with knowledge graphs and how such technology might be useful for tracing the relationship between legal cases. We now had a way to make such a thing practical, with minimal manual intervention.
We implemented several new features that added value to the core search experience:
First, briefcases. Users can store the results of their search in a folder we call a “briefcase”. From there, users can open up a chat dialogue to query the results in the briefcase, helping to digest complex information by, for example, generating summaries.
Second, the knowledge graph. Building on Ian’s earlier idea, we now have a feature that allows users to render search results as a knowledge graph, rather than just a list of entries. This allows users to visualise the relationships between entities, very useful for things like cases, which often refer to other cases. Using this technology, it becomes easy to see which cases are more central, and therefore more important, to a given body of law.
Third, research memos. Users can simply tell Elegant to go and research a specific legal topic. Elefant then returns a research memo summarising its findings and including links to primary sources.
Fourth, the reading list downloader. No longer do students and lawyers have to painstakingly search for and then download cases individually. Elefant users can simply upload their reading lists to Elefant, which will recognise case names and then download the entire list of cases, packaged conveniently in a single zip folder.
We’re even working on going back to our roots, to develop a sentencing tool that can read cases, extract data from them, and structure that data. Users can then discern sentencing trends and produce useful aids like graphs, tables or charts.
“But wait!” you might say, “What about hallucinations?”
Actually, we’re well aware of the limits of generative AI. In fact, for some time, we resisted branding Elefant as an AI company - we were very focused on improving the search experience.
We’ve now reached a stage, however, where we’re confident we’ve found ways to handle the hallucination problem. In fact, we think we do it better than others:
First, we focus on primary sources like cases, statutes and parliamentary records. We don’t generate answers for you based on a training dataset that encompasses the entire internet. We search first, narrow down to relevant primary sources - and this narrowing down takes a lot more domain expertise than one would think - and only then turn the generative AI loose to do its thing. The focus on primary sources at the start, rather than at a later stage like most Retrieval Augment Generation solutions, helps to significantly reduce the risk of hallucinations.
Second, we validate the results of anything generated by AI using a form of Logic Augmented Generation. We’re working with our friends at Legalese to develop computational law technologies that map the logical relationships between legal provisions. In other words, we’re also developing a deductive logic framework to complement the inductive logic work done by generative AI. The two approaches work hand-in-hand to check each other’s results, thereby further enhancing accuracy.
Third, we’re constantly tweaking and repurposing our existing features to enhance the overall Elefant experience. For example, knowledge graphs can also be used to enhance accuracy of AI-generated answers. Further, we’re tweaking the reading list downloader so it can be used as a final layer of verification - anything Elefant cites, it should also be able to find and download for you so that you can check that source.
We’re excited about AI and the possibilities it opens up, but we’ve never forgotten our beginnings as lawyers and our commitment to the truth.
 |
|