Posted on April 25, 2017
A Lawyer’s guide to Artificial Intelligence; or Why I Joined an AI Startup?
Although us lawyers sometimes hate to admit it in front of our non-lawyer friends, we all know that the glamorous image of the profession is often a gross misrepresentation of reality.
It’s not that the litigation shark or the savvy corporate lawyer depicted in television shows and movies do not exist. They do. But the movies often rush to the exciting cross-examination scene in which the lawyer brilliantly manages to protect its innocent client, while skipping over the long hours spent tediously reading thousands of pages and analyzing complex contracts. The truth is that a real lawyer’s life wouldn’t make much of an interesting movie at all.
The LawGeex Revolution
I think that is a big part of why I was so excited when I first met with LawGeex CEO and co-founder, Noory Bechor, and heard about some of the amazing work done here. I had just recently returned to Israel after earning an LL.M. degree at New York University and practicing in New York City when I met with Noory for the first time in late 2016. I immediately recognized that what LawGeex is offering is nothing short of a revolution in the way law is practiced. Since then I have been leading a team of senior attorneys and data scientists in tackling some of the most difficult challenges our company is facing.
Introduction to AI in the Law
Before touching on what exactly these challenges are, I should probably introduce what LawGeex is doing and why it is already today one of the most exciting and cutting edge companies.
LawGeex uses cutting-edge Artificial Intelligence (AI) technology, in reviewing and automatically approving contracts, streamlining workflow, saving valuable lawyer-hours and avoiding unnecessary mistakes. Our AI is a human-like legal issues spotter providing relevant information on contract terms, therefore allowing lawyers to focus their review on the relevant segments of each contract, saving countless lawyer-hours. It can allow for automatic approval of a submitted contract, or where this is not possible, a detailed report on missing, problematic or rare clauses, within One Hour.
By employing sophisticated Machine Learning and Deep Learning algorithms (developed by LawGeex research team, led by CTO and co-founder, Ilan Admon) we are making legal practice for in-house counsel more enjoyable, accessible, and efficient.
What is AI anyway?
AI algorithms were first theorized and developed all the way back in the 1950s, but it was only the computing power of the last decade that allowed us to widely implement them. These algorithms basically try to imitate the human brain so that rather than requiring a fixed set of convoluted rules to determine a result, they output a prediction based on examples. In other words, these are computers that learn by example, much like humans do.
Let’s bring on the Pandas and Koalas
LawGeex’s AI engine is programmed to solve what is called a classification problem, which basically means calculating the probability that certain information is either of type A or type B. Let’s take an overly simplified example in which a computer is taught to distinguish between pandas and koalas.
As mentioned above, the algorithm learns by example, so we would first need to present the computer with samples of pandas and koalas. Collectively, these samples are called a training set.
Once enough samples have been collected, we can ask the algorithm to build a model, which can be understood as outlining a boundary between the koalas and the pandas. This is a resource intensive iterative process called training. Once the training is complete, we would show the computer new samples that were not included in the training set and test how well our algorithm learned to distinguish between pandas and koalas. A successful model would predict that the red question mark located left of the blue boundary line is a panda, and the one on the right is a koala. While the panda-koala example may look pretty straightforward and simple, our actual training set is far more complicated, as you can see in the terrifying-looking image below.
You may think that this jungle of colorful dots is light-years away from your legal practice, but this is actually a visualization of how our AI tools perceive and analyze thousands of legal clauses included in thousands of contracts (our AI solution understands and interprets an exclusive corpus of 1000s of contracts). Our team inputs real contracts into our AI machines, which then “read” the contracts.
Since computers can’t actually read in the same sense as humans do, they basically convert the text into a numeric representation. Each dot in the image represents a specific segment of a contract included in the training set. The different colors represent different legal issues. The pink dots, for example, represent samples of non-compete clauses, and the purple ones represent governing law sections. Our brilliant algorithms allow us to make sense of this dot-jungle and solve the classification problem, but of course, these algorithms rely on good data and a good training set. This is where my team comes in.
Creating a good training set
Collecting training sets is a slow and expensive process, which even the largest technology companies struggle with. The problem is not only collecting the examples but mostly knowing if and when to stop. How do we know if enough samples were collected? Is our training set complete in the sense that it captures the entire scope of a legal concept? How do we find which samples are missing from the training set? These are all challenges we in the data team deal with on a daily basis, and solving them is crucial for building a really good AI.
Train AI Like a Lawyer
It is not enough to collect and label data if that data is just more of the same of what we already have. We strive to present our algorithms with a variety of examples that will allow it to distinguish between different legal concepts. Training an AI machine is very similar in this sense to training a new lawyer – exposure to different types of examples is crucial in developing a good understanding of the legal practice.
Hand labeling data is expensive and time-consuming, and we want our expert lawyers’ time to be well-spent labeling only quality samples that will make our AI smarter. In order to achieve that, we use a set of sophisticated processes for predicting which data (out of the enormous amount of text out there), might qualify to be included in our training set. We involve our expert lawyers in strategic junctions throughout the process to make sure we are being accurate.
Using AI to give Lawyers Back their time
Our Machine Learning and Deep Learning technology is what gives us our competitive edge and sets us apart from other companies in the legal-tech world. Learning by example, rather than a fixed set of rules, allows us to analyze complex information with human-like precision and help our clients move swiftly from negotiation to signing. This allows lawyers to get their time back. With the routine drudge work gone, lawyers can become more like the glamorous lawyers on screen (such as Harvey Specter in Suits).
This is why we are committed to investing so much in good data, mixing the best of expert lawyers and the latest technology, consistently making our AI smarter, more accurate and more human-like every day. For me personally, it is making me love law again.
Gil Rosenblum, Esq. is Data Team Leader at LawGeex. He trained as a lawyer in New York and currently lives in Tel Aviv. For more details or a demo get in touch.