Breaking Down Legal Artificial Intelligence

Hi and welcome to the second installment in the ultra series, where we’ll explore what we mean by AI. AI is a term with no definite meaning. The term artificial intelligence is applied when a machine mimics cognitive functions that we associate with human minds. Just learning and problem solving. » The scope of AI is uncertain at the best of times. And to make things even more confusing, it’s a moving target. As machines become increasingly capable, tasks considered as requiring intelligence are often removed from the definition. For instance, optical character recognition is now frequently excluded from artificial intelligence having become a routine technology. » It is much more helpful to break it down into different categories.

Different software may use multiple categories, but it helps to understand how the systems work. We will only discuss things relevant to legal AI. For example, robotics is usually considered a part of AI. But it is not yet relevant to lawyers. Here’s another quote for you to kick things off. It’s only AI when you don’t know how it works. Once you know how it works, it’s just software, Kurt Keutzer. » For our first dive into AI, let’s talk about machine learning. It broadly breaks down into two main parts, supervised learning and unsupervised learning. Supervised learning at its most basic is where a machine has an input to an output. It’s usually a number or a particular tack.

Let’s say you wanted to categorize millions of pictures of pets into subcategories, cats and dogs. You could take the first few hundred pictures and tell a machine, this is a picture of a cat. This is also a cat. This is a dog. The machine would then use that data to categorize the remaining few million pictures. In a legal context this may be used to identify a particular clause such as the jurisdiction clause or the amount of salary in an employment contract. » Unsupervised learning is where the machine clusters groups based on the level of similarity. So you may give it a particular clause or a document and the machine looks for similar items.

The machine will find clauses or documents that have similar wording to the original. And you can set the parameters so you get what you’re looking for. The machine is the one using various algorithms to find the groups rather than relying on human reinforcement. » Semi-supervised or reinforcement learning uses unsupervised learning technology with human reinforcement at stages along the way. Machine tries to group things and then the human says, you got this bit right you got the other bit wrong, learn from it and adjust. With most of these systems, there’s often a trade-off between getting everything you’re looking for and getting only the things that you’re looking for. Imagine a barrel full of marbles, all sorts of marbles, all different colors.

I ask a machine to find orange marbles. Now the machine has a fairly good idea what I mean by orange, but the problem comes around the edges. What about tangerine? Ginger, how about scarlet colored marbles? If the machine gives me everything that’s vaguely orange, I may have too many marbles to look through. If it’s more selective then it may miss out the marble that I need. To put it all in a legal context, and let’s be extreme about it. If I were looking for a particular clause in a contract, if the search returns every clause in the document, it will definitely have the one I need, but also lots that’s irrelevant.

Even though has brought me the right clause, the volume of irrelevant information makes it useless. On the flip side, if it returns no clauses, it will definitely not have any irrelevant clauses, but it will also be missing the right one. There’s often a balance between these two competing objectives. Including the right thing and not including anything irrelevant with a perfect balance being difficult to achieve.