I would have loved to have been in the audience when Harry Surden spoke at an afternoon CodeX Speaker Series event a few weeks ago at Stanford Law School. But, you know, California is way over there and I’m way over here. So, although I could not actually attend I was prompted to go back and read his recent paper, “Machine Learning and Law.” And, thanks to the folks at Codex, it turns out the session I missed was recorded and is also now available for viewing.
Surden has an interesting mix of experience. He has a background in software engineering and worked in that capacity at Cisco and Bloomberg before he entered the field of law. He’s currently Associate Professor of Law at the University of Colorado Law School where he teaches intellectual property law, information privacy law and brings his software engineering experience to bear on his thinking about the application of computer technology and legal practice.
This relatively unique set of experiences informs one of his central research questions: “What impact might artificial intelligence (AI) have upon the practice of law?”
When you hear that term “artificial intelligence” you might, like me, think back to the work begun in the ’60s where attempts were made to create the equivalent of a thinking human brain. The field of artificial intelligence has evolved since then to include a “softer” outcomes based approach focusing on computer algorithms capable of performing specific tasks that improve over time.
This process has been called “machine learning” and Surden gives a nice introduction to this idea in his Codex talk and provides an excellent overview in the first part of his article. Generally speaking he aims to start readers off with “some basic principles of machine learning in a manner accessible to non-technical audiences in order to express a larger point about the potential applicability of these techniques to tasks within the law.”
To illustrate this “potential applicability” he begins with a practical example familiar to most of us: the email spam filter. The algorithm used in a typical spam filter is designed to detect patterns found in a collection of data (i.e. the incoming emails). The identification of these patterns allows the program to “decide” whether an email is wanted or unwanted. So for example, the phrase, “Earn Extra Cash!,” is more likely to be found in an unwanted spam message than it would be in an authentic and wanted email message. Over time the program builds on its experience with the email messages it collects and improves its ability to identify and remove spam from our email traffic.
Because the filtering process improves over time the term “learning” has been applied to this process. But Surden cautions that this is not learning in the human, cognitive sense. There is no “intelligence” involved here. Rather this is learning in a metaphorical sense based on the functions performed by the computer and the outcomes that are produced.
This illusion of intelligence succeeds because “machine learning often (but not exclusively) involves learning from a set of verified examples of some phenomenon.” For an email spam filter that might mean drawing from a dataset that contains unwanted emails first identified by human readers. That dataset then initially serves to inform the algorithm giving it something to learn from.
Surden describes three main features of machine learning:
- data is required (e.g. the incoming emails)
- an algorithm with sophisticated pattern detection allowing predictions that appear similar to decisions a person would make given the same information, and
- a system that can realign itself, modify its behaviour and adjust to changes found in the environment it works in
We know that the practice of law will usually require advanced cognitive abilities like reasoning and judgment based on professional intuition and experience. However, as Surden notes, we also know “that even the most advanced artificial intelligence systems can’t match the analytical skills of even very young children in certain circumstances let alone trained attorneys.”
So given these inherent limitations in computer processing what types of legal tasks would lend themselves to automation through machine learning? I’ll continue with Surden‘s suggestions in my next post.