Last week I looked at Harry Surden’s paper on the application of “machine learning” techniques to the practice of law and the recent talk he gave at a Codex Speaker Series in Stanford. After introducing the concept of “machine learning,” Surden notes that although artificial intelligence is still unable to stand in for complex human thought processes we can still get “intelligent results without intelligence.” He also points out that, the goal here is not to replace attorneys with machines. Instead these algorithms “act as a compliment” which can help to make litigation processes and attorneys themselves more efficient.
He summed this idea up nicely in his Codex presentation:
“There are some tasks that, when they’re done by people, involve the use of higher-order skills such as abstract reasoning or problem-solving. However for some subset of these tasks, not all of them, for some of them, you can occasionally find patterns in data that can serve as proxies for some underlying cognitive judgment and can be harnessed to produce reliably useful automated results. In other words, these proxies lead to outcomes and results that approximate those that would have been done by a similarly situated person.”
So in this sense machine learning algorithms, like those found in the email spam filter example that Surden uses in his introduction, can provide useful results and produce outcomes that are similar to those arrived at using higher-level reasoning or problem-solving skills. To do this these information systems don’t, and don’t need to, artificially replicate human cognitive processes.
Surden frames this machine learning process as an “outcome-oriented view of intelligence” and sees a lot of potential for applying these types of techniques in legal practice. Especially promising when we realize that legal documents serve as the data that informs the algorithm.
“The practice of law is intertwined with the production, analysis, and organization of text documents. These include written legal opinions, discovery documents, contracts, briefs, and many other types of written legal papers.”
Surden mentions in his introduction that data is one of the key components for machine learning and legal practice offers no shortage of legal documents from which to draw data from.
One place where machine learning techniques have been successfully applied is in the discovery process. Given a known set of relevant types of documents an algorithm can be used to sort through a collection of documents and match on patterns to find other documents that might be similar and relevant to the case. Algorithms might act as filters as well excluding documents that fail to fall within the specific range of dates that an action’s “core incident” took place. Lawyers will still need to analyse the resulting smaller set of documents to confirm “relevance or privilege”, but the computer has succeeded in “shrinking the haystack” or “picking lint” as mentioned in a comment to my previous post.
However, in addition to classifying and clustering documents into piles of relevant and irrelevant documents there are “predictive coding” techniques that can be used to infer legal consequences and outcomes for clients, for example, identifying elements of risk, or degrees of liability, etc. Prediction patterns might also reveal the potential for a settlement or the amount of damages that might likely be awarded under certain parameters.
Surden refers to some work done by Daniel M. Katz and others in this area including his paper on “quantitative legal prediction” and his more recent study that analysed patterns of voting behaviour in the U.S. Supreme Court.* Surden describes the machine learning model developed as able to make “very good predictions about the outcomes of cases” based on the patterns express in previous vote count behaviours of Supreme Court justices. He also notes that Katz suggests that this “combination of human intelligence and computer-based analytics will likely prove superior to that of human analysis alone, for a variety of legal prediction tasks.”
Other examples of potential machine learning applications include: the discovery and identification of “non-obvious relationships” within large document collections extracting “subtle but useful patterns that can be employed to automate certain complex tasks”; analysing contracts for both structural aspects and potential correlations**; using automated document clustering techniques to assist in finding “prior art” in patent law cases to determine whether a patent application is new or not.
Surden does identify a number of limitations with these automated approaches, but he concludes that machine learning can be applied to “certain typical ‘easy-cases’ so that the attorney’s cognitive efforts and time can be conserved for those tasks likely to actually require higher-order legal skills.” So no suggestion that lawyers should or need to be replaced. Instead he asserts that technology can offer a variety of techniques that can enhance the skills and knowledge that lawyer’s already possess.
* See Katz, Daniel Martin. Quantitative Legal Prediction – or – How I Learned to Stop Worrying and Start Preparing for the Data Driven Future of the Legal Services Industry, 2012, and Katz, Daniel Martin, Michael James Bommarito, and Josh Blackman. Predicting the Behavior of the Supreme Court of the United States: A General Approach, 2014.
** See also Surden’s presentation on “Computable Contracts” from May of this year.