Artificial Intelligence (AI) technology is gradually and surreptitiously permeating diverse aspects of human life. AI is now being utilized in performance of tasks once considered to be within the exclusive domain of human intelligence. From high-tech driverless cars and trucks to basic facial recognition technology in your Facebook profile. The tentacles of AI are spreading like an octopus and even the practice of law has not been out of reach. With this increasing spread, lawyers can only hope that we do not get to that point where the work of lawyers is completely and efficiently replaced using AI technology.
Meaning and scope of AI
AI is probably one concept that has defied a universally accepted definition even among experts in the field of computing. As noted by Scherer, this difficulty lies with the rather “conceptual ambiguity of intelligence” which is often associated with human intelligence. But John McCarthy a pioneer in the field of AI countered by stating that “AI does not have to confine itself to methods that are biologically observable.” He went on to define AI as “the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence”.
At the early stage of AI, ability to perform intellectual tasks seem to be the focal point of definitional approach at AI. But, Scherer has rightly observed that the “concepts of what constitutes artificial intelligence have shifted over time as technological advances allow computers to perform tasks that previously were thought to be indelible hallmarks of intelligence.” He also went on to note that “[t]oday, it appears that the most widely-used current approaches to defining AI focus on the concept of machines that work to achieve goals.” Thus, the concept of intelligence has gradually moved from the sole emphasis on human cognitive ability to incorporate the rationale ability to achieve defined goals. According to Scherer, AI “refers to machines that are capable of performing tasks that, if performed by a human, would be said to require intelligence.” Scherer’s definition though could be extended to include not just machines (as in hardware) but also software that has the capacity to perform ‘tasks that, if performed by a human, would be said to require intelligence.’
Machine Learning is one of the most important branches of AI. Machine learning is a branch of AI in which computers ‘learn’ to perform some tasks and improve in the performance of the task over time through a training using ‘seed sets’. Thus, machine learning enable computers to perform tasks for which they are not explicitly programed by developing intelligence from data analysis. This process makes it possible for researchers to design computer programs to perform tasks that were once considered to be only capable of performance using human cognitive intelligence.
Machine learning has been associated with the ability of computers to ‘learn’ from experience, and subsequently improve their performance as a result of executing same task over a period of time. Surden noted that the use of ‘learning’ with reference to machine learning does not imply that computers are capable of possessing human cognitive abilities. Rather the concept of ‘learning’ in machine learning is in a functional sense – machines develop the capacity to change and better perform a given task as a result of experience acquired in the performance of similar or related tasks. In highlighting the ‘intelligence’ associated with machine learning, Surden stated that “[i]f performing well, machine learning algorithms may produce automated results that approximate those that would have been made by a similarly situated person.” Machine learning algorithm has been employed in many modern technologies such as speech recognition, facial recognition, auto-corrects, spam filters in emails etc.
The spam email analogy
To a lay person, perhaps the best example that has been used to illustrate the basic features of machine learning is the email analogy used by Surden in his article Machine Learning and Law. Spam emails usually constitutes a nuisance to the recipient. Hence email service providers provide their users with the option to flag an email as spam to enable the service to track similar emails and send them directly to the spam email folder rather than the user’s inbox. This process entails the use of machine learning algorithms, and the training of such algorithms to detect the unique characteristics of spam emails by feeding the algorithms with examples (seed set) of junk emails.
For example, if a user receives a junk email from an online pharmacy requesting the user to place an order for ‘Viagra or Cialis’. The user could simply flag the email as junk by sending it to the spam mail folder. This process of flagging the email and sending it to the spam folder serves to “train” the machine learning algorithm by providing it with seed set of spam email to analyze. In doing this, the machine learning algorithm will identify certain characteristics in the email from which it will learn to identify subsequent emails possessing those characteristics as spam emails. Such characteristics could include the domain name from which the email originated, words or phrases in the body of the email such as “Cialis” or “Viagra” or “Online pharmacy”. The algorithm can use these and other characteristics to determine whether an incoming email is spam or not.
According to Surden “machine learning algorithms are able to automatically build such heuristics by inferring information through pattern detection in data. If these heuristics are correct, they will allow the algorithm to make predictions or automated decisions involving future data.” The ability of the machine learning algorithm in this case to identify spam emails from other sources (aside the online pharmacy) will improve with spam emails from other sources being flagged and sent to the spam email folder for analysis.
Surden also noted that apart from learning characteristics that will enable it to identify an email as spam, the algorithm may also learn other characteristics that will enable it to identify an email as not being spam. Thus, it could learn that emails from individuals that the user had previously communicated with are not spam even if emails from such individuals contain phrases like “Viagra” or “Cialis”. Hence, the rate of accuracy in identification of spam email by the machine learning algorithm improves with more ‘data set’ being feed to and analyzed by the algorithm. According to Surden:
This capability to improve in performance over time by continually analyzing data to detect additional useful patterns is the key attribute that characterizes machine learning algorithms. Upon the basis of such an incrementally produced model, a well-performing machine learning algorithm may be able to automatically perform a task—such as classifying incoming emails as either spam or wanted emails—with a high degree of accuracy that approximates the classifications that a similarly situated human reviewer would have made.
This is not to imply though that machine learning algorithm possess perfect accuracy, it is possible to have cases of false positives and false negatives. Examples are situations where spam emails fail to be identified and flagged as such, and situations where emails that are not spam are wrongly flagged as spam and sent to the spam folder. What is evident though is that with adequate training, machine learning algorithms can achieve accurate results that meet or exceed the rate achieved by humans in much shorter time.
Predictive coding is a machine learning process that relies on analysis of a sample data set to make a determination or classification of a larger dataset. Hampton noted that the process involves ‘machine learning and a combination of different algorithmic tools.’ Common tools employed in predictive coding include metadata searching, contextual searching, and concept searching.
Unlike keyword search which focuses on specific search term irrespective of the context. Concept searching relies on the context in which the specific word is used. Thus, predictive coding is now increasingly used in review of large document sets – sorting same into predetermined categories. Yablon and Landsman-Roos has described predictive coding as “a process whereby computers are programmed to search large quantities of documents using complex algorithms to mimic the document selection process of knowledgeable, human document review.”
At its early stage of use in document review, predictive coding attracted little interest until 2010 when the results from two pilot research projects which compared predictive coding to manual document review were published. Both studies concluded that the use of predictive coding in document review achieved a higher level of result than manual review. Since then, support for use of predictive coding in document review has continue to grow. Today, the process has come to be variously referred to as technology-assisted review (TAR), or computer-assisted review (CAR).
The next edition of this blog will examine how predictive coding, an aspect of AI could be applied in civil litigation especially in the area of document review.