Lawyers’ Ethics and the Use of Artificial Intelligence in Legal Services

By Michael Legg, Professor, UNSW Faculty of Law & Justice, Sydney

New technology, such as artificial intelligence (AI), is driving change in society and business. AI’s advancement also impacts on legal practice, not just through efficiencies and profitability, but also through the ethical duties of lawyers. Ethical responsibilities both limit and require the use of AI. Lawyers need to be able to use AI tools to be able to discharge their duties of competence and acting in the best interest of clients. AI tools may also facilitate access to justice. AI also creates risks when it comes to maintaining confidentiality, legal professional privilege, and independence.

What is Artificial Intelligence?

AI, as a term or field of computer science, is employed where processes are used to carry out tasks which, if performed by a human, would be seen as evidence of intelligence ‒ i.e. the processes mimic, imitate or simulate intelligence. AI is also an umbrella term. There are different branches of AI as shown by figure 1. This article focuses on the branches of AI relevant to legal practice, namely expert systems, machine learning and Natural Language Processing (NLP).1

In brief, expert systems, the oldest and most simple form of AI, are pre-programed systems which can guide users through a sequence or series of steps, similar to a decision tree. The system involves obtaining and deconstructing human expert knowledge into a computable form that can then be accessed more cheaply and widely. Machine learning refers to data-driven programs which use pattern recognition in data and statistics to produce their outputs. There are three types of machine learning. In supervised learning, the data is already labelled (for example, a picture is labelled as a dog or a cat), and the program is trained on that data to identify associations between the data and the labelled outcome, or classification. The program can then classify new data. Unsupervised learning involves no labelling, instead the software