Legal AI: A Lawyer’s Friend or Foe?

‘The robots are taking our jobs!’ An overused quip thrown out whenever Legal Artificial Intelligence and automation are mentioned. We laugh it off, but an uneasiness lingers because, undoubtedly, there is more than a hint of truth to it. As the incoming crop of clerks and graduates, the sentiment is particularly concerning. The confronting articles claiming that ‘X’ percent of law jobs will soon be automated doesn’t help either.

It may be cliché coming from another technology advocate, but nevertheless I implore you to reframe it as potential for an array of new opportunities. The most impactful of those opportunities offered is the ability to drastically improve access to justice.

The Access to Justice Arrangements inquiry conducted by the Australian Productivity Commission revealed that disadvantaged Australians are more susceptible to, and less equipped to deal with, legal disputes. The reasons are well known. Excessive costs and increasing complexity preclude those who need them the most from accessing legal services.

The two types of legal automation

Automation of repetitive tasks has been progressing rapidly for the better part of the last two decades. When it comes to the legal profession however, the adoption of AI technology is still in its relative infancy.

Legal technology as it exists now can broadly be divided into two categories: technology that services a lawyer’s needs, and technology that services a consumer’s needs.

The first category is likely the most familiar and involves tools that help a lawyer be more efficient at doing what they do. Contract management, matter management, e-Discovery and other lawyer-centred products are steadily becoming staples of legal practice.

The second category is much more fascinating and has the potential to resolve some of the main issues associated with legal accessibility. The goal of this technology is to perform the role of a lawyer, providing legal services to customers more efficiently and at a much lower cost. If the cost of accessing legal services exceeds any potential benefit obtained for consumers, then there is no other option than to accept the circumstances as they stand. I believe development of legal technology would allow us to overcome this deficiency in justice.

The type of work that lawyers do is well suited for a technological system. Logical problem solving, based on sets of rules and input information, is fundamentally what computers are used for.

Traditional programming involves a manual process whereby a programmer encodes the rules and parameters within which the computer should interpret the input information to produce a result. The digitisation of legal documents is an example of how traditional programming has existed in the legal practice. Searchable databases have been great contributors to improving accessibility to legal texts. While these will still remain fundamental tools, machine learning offers a massive leap beyond what we have been able to achieve thus far. Instead of pre-programing a set of rules, a machine learning model generates them by detecting a pattern between the input and output data fed to it. These models can theoretically analyse cases, using the facts, arguments and judgments to recognise a pattern in order to predict outcomes of future cases.

Legal A.I in practice

Why is it then, despite these grand opportunities, does this form of legal AI still (for the most part) exist only in the realm of theory?

The simple answer is that our technology just isn’t there yet. The primary barrier is the machine’s inability to understand and process legal data. To better understand the challenge, some context on how machine learning models currently function in the legal practice is needed.

The typical machine learning model used in legal practice is composed of two key components: the model’s algorithm, and a collection of pre-selected and pre-processed legal documents from which the model can “train” itself. The collection of documents, or corpus, is labelled; each document in the corpus is given a desired label that the lawyer wants it to be associated with. For instance, a lawyer will label a given application to the court as being either successful or unsuccessful. Each time a label is given to a document, the lawyer is providing a training example for the algorithm. The algorithm processes the data within the documents to recognise patterns according to their labels. In this training phase, the algorithm learns to classify the documents with their respective labels through (effectively) trial and error. Once trained, provided the algorithm is well-designed and the training data is sufficient, the final machine learning model can extrapolate what it has learnt to predict the likelihood of success for future applications. An everyday example of this process is the spam filter that most email software use. Each time an email is marked as spam, the machine learning system recognises patterns in the contents of those emails to pre-emptively divert similar incoming emails.

However, despite this progress in data processing and generation, machine learning in legal practice is still a way off from being at the level where they can adequately perform the role of lawyers. In their current form, machine learning models are not cognitively sophisticated enough to approach legal problems with the same analytical rigour a human lawyer would be capable of. A human can analyse and understand the nuances of a legal document through comprehension of that single document and – by virtue of years of practice and study – reflect on their own knowledge. A machine learning model on the other hand, can only approach legal problems through detecting recurring patterns found in a large corpus of similar documents.

What is Natural Language Processing (NLP) in Legal A.I?

This fundamental difference in approaches stems from the system’s inability to comprehend a text. Unlike a human, a machine learning model does not “read” and understand a text, instead it uses pre-processing techniques to parse and extract only the most essential features of the information. These techniques are known as Natural Language Processing (NLP).

NLP takes human phrasing, or “natural language”, and converts it into information that the computer can recognise and process.

While NLP techniques have seen significant leaps in progress, they are still limited by the computer’s fundamental inability to understand nuance and context. This is the major challenge for legal AI as legal language is heavily contextual and filled with nuance. If you have ever wondered why translating a phrase through Google Translate can sometimes produce a humorous result, that is the limitation of NLP. The computer isn’t able to understand the nuances of the phrase, it can only translate an approximation of the same information. However, there are ways we can supplement the data we input to help machine learning models better understand context. One of these tools are ontologies.

Ontologies are a way to represent knowledge in a machine-readable format. They are an explicit, structured description of concepts and relations between them in a domain. Legal concepts are known as being open textured concepts, meaning that their definition may vary depending on many factors. We, as humans, can use context to understand these concepts and their place in the broader text. Ontologies can help bridge the contextual gap for a machine learning model. Take for example the concept of a ‘bottle’. It has both a real world meaning and a legal meaning. In the real world, a bottle exists as an object, a verb and can also mean an action associated with the object. As a legal concept, a bottle can be an element to aggravated assault or a threat. By creating an ontology that maps the relationships between ‘bottle’ and other related terms, the machine can more accurately detect how it should be interpreted in each situation. In theory, if the machine learning model detects a ‘bottle’ in the presence of the elements of assault, the model will recognise to process said bottle as a weapon and search for other related terms.

However, development of tools like ontologies present their own challenges. Mapping the relationships between all legal concepts across multiple jurisdictions is, as you can imagine, a massive undertaking. It requires not just the efforts of software developers and engineers but also legal practitioners who understand these relationships best. Development of legal AI tools that improve access to justice is now more a matter of time than possibility. Of course, the challenges discussed here are only a very broad overview of a segment, of the whole picture. But if we as the new generation of lawyers can drive interest and investment in this area, we can improve access to legal services for more people even sooner.

Interested in learning more about legal A.I? Read the other articles in our series including Steak, muffins and chihuahuas: The unkept promise of ‘game-changing’ Legal A.I. and Legal A.I: High on artificial, low on intelligence.