Legal AI beyond the hype: understanding what AI in legal actually means

There is a lot of hype about artificial intelligence (AI) in legal as a way to reduce costs, expand access to justice, review documents faster, and replace all humans with their robotic doppelgänger. This post is my attempt to help define and demystify Legal AI so that it’s something specific we can talk about, not just hand-wavy magic.

What Are the Elements of AI in Legal?

Legal AI is a convenient shorthand for the combination of multiple machine learning algorithms that are often used together with natural language processing to make sense of legal documents and text. AI in legal is not really AI, but “AI” is easier to say than all of that.

The term AI is somewhat misleading. AI is artificial intelligence. It’s the idea of giving a computer the ability to think, reason, and learn about the world around us. It’s replicating the capability of the biology of our brains onto silicon chips. The goal of truly advanced artificial intelligence efforts is to create a general algorithm that can learn to do basically anything. None of the legal technology companies have that type of AI. If they did, their robot lawyers would quickly get bored reviewing legal documents and sign up for user accounts on Pinterest and Snapchat.

Now that we agree that Legal AI is used as an umbrella term encompassing several components, let’s demystify each of them.

Natural Language Processing

In the legal industry, natural language processing technology can speed up tasks such as legal research or contract reviews.

Natural language processing, or NLP, is how software makes sense of written words. While you and I learned language by example, computers take a more algorithmic approach.

Take this sentence, for instance, “On safari, I took a picture of a giraffe in my pajamas.” Although the word order and structure is a little ambiguous, people can parse out the intended meaning. A human will correctly interpret that the speaker is the one on safari taking pictures and will also understand that the speaker, not the giraffe, is the one wearing pajamas. A human makes this correct interpretation of the sentence based on their knowledge about giraffes and pijamas.

Natural language processing, on the other hand, breaks the sentence into nouns, verbs, and other parts of speech. Then, NLP technology transforms the imprecise text in documents, contracts, spoken language, and silly sentences about giraffes into a precise hierarchy of related and labeled components. These sentence structures can be used on their own, but they are often used as inputs into machine learning algorithms to predict outcomes.

Legal professionals might use NLP to speed up document review. For example, you could feed a contract through a natural language processing program to help identify specific clauses, double-check the legal language, and ensure the contract includes all clauses that you need to stay compliant with your standard operating procedures.

Machine Learning

Machine learning is broadly broken down into two major categories: supervised and unsupervised. Legal tech uses both types to do a wide range of things, for example:

• Identify key clauses in non-disclosure agreements (mixed)
• Review an invoice for guideline violations (primarily supervised)
• Find similarities across legal briefs to find other relevant case citations (primarily unsupervised)

Supervised Machine Learning

The primary goal of supervised machine learning algorithms is to make predictions. It’s called “supervised” machine learning because you provide the answers to teach the algorithm. You supply lots of example data inputs with known outcomes (called “training data”) so that the algorithm can learn how to predict future outcomes. By feeding the algorithm previously observed data, you train the algorithm to make future predictions.

Supervised machine learning can be thought of as a very sophisticated set of if/then rules. The key difference is that with if/then rules, a person has to define specifically how a set of inputs lead to a result; ”if this, then that”. With supervised machine learning, the person doesn’t write the rules. Instead the person provides examples of all the inputs and results, then the computer figures out the rules that determine how the inputs lead to results.

Supervised machine learning has many applications in the legal industry. For instance, you could train an algorithm to review intellectual property (IP) case law and the outcomes of patent prosecution cases. Then, that algorithm could make educated predictions about future patent prosecution case outcomes.

Unsupervised Machine Learning

In unsupervised machine learning, the input data do not have answers or outputs. Instead, these algorithms, often called clustering algorithms, try to identify clusters of similar data inputs or observations. In the legal industry, an unsupervised machine learning algorithm might use clustering to identify certain types of contract clauses by finding similarities to other similar clauses.

Not quite seeing the difference between supervised and unsupervised? Let’s say you’re brand new to Spotify. Without months of listening history to refer to, Spotify’s algorithm will base its recommendations on the first songs and artists you listen to instead of your personal preferences. Recommendations are based on “clusters” or similar data, such as “female pop singers” or “80s hair bands.”

The legal industry uses unsupervised machine learning to analyze text. For instance, if you’re trying to use standard language for limited liability clauses in third-party contracts, you could use an unsupervised machine learning algorithm to review those contracts. The clustering algorithm will find those clauses and language related to liability in other clauses throughout a contract, making it easier (and faster) for a human to review the information. Not only can clusters presort information for people, but the clustered information can also be used to create training data for supervised machine learning algorithms.

How Does AI in Legal Work?

In practice, combinations of multiple supervised and unsupervised algorithms are often linked together to make a prediction or produce an outcome. Legal is a text-heavy world with a particular focus on documents and contracts. In those cases, natural language processing is used to turn paragraphs of text into structured data to be used as inputs into those algorithms. So, once again, “AI” makes for a convenient shorthand.

Additional use cases of AI in legal include:

• Risk-scoring of sales contracts for enterprise software companies
• Routing lease agreements for approval for a property management company
• Predicting outcomes of litigation for litigation financing companies
• Reviewing legal bills for compliance with your billing guidelines especially for law firms that bill hourly

Legal AI Aims to Help Humans, Not Replace Them

While there are many great practical uses for Legal AI tools today, they’re not substitutes for human lawyers. AI in Legal is not magic or even truly AI — it’s just math. But the components that make up what we call “Legal AI” are still extremely useful. Together, the many aspects of Legal AI allow legal departments to save money, accelerate revenue, close deals faster, get contracts signed sooner, and apply business rules more consistently.

Despite the hype, AI technology is here, and it’s here to help legal serve its customers.