Gaining AI Fluency in the Legal Profession: Understanding Machine Learning

Michael Callier
July 7, 2023

Artificial intelligence (AI) has become a significant influence in our lives and professions, yet its inner workings often remain mysterious to many. This article aims to demystify machine learning, a crucial branch of AI for the legal community, by explaining it in simple terms. By drawing parallels between how machines and humans learn, we can gain insights into how AI functions, understand its relevance to our own cognitive processes, and make smarter decisions about how we use AI generally.

Understanding Machine Learning 

Machine learning is a subset of AI that aims to mimic human learning processes. It involves feeding large amounts of data into algorithms that learn and identify patterns. This enables the algorithms to apply their learnings to new information they encounter. For example, machine learning helps computers differentiate between different agreement types, identify specific clauses within those agreements and derive meaning from them. Every prominent legal technology that leverages AI today uses some form of machine learning including tools like Evisort, Brightflag and Casetext.  

Three Basic Paradigms of Machine Learning 

  1. Supervised Learning: In supervised learning, computers are trained using labeled examples, much like students learning from teachers. The computer receives input data along with the correct output or label associated with it. By learning the underlying patterns or rules connecting the input data to the output, the computer can make predictions or classifications on new, unseen data. In the legal field, this can assist with tasks such as document classification, contract analysis or predicting legal outcomes based on historical data. We of manually label documents with relevant metadata to support search and typically refer to it as “tagging”. 
  2. Unsupervised Learning: Unsupervised learning occurs when computers learn from unlabeled data, exploring it to find patterns or structures without explicit guidance. It's like a detective finding hidden clues and solving a mystery. In the legal field, unsupervised learning can help cluster similar legal documents or identify topics within large sets of legal texts. Tools like Relativity and Kira leverage unsupervised learning to assist legal professionals in gaining insights from complex, unstructured data. Many in the legal profession have used unsupervised learning to support eDiscovery and mergers and acquisition projects. 
  3. Reinforcement Learning: Reinforcement learning involves a computer system learning through trial and error, similar to how we learn from rewards and consequences. The computer, acting as an agent, interacts with an environment and receives feedback based on its actions. Over time, the agent learns which actions lead to favorable outcomes and adjusts its behavior accordingly. In the legal context, reinforcement learning can help develop systems for legal decision-making, providing guidance on the best course of action in a legal dispute or offering advice based on previous outcomes. It's like having a virtual legal assistant that learns and improves its decision-making abilities over time. For example, AlphaGo learned to outperform humans by playing millions of games against itself and ChatGPT learned how to stick to human values and have human-like dialogue by chatting for months with humans. 

Transfer Learning: Building on Prior Knowledge 

Transfer learning is a technique that leverages knowledge gained from solving one problem and applies it to a different but related problem. It's like using past experience to tackle new situations. By using pre-trained models that have learned from vast datasets, transfer learning enables more efficient learning on new tasks with less labeled data. This approach accelerates the development of natural language processing (NLP) applications in the legal industry. For example, BERT (Bidirectional Encoder Representations from Transformers), a large language model, is used in Google's search engine to understand the context and meaning of search queries and web pages. LEGAL-BERT is a family of BERT models fine-tuned on publicly available legislation, court cases and contracts. It performs better than BERT out of the box for NLP tasks on law-related information.  

Zero-Shot Learning: Embracing the Unseen 

Zero-shot learning empowers models to recognize and understand new types of data they have never encountered before. By utilizing additional information or contextual cues about these new types, the models can make accurate predictions. However, it is important to recognize that zero-shot learning has limitations and human-in-the-loop quality control is essential to ensure reliable results. Factors like volume of training data, model size (i.e., how many parameters), task complexity and optical character recognition or OCR scan quality heavily influence zero-shot learning success. As a general rule, zero-shot learning does not work 100 percent of the time so users should rely on human-in-the-loop quality control. 

While no model or approach is perfect, understanding the basics of AI and machine learning empowers informed decision-making. When considering AI tools or applications, legal professionals can ask vendors about the machine learning methods employed, the use of pre-trained models and their fine-tuning process. By gaining AI fluency, legal professionals can harness the power of technology to enhance their work. We can also make better decisions about which AI technology to buy and better manage expectations around how well it works. By embracing the principles of machine learning and understanding its practical applications, legal professionals can navigate the evolving landscape of AI and leverage its benefits for professional success. 

Here are a few questions you can ask technology vendors when exploring AI-enabled tools: 

  • What machine learning methods do you use to achieve what outcomes?  
  • What large language models (LLMs) are incorporated into the tool? and whether they are out-of-the-box?  
  • Are the LLMs fine-tuned for legal? If so, tell me about the fine-tuning process including length of time and overall capital investment.   
  • What is the total number of model parameters (GPT-3.5 has over 175 billion parameters)? 
  • What are typical scores for zero-shot learning?  

To explore how AI can enhance your department's operating environment or increase your AI fluency, please contact me at