Artificial Intelligence is not a singular technology but rather a field composed of several specialized subsets, each contributing uniquely to the overall goal of replicating aspects of human intelligence in machines. 

Subset when defined is a set that is part of another, larger set. When it comes to AI, subsets are categories of more specific AI systems that have distinct elements but nonetheless comprise the overall AI field.

While these subsets interconnect through shared technologies and overlapping functions, they vary in scope and application. For example, Machine learning (ML), a key AI subset which is explained further, underpins many advancements, including specialized techniques like deep learning. Its versatility and integration with subsets like natural language processing and computer vision highlight AI's interconnected and collaborative nature.

Machine Learning (ML)

Machine learning, a significant branch of AI, focuses on creating algorithms that allow machines to identify patterns in data and make decisions or predictions without explicit programming. The original definition provided by Arthur Samuel in the 1950s reflects the description given above for Machine Learning. The branch of ML is foundational in many AI applications today, from conversational chatbots to spam filters.

An intuitive way to understand ML’s mechanism is through interacting with an AI chatbot: you might give it an initial, complex prompt, followed by simpler commands. The chatbot, trained on an ML algorithm, can interpret and respond to each new command by drawing from prior prompts, showing how ML algorithms analyze and adjust based on previous interactions. Spam detection in email, a real-life example of ML, functions similarly by recognizing spam through patterns based on historical data. Another common ML application is fraud detection in banking, where algorithms monitor transactions and flag potential fraud based on anomalous patterns. For instance, JP Morgan integrated an AI model with its banking system. This integration ensured the AI model complemented rather than disrupted the bank’s operations. How it works in practice, is that the AI model analyzes transactions by building profiles of usual customer behavior from historical data. When a transaction noticeably differs from the norm, it gets flagged for review. Over time, AI learns and adapts, enhancing its accuracy and lowering false positives. The PayPal flagging system uses machine learning to detect fraud. 

‍

How Machine Learning Works

Machine learning operates by training algorithms on datasets to achieve specified objectives like pattern detection or object recognition. The effectiveness of ML depends on high-quality training data and repeated cycles of model adjustment, known as the fitting process.

Types of Machine Learning

• Supervised Learning: This type relies on labeled data to train algorithms, refining model parameters through cross-validation to avoid overfitting. Widely used in spam classification, techniques include neural networks, naïve Bayes, linear and logistic regression, random forests, and support vector machines (SVMs). One example of supervised learning is facial recognition in social media applications, which labels data points based on user-provided inputs.
• Unsupervised Learning: Used to identify hidden patterns or clusters within unlabeled data, this type is useful in tasks such as exploratory data analysis and customer segmentation. Techniques include neural networks, k-means clustering, and principal component analysis (PCA).  An example of unsupervised learning is customer segmentation by e-commerce sites that categorize shoppers based on purchasing behaviors.
• Semi-Supervised Learning: Combining supervised and unsupervised approaches, this technique uses a small amount of labeled data alongside a large amount of unlabeled data to improve classification accuracy. For instance, Social media platforms analyze both labeled and unlabeled user data to improve content recommendations.
• Reinforcement Learning: This type differs by teaching algorithms through trial and error, reinforcing successful strategies. Reinforcement learning is well-suited for decision-making tasks. For example, Google’s AlphaGo used reinforcement learning to improve its strategy and ultimately defeated human champions in the game Go.

‍

Deep Learning

Deep learning is a subset of machine learning that uses multilayered neural networks, called deep neural networks, to stimulate the complex decision-making power of the human brain. Simply put, a neural network is a machine learning program or model that makes the decisions. Every neural network consists of layers of nodes or artificial neurons - an input layer, one or more hidden layers, and an output layer. All nodes connect to each other, they get activated and data passes from one layer to the next one of the network. An example of deep learning is autonomous vehicles use deep learning to analyze visual data and navigate in real time. 

Another example is contract review platforms powered by deep learning that can identify key clauses, assess risks, and flag unusual terms in large volumes of contracts, saving time and reducing errors. Deep learning is particularly relevant to contract review because it automatically extracts features from complex datasets, requiring less human intervention than traditional machine learning. While ML often relies on manual adjustments by engineers, DL can independently refine its understanding, enabling it to identify nuanced clauses and assess risks with greater precision.
‍

Other Subsets of AI

The versatility of machine learning not only establishes it as a core component of AI, but it provides a foundation to other subsets of AI, as they share functions and training mechanisms at times. This interconnected framework enhances the specialized functions of each subset, enabling AI systems to perform diverse tasks with greater precision and adaptability across legal and other complex domains.

‍

Natural Language Processing (NLP)

NLP enables AI systems to interpret and generate human language, crucial for tools that streamline legal document drafting, automate client communication, and assist with legal research. NLP helps lawyers interact with AI tools naturally, enhancing usability and efficiency. For instance, Google Translate uses NLP to understand text structure and meaning across languages, while legal research tools like Westlaw leverage NLP to interpret search queries within legal contexts.

‍

Computer Vision

Although commonly seen as an application of ML, computer vision is focused on enabling machines to understand and process visual information. It is a foundational technology for tasks like facial recognition and autonomous driving. 

In legal practice, computer vision can analyze visual content from evidence such as images and videos, especially in cases involving forensic analysis or IP disputes. While often powered by ML, computer vision applications stand out in cases where visual data is critical. One example is that in IP cases, computer vision algorithms can detect trademark infringements by analyzing images of logos and designs, identifying unauthorized uses that might otherwise go unnoticed. Additionally, Apple’s Face ID technology, used for secure device access, relies on computer vision to identify user features accurately.

‍

Expert Systems

Expert systems stimulate human expertise in specific legal domains, applying rule-based logic to aid decision-making. These systems are widely used in legal document assembly and compliance verification, where consistent rule application is essential. For example, Document automation platforms use expert systems to generate standard legal documents, like wills or non-disclosure agreements (NDAs), by following predefined legal rules and conditions based on input from the user. Another example is IBM’s Watson Health aids in diagnosing medical conditions by analyzing symptoms against a database of medical knowledge.

‍

Generative AI

Generative AI (GenAI) is a subset of artificial intelligence technology that can produce various types of content, including text, imagery, audio, and synthetic data. 

Simply put, Generative AI creates new content based on existing data, including document drafting, legal research summaries, and client correspondence. This facet has been explored extensively in the last years and the use of GenAI will be able to revolutionize legal work. A survey from LexisNexis indicated that 92% of lawyers stated that it would impact their line of work to some extent. Half of these lawyers also answered that they believe GenerativeAI will significantly transform their business. 

Large Language Models (LLMs) are a category of foundation models trained on immense amounts of data making them capable of understanding and generating natural language and other types of content to perform a wide range of tasks.”  To best understand the new tools that have emerged, it is key to understand the term of foundation models. 

Foundation models are models trained on broad data at scale such that they can be adapted to a wide range of downstream tasks.

LLMs provide insights by synthesizing vast amounts of legal information.  As the model refines, it becomes increasingly accurate in its predictions, learning from data patterns to produce relevant results.  For example, in legal work, AI can be trained to recognize the IRAC (Issue, Rule, Application, Conclusion) structure commonly used in judgments and legal writing.

Vertical LLMs are a sub-category within LLMs, which also are based on foundation models and are industry-specific.

Currently, Vertical LLMs are seen as transformative for many industries. In fintech, vLLMs companies are helping automate tax compliance for CPAs and tax firms, in healthcare companies are leveraging AI speech recognition to automate real-time documentation of clinician-patient conversations. In legal we, at casepal, are making accurate and secure AI tools for law firms and in-house teams that address key legal tasks such as research & analysis, drafting, and summarization.

‍

Robotics and Artificial General Intelligence (AGI)

Robotics combines AI with engineering to create autonomous systems that perform physical tasks, while AGI represents a theoretical subset aiming to achieve general, human-like intelligence. While AGI remains a future goal, robotics is already transforming physical aspects of legal work, such as in litigation support for document handling and courtroom setups. Robotics could also play a role in law firm logistics, handling routine tasks that free up lawyers for more complex client work. For instance, law firms use automated robots in document storage facilities to retrieve files quickly, enhancing efficiency in large firms with extensive physical archives. In warehouses too, Amazon’s robots manage inventory using AI to navigate spaces and optimize order processing.

‍

Artificial Intelligence encompasses diverse subsets, each uniquely advancing the replication of human intelligence in machines. In law, subsets like machine learning, natural language processing, and deep learning enable powerful tools for tasks such as document drafting, case analysis, and client communication. These technologies enhance efficiency and precision, offering lawyers new ways to tackle complex tasks and improve service delivery.

‍

References:

Google Cloud, 'What is Machine Learning?' (Google Cloud) https://cloud.google.com/learn/what-is-machine-learning#how-does-machine-learning-work accessed 19 October 2024.

MIT Sloan, 'Machine Learning Explained' (MIT Sloan, 15 May 2020) https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained accessed 19 October 2024.

IBM, 'What is Machine Learning?' (IBM) https://www.ibm.com/topics/machine-learning accessed 19 October 2024.

TechTarget, 'What is Artificial Intelligence (AI)?' (TechTarget, 2024) https://www.techtarget.com/searchenterpriseai/definition/AI-Artificial-Intelligence accessed 19 October 2024.

IBM, 'What Are Neural Networks?' (IBM) https://www.ibm.com/topics/neural-networks accessed 1 November 2024.

OpenAI, 'Models' (OpenAI, 2024) https://platform.openai.com/docs/models#o1 accessed 12 November 2024

Jeyadev Needhi, 'How AI Transformed Financial Fraud Detection: A Case Study of JP Morgan Chase' (Medium, 28 August 2023) https://medium.com/@jeyadev_needhi/how-ai-transformed-financial-fraud-detection-a-case-study-of-jp-morgan-chase-f92bbb0707bb accessed 12 November 2024

Alexander S. Gillis, 'What is Generative AI?' (TechTarget, 2024) https://www.techtarget.com/searchenterpriseai/definition/generative-AI accessed 12 November 2024

LexisNexis, 'LexisNexis Launches Generative AI Solution for Legal Professionals' (LexisNexis, 2024) https://www-lexisnexis-com.mu.idm.oclc.org/html/lexisnexis-generative-ai-story/ accessed 12 November 2024

Bernard Marr, 'How Generative AI Will Change the Jobs of Lawyers' (Forbes, 14 March 2024) https://www.forbes.com/sites/bernardmarr/2024/03/14/how-generative-ai-will-change-the-jobs-of-lawyers/ accessed 12 November 2024

LexisNexis, 'International Legal Generative AI Survey Shows Nearly Half of the Legal Profession Believe Generative AI Will Transform the Practice of Law' (LexisNexis, 2024) https://www-lexisnexis-com.mu.idm.oclc.org/community/pressroom/b/news/posts/lexisnexis-international-legal-generative-ai-survey-shows-nearly-half-of-the-legal-profession-believe-generative-ai-will-transform-the-practice-of-law accessed 12 November 2024

National Network of Libraries of Medicine, Natural Language Processing (NNLM, 2024) https://www.nnlm.gov/guides/data-glossary/natural-language-processing accessed 12 November 2024

Thomson Reuters, Legal Research Meets Generative AI (Thomson Reuters, 2024) https://legal.thomsonreuters.com/blog/legal-research-meets-generative-ai/ accessed 12 November 2024.

Thomson Reuters, Natural Language Processing (Thomson Reuters, 2024) https://www-thomsonreuters-com.mu.idm.oclc.org/en/artificial-intelligence/natural-language-processing.html accessed 13 November 2024

Xiaoxi Kang and others, 'Can ChatGPT Perform Reasoning Using the IRAC Method in Analyzing Legal Scenarios Like a Lawyer?' (Monash University, 2024).

PayPal, 'Payment Fraud Detection and Machine Learning' (PayPal, no date) https://www.paypal.com/us/brc/article/payment-fraud-detection-machine-learning accessed 21 November 2024.

Google Cloud, 'Deep Learning vs. Machine Learning' (Google Cloud, no date) https://cloud.google.com/discover/deep-learning-vs-machine-learning#artificial-intelligence-vs-machine-learning-vs-deep-learning accessed 21 November 2024.

IBM, 'Large Language Models' (IBM, no date) https://www.ibm.com/topics/large-language-models accessed 21 November 2024.

‍

‍