Artificial Intelligence: AI Glossary

The American Association of Law Libraries (AALL) AI and Legal Information Special Committee created this useful AI Key Term Glossary (in the righthand column) in May 2024.

These other glossaries are also helpful in understanding the AI lexicon.

Application Programming Interface (API): This is a way for two or more computer programs to communicate with each other. An API is like waitstaff in a restaurant, taking your order and delivering it to the kitchen, and returning with food.

Artificial Intelligence (AI): A non-human program or model that can solve sophisticated tasks.

Chatbot: A chatbot is a computer program that simulates human conversation with an end user. Not all chatbots are equipped with AI, but modern chatbots increasingly use conversational AI techniques. Imagine a chatbot is like a magic talking book. Some are like simple chapter books that can only tell you things they already know. Some are choose-your-own-adventure books that can create new stories. These chatbots may use AI.

Context Window: The span of text that a large language model (LLM) can “see” at any given moment. It’s like looking through a camera lens, where only part of the scene is visible.

Deep Learning: A subset of machine learning that models high-level abstractions in data using multiple processing layers. It’s like a detective solving complex cases by finding patterns in clues over time.

Generative AI (GAI): This involves language models that can generate content that is complex, coherent, and original. Like someone's imagination, it takes things it believes it knows and creates new things that it has not seen, heard, or written before.

Generative Pre-Trained Transformer (GPT): GPT is a generative AI technology that has been previously trained to transform its input into a different type of output. GPT models are general-purpose language prediction models. They are like a skilled storyteller who crafts tales from just a few words given by the audience.

Grounding: This is the ability of AI to connect output to verifiable sources of information, things that actually exist. It is as if a robot read the word "apple" but did not really understand what an apple represented: that it is eaten, what it tastes like, and that the core is often not eaten.

Hallucination: An incorrect response from AI, where the model generates outputs not grounded in its input data. It’s like a mirage in the desert showing water where there is none.

Human in the Loop: This is a phrase that clarifies human involvement in training of an AI model to make the correct decision. Unlike supervised training, it can occur at any stage of the AI's operation. For example, when generative AI is used to create content that could have ethical implications, such as deepfakes, human oversight is crucial to prevent misuse and that output aligns with ethical guidelines.

Large Language Models (LLM): A language model is a computer program that has been taught how to understand and use language and figure out what your words mean. A large language model can look at more of your words to try to understand your meaning. Think of a large tree that has rings that tell its age. A language model has layers of information that it uses to understand and generate language. Just as older trees have more rings, large language models can handle more layers of information.

Machine Learning: A program or system that trains a system like a language model from input data. Like someone who is learning to cook a new dish. Over time, the more attempts they make, the more they learn about how to perfect the dish.

Model Brands: When we talk about ChatGPT, Llama, or Midjourney, we are referring to the brand of an AI language model.

Natural Language Processing (NLP): Natural language processing combines rule-based modeling of human language with statistical and machine learning models to enable computers recognize, understand, and generate text and speech. It's like when a librarian is asked, "Can you find me stories about pets?” and you provide books on animals that people might keep at home, like cats or dogs, even if you were not asked for those exact words.

Neural Network: Computational models that replicate the functioning of the human brain, designed to recognize patterns in data. Imagine it as a network of roads connecting information like cities with traffic flow directing decisions.

Prompt: The directions or what you are asking the generative AI tool to do or create. It’s like giving a musician a few starting notes and asking them to play a song based on those notes.

Prompt Design/Engineering: Prompt engineering is the practice of designing inputs for AI tools that will produce optimal outputs. Imagine it like writing a good exam question that guides students to show their true understanding.

Retrieval-Augmented Generation (RAG): This is the process of optimizing the output of a large language model, so it references an authoritative knowledge base outside of its training data sources before generating a response. It's a form of grounding. It’s like having a librarian who quickly fetches the best books relevant to your query to help you write a detailed report.

Search Algorithm: It is a set of steps or rules that computers use to quickly find information in a huge amount of data.

Supervised Learning: Supervised learning is a type of machine learning that uses labeled data to train algorithms to recognize patterns and predict outcomes. It’s a bit like when you teach a dog tricks by showing it what to do and then rewarding it.

Token: Tokens can be thought of as pieces of words. Before the API processes the request, the input is broken down into tokens. Think of tokens as a collection of puzzle pieces in a larger picture that the AI is trying to assemble.

Unsupervised Learning: Unsupervised learning in artificial intelligence (AI) is a type of machine learning that learns from data without human guidance. It’s like a child exploring a toy box to figure out what each toy does without any other input.