Sourced from the Center for Integrative Research in Computing and Learning Sciences' (CIRCLS) Glossary of Artificial Intelligence Terms for Educators

• A
  • Adaptive Learning: Subject or course material is adjusted based on the performance of the learner. The difficulty of material, the pacing, sequence, type of help given, or other features can be adapted based on the learner’s prior responses.
  • Algorithm: Algorithms are the “brains” of an AI system and what determines decisions in other words, algorithms are the rules for what actions the AI system takes. Machine learning algorithms can discover their own rules (see Machine learning for more) or be rule-based where human programmers give the rules.
  • Artificial General Intelligence (AGI): Artificial general intelligence has not yet been realized and would be when an AI system can learn, understand, and solve any problem that a human can.
  • Artificial Intelligence (AI): AI is a branch of computer science. AI systems use hardware, algorithms, and data to create “intelligence” to do things like make decisions, discover patterns, and perform some sort of action. AI is a general term and there are more specific terms used in the field of AI. AI systems can be built in different ways, two of the primary ways are: (1) through the use of rules provided by a human (rule-based systems); or (2) with machine learning algorithms. Many newer AI systems use machine learning (see definition of machine learning below).
  • Artificial Narrow Intelligence (ANI): AI can solve narrow problems and this is called artificial narrow intelligence. For example, a smartphone can use facial recognition to identify photos of an individual in the Photos app, but that same system cannot identify sounds.
• B
  • Bard: Bard is a conversational chatbot released by Google. It has similar functionality to OpenAI’s ChatGPT.
  • Black Boxes: We call things we don’t understand, “black boxes” because what happens inside the box cannot be seen. Many machine learning algorithms are “black boxes” meaning that we don’t have an understanding of how a system is using features of the data when making their decisions (generally, we do know what features are used but not how they are used)There are currently two primary ways to pull back the curtain on the black boxes of AI algorithms: interpretable machine learning and explainable machine learning.
• C
  • Chatbot: A program that runs within a website or app that interacts directly with a user through a lifelike “conversation.”
  • ChatGPT: ChatGPT is a conversational chatbot released by OpenAI. It has very broad functionality, ranging from creative storytelling to writing code that is used in computer programming.
  • Computer Vision: Computer Vision is a set of computational challenges concerned with teaching computers how to understand visual information, including objects, pictures, scenes, and movement (including video). Computer Vision (often thought of as an AI problem) uses techniques like machine learning to achieve this goal.
  • Critical AI: Critical AI is an approach to examining AI from a perspective that focuses on reflective assessment and critique as a way of understanding and challenging existing and historical structures within AI.
• D
  • DALL-E: The DALL-E family of models are text-to-image generative models created by OpenAI starting in 2021. These models take some text input and produce an image.
  • Data: Data are units of information about people or objects that can be used by AI technologies.
  • Deep Learning: Deep learning models are a subset of neural networks. With multiple hidden layers, deep learning algorithms are potentially able to recognize more subtle and complex patterns. Like neural networks, deep learning algorithms involve interconnected nodes where weights are adjusted, but as mentioned earlier there are more layers and more calculations that can make adjustments to the output to determine each decision. The decisions by deep learning models are often very difficult to interpret as there are so many hidden layers doing different calculations that are not easily translatable into English rules (or another human-readable language).
• E
  • Explainable Machine Learning (XML) or Explainable AI (XAI): Researchers have developed a set of processes and methods that allow humans to better understand the results and outputs of machine learning algorithms. This helps developers of AI-mediated tools understand how the systems they design work and can help them ensure that they work correctly and are meeting requirements and regulatory standards.
    • It is important to note that the term “explainable” in the context of explainable machine learning or explainable AI, refers to an understanding of how a model works and not to an explanation of how the model works. In theory, explainable ML/AI means that an ML/AI model will be “explained” after the algorithm makes its decision so that we can understand how the model works. This often entails using another algorithm to help explain what is happening as the “black box.” One issue with XML and XAI is that we cannot know for certain whether the explanation we are getting is correct, therefore we cannot technically trust either the explanation or the original model. Instead, researchers recommend the use of interpretable models.
• F
  • Fine Tuning:  It is often desirable to take an existing deep learning model that has already been trained, and repurpose it for a slightly different application. Suppose that there is a pre-trained model that classifies images into one of 1000 different categories (e.g., dog, cat, human, table, potato). Fine tuning would classify images of some subset of those 1000 categories (e.g., only dogs and cats). Creating such a specialized model can improve classification accuracy and performance for its specific purpose.
  • Foundation Models: Foundation Models represent a large amount of data that can be used as a foundation for developing other models. For example, generative AI systems use large language foundation models. They can be a way to speed up the development of new systems, but there is controversy about using foundation models since depending on where their data comes from, there are different issues of trustworthiness and bias. Jitendra Malik, Professor of Computer Science at UC Berkeley once said the following about foundation models: “These models are really castles in the air, they have no foundation whatsoever.”
• G
  • Generative AI (GenAI) and chat-based generative pre-trained transformer (ChatGPT) models: A system built with a neural network transformer type of AI model that works well in natural language processing tasks (see definitions for neural networks and Natural Language Processing below). In this case, the model: (1) can generate responses to questions (Generative); (2) was trained in advance on a large amount of the written material available on the web (Pre-trained); (3) and can process sentences differently than other types of models (Transformer).
  • GitHub CoPilot: GitHub’s Copilot is a tool that aids in computer programming. It can be integrated as a plugin to several commonly used text editors and IDEs (e.g., VS Code, Neovim, and JetBrains).
  • GPU: Deep learning models are generally trained on graphics processing units (GPUs). GPUs were originally developed for rendering graphics—especially 3-dimensional graphics—on computers. In the early 2000s researchers began experimenting with using GPUs for scientific computing. Use of GPUs for scientific computing is extremely effective because GPUs can perform many operations in parallel. Being able to parallelize linear algebraic operations has been particularly transformative for deep learning, and is why GPUs are such powerful tools to drive the development of generative AI technology.
• H
  • Hallucinations: The term hallucination is frequently used to describe those instances in which ChatGPT or similar conversational chatbots provide output that is factually incorrect, either in part or in whole. It is very common for generative AI tools to invent inaccurate “facts” that sound convincing, but have no basis in reality.
  • Human-centered Perspective: A human-centered perspective sees AI systems working with humans and helping to augment human skills. People should always play a leading role in education, and AI systems should not replace teachers.
• I
  • Inference: The inference phase involves using models to make predictions. Inference is conducted when a model is trained. When ChatGPT produces a reply to a question, that is the model running inference. When an iPhone’s camera places square bounding boxes around human faces, that is a model running inference.
  • Intelligence Augmentation (IA): Augmenting means making something greater; in some cases, perhaps it means making it possible to do the same task with less effort. Maybe it means letting a human (perhaps teacher) choose to not do all the redundant tasks in a classroom but automate some of them so they can do more things that only a human can do. It may mean other things. There’s a fine line between augmenting and replacing and technologies should be designed so that humans can choose what a system does and when it does it.
  • Intelligent Tutoring Systems (ITS): A computer system or digital learning environment that gives instant and custom feedback to students. An Intelligent Tutoring System may use rule-based AI (rules provided by a human) or use machine learning under the hood. By under the hood we mean the underlying algorithms and code that an ITS is built with. ITSs can support adaptive learning.
  • Interpretable Machine Learning (IML): Interpretable machine learning, sometimes also called interpretable AI, describes the creation of models that are inherently interpretable in that they provide their own explanations for their decisions. This approach is preferable to that of explainable machine learning (see definition below) for many reasons including the fact that we should understand what is happening from the beginning in our systems, rather than try to “explain” black boxes after the fact.
• L
  • Large Language Model (LLM): Language models are probabilistic models for predicting the next word in a sequence of words. Large language models are essentially hyper-parametric versions of language models—often having millions or even billions of parameters. ChatGPT, Bard, and other popular generative AI tools are powered by large language models.
  • LLaMA: The LLaMA—short for Large Language Model, Meta AI—family of models are large language models (LLMs) similar to the GPT family of models. The original LLaMA models were created by Facebook, who then released the models’ weights to the research community. These were subsequently leaked via BitTorrent.
• M
  • Machine Learning (ML): Machine learning is a field of study with a range of approaches to developing algorithms that can be used in AI systems. AI is a more general term. In ML, an algorithm will identify rules and patterns in the data without a human specifying those rules and patterns. These algorithms build a model for decision making as they go through data. (You will sometimes hear the term machine learning model.) Because they discover their own rules in the data they are given, ML systems can perpetuate biases. Algorithms used in machine learning require massive amounts of data to be trained to make decisions. 
    • It’s important to note that in machine learning, the algorithm is doing the work to improve and does not have the help of a human programmer. It is also important to note three more things. One, in most cases the algorithm is learning an association (when X occurs, it usually means Y) from training data that is from the past. Two, since the data is historical, it may contain biases and assumptions that we do not want to perpetuate. Three, there are many questions about involving humans in the loop with AI systems; when using ML to solve AI problems, a human may not be able to understand the rules the algorithm is creating and using to make decisions. This could be especially problematic if a human learner was harmed by a decision a machine made and there was no way to appeal the decision.
  • Model Collapse: Large language models (LLMs) are trained on massive text data sets. One mechanism of assembling these training sets is crawling popular sites on the internet. The concept of model collapse has been proposed by researchers as a mechanism by which LLMs may gradually degrade in quality because of the nature of the text on which they are trained. Specifically, model collapse describes the degradation of LLMs by a “feedback loop” that will be created as a result of training an LLM on data that was generated by another LLM—rather than the human-generated data that currently makes up most of the text on the internet. As LLMs are used more and more, their output will become an increasingly large proportion of the text on the internet. And as such LLMs trained in the future will be trained less on human-generated content, and more on content generated by earlier LLMs.
• N
  • Natural Language Processing (NLP): Natural Language Processing is a field of Linguistics and Computer Science that also overlaps with AI. NLP uses an understanding of the structure, grammar, and meaning in words to help computers “understand and comprehend” language. NLP requires a large corpus of text (usually half a million words).
    • NLP technologies help in many situations that include: scanning texts to turn them into editable text (optical character recognition), speech to text, voice-based computer help systems, grammatical correction (like auto-correct or grammarly), summarizing texts, and others.
  • Neural Networks (NN): Neural networks also called artificial neural networks (ANN) and are a subset of ML algorithms. They were inspired by the interconnections of neurons and synapses in a human brain. In a neural network, after data enter in the first layer, the data go through a hidden layer of nodes where calculations that adjust the strength of connections in the nodes are performed, and then go to an output layer.
• O
  • OpenAI: OpenAI is an artificial intelligence startup responsible for ChatGPT and the various generations of the GPT family of large language models (e.g., GPT 3, GPT 4) that have powered ChatGPT.
• P
  • Pre-Trained Model: A deep learning model is considered “pre-trained” when the process of estimating the model’s parameters is complete. This is a very computationally expensive process, and can take hours, days, or weeks to complete, even when running on a supercomputer.
• R
  • Reinforcement Learning: A type of machine learning in which an agent learns to behave in an environment by performing actions and receiving rewards or punishments.
  • Robots: Robots are embodied mechanical machines that are capable of doing a physical task for humans. “Bots” are typically software agents that perform tasks in a software application (e.g., in an intelligent tutoring system they may offer help). Bots are sometimes called conversational agents. Both robots and bots can contain AI, including machine learning, but do not have to have it. AI can help robots and bots perform tasks in more adaptive and complex ways.
• S
  • Self-attention mechanism: These mechanisms, also referred to as attention help systems determine the important aspects of input in different ways. There are several types and were inspired by how humans can direct their attention to important features in the world, understand ambiguity, and encode information.
  • Supervised Learning: A type of machine learning in which a model learns from labeled data to make predictions or classifications.
• T
  • Training: Broadly speaking, the process of training a model amounts to estimating the model’s parameters. In the context of machine learning, training can be a very time consuming process—often taking hours, days, or weeks. In the case of deep learning specifically, this involves many “passes” through a training data set in small batches. The results of each batch are reviewed for the quality of its results, to tune the model. It is common to make dozens of passes through the data.
  • Training Data: This is the data used to train the algorithm or machine learning model. It has been generated by humans in their work or other contexts in their past. While it sounds simple, training data is so important because the wrong data can perpetuate systemic biases. If you are training a system to help with hiring people, and you use data from existing companies, you will be training that system to hire the kind of people who are already there. Algorithms take on the biases that are already inside the data. People often think that machines are “fair and unbiased” but this can be a dangerous perspective. Machines are only as unbiased as the human who creates them and the data that trains them.
  • Transformers:  In the context of machine learning, transformers are a species of layer within a deep learning model. Specifically, they use an “attention” mechanism first proposed in 2014 by Bahdanau (et al.). The transformer architecture has proven to be extremely flexible in deep learning models working with both text and image data.
  • Transformer models: Used in GenAI (the T stands for Transformer), transformer models are a type of language model. They are neural networks and also classified as deep learning models. They give AI systems the ability to determine and focus on important parts of the input and output using something called a self-attention mechanism to help.
• U
  • Unsupervised Learning: A type of machine learning in which a model learns patterns in data without being given any specific labels or guidance.
  • User Experience Design/User Interface Design (UX/UI): User-experience/user-interface design refers to the overall experience users have with a product. These approaches are not limited to AI work. Product designers implement UX/UI approaches to design and understand the experiences their users have with their technologies.