AI Glossary

75 words you will meet when you learn AI, each explained in plain English with an everyday example. Every term has its own page. Start with the basics, or jump to a section.

New to all of this? Start the AI Basics course →

The basics

AlgorithmAn algorithm is a precise, step-by-step set of instructions that a computer follows to solve a problem or complete a task.APIAn API (application programming interface) is a set of rules that lets one program ask another for data or actions, such as sending a question to an AI model.Artificial general intelligenceArtificial general intelligence (AGI) is a hypothetical AI that could learn and perform any intellectual task a human can, rather than only specific tasks.Artificial intelligenceArtificial intelligence (AI) is software that does tasks that usually need human thinking, like recognizing faces, understanding speech, or making decisions.ChatbotA chatbot is a computer program that holds a conversation with people through text or voice, ranging from simple scripted menus to AI assistants.Deep learningDeep learning is machine learning that uses neural networks with many layers, so a computer can learn very complex patterns from huge amounts of data.Generative AIGenerative AI is artificial intelligence that creates new content, such as text, images, music, code, or video, from patterns it learned in existing examples.InferenceInference is the stage where a trained AI model is actually used: it takes a new input and produces a prediction or an answer.Machine learningMachine learning is a way of building software where the computer learns patterns from examples, instead of a programmer writing every rule by hand.ModelIn AI, a model is the trained program that has learned patterns from data and can now make predictions or generate output for new inputs.TrainingTraining is the process of teaching an AI model by showing it many examples and adjusting its internal numbers until its answers become accurate.Turing testThe Turing test is a 1950 thought experiment by Alan Turing: if a person chatting by text cannot tell a machine from a human, the machine passes.

How machines learn

AccuracyAccuracy is the share of a model's predictions that are correct, calculated as the number it got right divided by the total number it tried.BackpropagationBackpropagation is the algorithm that works out how much each weight in a neural network contributed to an error, so every weight can be adjusted correctly.BenchmarkA benchmark is a standard set of test tasks used to measure and compare AI models, so different systems can be scored on exactly the same questions.ClassificationClassification is a machine learning task where the model sorts each input into one of a fixed set of categories, such as spam or not spam.ClusteringClustering is an unsupervised learning method that groups similar data points together automatically, without being told in advance what the groups are.False positiveA false positive is a mistake where a model says 'yes' when the true answer is 'no', like flagging a safe email as spam or a healthy patient as sick.Gradient descentGradient descent is the step-by-step method used to train models: it repeatedly adjusts the model's numbers in the direction that most reduces its error.Loss functionA loss function is the score that measures how wrong a model's predictions are; training adjusts the model to make that score as small as possible.OverfittingOverfitting is when a model memorizes its training examples, including noise and quirks, so it performs well on them but poorly on new, unseen data.RegressionRegression is a machine learning task where the model predicts a number, such as a price, a temperature, or a delivery time, rather than a category.Reinforcement learningReinforcement learning is machine learning where a program learns by trial and error, receiving rewards for good actions and penalties for bad ones.Supervised learningSupervised learning is machine learning where the model trains on examples that come with the correct answers, so it can learn to predict answers for new data.ThresholdA threshold is the cutoff score a model's confidence must reach before it counts as a 'yes'; moving it trades fewer false alarms for more missed cases.Unsupervised learningUnsupervised learning is machine learning on data that has no labels, where the model must find structure, groups, or patterns on its own.

Neural networks

Language models & chatbots

AI agentAn AI agent is a system that uses a language model to plan and carry out multi-step tasks on its own, deciding which actions and tools to use along the way.ChatGPTChatGPT is OpenAI's chatbot, released in November 2022, that lets people talk to a large language model; GPT stands for Generative Pre-trained Transformer.Context windowThe context window is the maximum amount of text, measured in tokens, that a language model can consider at once, including your prompt and its own reply.EmbeddingAn embedding is a list of numbers that represents the meaning of a piece of text, image, or other data, arranged so similar meanings produce similar numbers.Fine-tuningFine-tuning is taking an already-trained model and training it a bit more on a smaller, specific set of examples so it performs better at a particular job.HallucinationAn AI hallucination is when a language model states something false or invented as if it were a fact, such as a fake quote, citation, or event.Large language modelA large language model (LLM) is an AI trained on vast amounts of text to predict the next word, which lets it write, answer questions, summarize, and converse.Open weightsAn open-weights model is one whose trained parameters have been published, so anyone can download it, run it on their own hardware, and modify it.PretrainingPretraining is the first and biggest stage of building a language model, where it learns general language and knowledge by predicting text across huge datasets.PromptA prompt is the text you give an AI model to tell it what you want; it can be a question, an instruction, an example, or a whole document to work with.Prompt engineeringPrompt engineering is the practice of writing and refining prompts so that an AI model gives more accurate, useful, and consistent results.RAGRAG (retrieval-augmented generation) is a technique where an AI looks up relevant documents first, then uses them to write an answer grounded in real sources.RLHFRLHF is a training step where human raters compare a model's answers and the model is then adjusted to produce the kind of replies people prefer.Semantic searchSemantic search is search that finds results by meaning rather than exact keywords, so a query and a document can match even when they share no words.System promptA system prompt is hidden instruction text set by the app developer that tells an AI model how to behave, before the user's own message is added.TemperatureTemperature is a setting that controls how random a language model's replies are: low values give focused, predictable text; high values give more variety.TokenA token is the unit of text a language model reads and writes; it is usually a word, part of a word, or a punctuation mark, roughly four characters in English.Tool callingTool calling lets a language model request that an outside program run, such as a calculator, database, or web search, and then use the result in its answer.VectorA vector is an ordered list of numbers; in AI it is how meaning, images, and other data are represented so computers can compare and calculate with them.Vector databaseA vector database is a storage system built to hold embeddings and quickly find the ones most similar to a query, which powers search by meaning.

Data

Generative AI & the senses

Safety & ethics

Everyday AI