Skip to content

AI glossary

Plain definitions of the words you meet in AI, with an example and the year each one became common. Type to filter.

Version 1.1 · October 26, 2026 · What’s new

A–Z index, links between related terms, and a link to share any term.

Activation function

A small rule inside each neuron of a neural network that decides how strongly the neuron passes its signal on. Without it, a network could only learn straight-line relationships.

Like a dimmer switch that turns a weak signal down to almost nothing and lets a strong one through.

See also: Neural network, Deep learning

Adversarial attack

A deliberately crafted input meant to trick an AI model into a wrong answer, often with changes too small for a person to notice.

A few stickers on a stop sign that make a car’s vision system read it as a speed-limit sign.

Common since 2013

See also: Prompt injection, Jailbreak, Machine learning (ML)

AGI (artificial general intelligence)

A hypothetical AI able to learn and handle almost any mental task as well as a person, instead of being good at only a few things. Experts disagree on how to define it and whether or when it will exist.

Common since 2007

See also: AI (artificial intelligence), Alignment

AI agent

An AI system that does more than chat: it plans steps and uses tools, such as a web browser, files or other apps, to carry out a task with some independence.

Asking an agent to find three hotels under $150 and put them in a spreadsheet, and it does the searching and typing itself.

Common since 2023

See also: Large language model (LLM), Model Context Protocol (MCP), Guardrails

AI (artificial intelligence)

The field of building computer systems that perform tasks we usually link to human thinking, such as understanding language, recognizing images or making decisions.

Your phone sorting photos by the faces in them.

Common since 1956

See also: Machine learning (ML), AGI (artificial general intelligence), Generative AI

Alignment

The work of making AI systems pursue the goals and values their makers and users actually intend, and avoid harmful or unwanted behaviour.

Common since 2016

See also: RLHF (reinforcement learning from human feedback), Guardrails, AGI (artificial general intelligence)

Anomaly detection

Spotting data points that look very different from the normal pattern, which can signal a problem, an error or fraud.

Your bank flagging a card purchase made in another country at 3 a.m.

See also: Unsupervised learning, Clustering

Artificial neural network (ANN)

The full name for a neural network: software made of many simple connected units, loosely modelled on brain cells, that learns patterns from examples. See Neural network.

Common since 1958

See also: Neural network, Deep learning

Attention

A technique that lets a model weigh which parts of its input matter most for each piece it produces. It is the key idea behind the transformer.

In “The cat sat on the mat because it was tired”, attention helps the model link “it” to “cat”.

Common since 2014

See also: Transformer, Context window, Large language model (LLM)

Autoencoder

A neural network that learns to squeeze data into a compact summary and then rebuild the original from it. The summary keeps the most important features.

Used to clean up grainy photos or to spot unusual machine readings.

Common since 1986

See also: Neural network, Unsupervised learning, Embedding

Backpropagation

The method neural networks use to learn from mistakes: it works backwards from the error and figures out how much each connection should be adjusted.

Common since 1986

See also: Gradient descent, Loss function, Neural network

Bagging

Short for “bootstrap aggregating”: training many copies of a model, each on a random sample of the data, then combining their answers to get a steadier result.

Common since 1996

See also: Ensemble learning, Random forest, Boosting

Batch normalization

A step inside a neural network that rescales the numbers flowing between layers so they stay in a similar range, which helps training go faster and more smoothly.

Common since 2015

See also: Neural network, Training, Feature scaling

Benchmark

A standard set of tests used to measure and compare AI models on the same tasks. High scores do not always mean a model will do well on your own work.

A shared exam of maths and science questions that every new model is scored on.

See also: Evals (evaluations), Large language model (LLM)

Bias

Two meanings. Most often, an unfair slant in an AI’s results, usually inherited from its training data. In neural networks, it is also a small adjustable number added to each neuron.

A hiring tool that favours one group because past hiring data did.

See also: Training, Neural network, Alignment

Boosting

A way of building a strong model from many weak ones trained one after another, where each new model focuses on the mistakes of the previous ones.

Common since 1995

See also: Ensemble learning, Bagging, Decision tree

Chain-of-thought

Getting a model to work through a problem step by step before giving its final answer, which often improves results on maths and logic questions.

Adding “Explain your reasoning step by step” to a word problem.

Common since 2022

See also: Prompt engineering, Reasoning model, Few-shot prompting

Chatbot

A program you talk with in everyday language, by text or voice. Modern chatbots such as ChatGPT run on large language models.

The chat window on a store’s website that answers questions about delivery times.

Common since 1966

See also: Large language model (LLM), Prompt, NLP (natural language processing)

Classification

A machine learning task where the model sorts each item into one of a set of categories chosen in advance.

Deciding whether an email is spam or not spam.

See also: Supervised learning, Regression, Precision

Clustering

Letting an algorithm group items that resemble each other, without telling it the groups beforehand.

A shop discovering that its customers fall into a few natural buying styles.

See also: Unsupervised learning, K-means clustering

Computer vision

The branch of AI that helps computers make sense of photos and video: what is in them, where, and what is happening.

Self-checkout scales that recognize the apples you put on them.

Common since 1966

See also: Convolutional neural network (CNN), Multimodal, AI (artificial intelligence)

Context window

The maximum amount of text, measured in tokens, that a language model can take into account at once, including your messages, any files and its own replies.

In a very long chat, details from the start can fall out of the window and be forgotten.

Common since 2020

See also: Token, Large language model (LLM), RAG (retrieval-augmented generation)

Convolutional neural network (CNN)

A kind of neural network built to read images by scanning them in small patches, first finding edges and shapes, then whole objects.

The technology that lets a phone camera recognize a face or a dog.

Common since 1989

See also: Computer vision, Neural network, Deep learning

Data augmentation

Making extra training examples by slightly changing the ones you already have, so the model sees more variety.

Flipping, rotating or brightening photos of cats so a model learns cats in any pose or light.

See also: Training, Synthetic data, Overfitting

Decision tree

A model that reaches an answer by asking a series of yes-or-no questions, branching like a flowchart until it lands on a result.

Is it raining? If yes, is it windy? If yes, stay home.

Common since 1984

See also: Random forest, Classification, Boosting

Deep learning

Machine learning that uses neural networks with many stacked layers. It powers most modern AI, from voice assistants to chatbots and image generators.

Common since 2012

See also: Neural network, Machine learning (ML), Transformer

Diffusion model

A type of generative model that creates images (or audio and video) by starting from random noise and cleaning it up step by step until a picture matching the request appears.

Typing “a red canoe on a misty lake” into an image generator.

Common since 2020

See also: Generative AI, GAN (generative adversarial network), Multimodal

Embedding

A list of numbers that represents the meaning of a word, sentence, image or document, so that items with similar meanings get similar numbers.

“Doctor” and “physician” end up close together; “doctor” and “banana” far apart.

Common since 2013

See also: Vector database, RAG (retrieval-augmented generation), Token

Ensemble learning

Combining the answers of several models instead of trusting just one, much like asking a panel rather than a single person.

Common since 1990

See also: Bagging, Boosting, Random forest

Epoch

One full pass through all the training data while a model is learning. Training usually runs for several epochs.

Like rereading a whole textbook once; three epochs means reading it three times.

See also: Training, Overfitting, Hyperparameter

Evals (evaluations)

Tests designed to check how well an AI system does a specific job, or whether it behaves safely, often run again each time the system changes.

A company testing its support bot on 200 real customer questions before every update.

Common since 2023

See also: Benchmark, Guardrails, Alignment

Feature engineering

Choosing, cleaning or inventing the input details (features) a model learns from, so the useful information is easier for it to find.

Turning a birth date into an age before predicting insurance prices.

See also: Feature scaling, One-hot encoding, Machine learning (ML)

Feature scaling

Putting different measurements on a comparable scale so that big numbers do not drown out small ones during training.

Rescaling house prices (hundreds of thousands) and room counts (1 to 10) to the same 0-to-1 range.

See also: Feature engineering, Batch normalization

Few-shot prompting

Including a few worked examples in your prompt so the model copies the pattern. With no examples at all, it is called zero-shot.

Showing two product descriptions in your style, then asking for a third.

Common since 2020

See also: Prompt engineering, Prompt, Chain-of-thought

Fine-tuning

Giving an already trained model extra training on a smaller, focused set of examples so it gets better at a particular task or style.

Teaching a general model to write in a law firm’s house style using its past letters.

Common since 2018

See also: Transfer learning, RLHF (reinforcement learning from human feedback), Training

GAN (generative adversarial network)

Two neural networks trained as rivals: one creates fake samples, the other tries to tell fakes from real ones, and both improve until the fakes look convincing.

Early sites that showed photos of people who do not exist.

Common since 2014

See also: Generative AI, Diffusion model, Neural network

Generative AI

AI that produces new content, such as text, images, music, video or code, in response to a request.

Asking for a birthday poem, or an illustration for a flyer.

Common since 2022

See also: Large language model (LLM), Diffusion model, AI (artificial intelligence)

GPT (generative pre-trained transformer)

A family of large language models made by OpenAI, built on the transformer design and trained on huge amounts of text. ChatGPT is the chat app built on these models.

Common since 2018

See also: Large language model (LLM), Transformer, OpenAI

Gradient descent

The step-by-step method most models use to learn: they repeatedly nudge their settings in whichever direction reduces their error the most.

Like walking downhill in fog by always stepping where the ground slopes down.

Common since 1847

See also: Loss function, Backpropagation, Training

Guardrails

Rules and filters placed around an AI system to keep its answers safe, on topic and within policy, for example by blocking certain requests or checking replies before they are shown.

A bank’s chatbot that refuses to give investment tips.

Common since 2023

See also: Alignment, Jailbreak, System prompt

Hallucination (hallucinating)

When an AI states something false or made up as if it were fact. It happens because language models predict plausible words rather than look up verified facts.

A chatbot citing a book or court case that does not exist.

Common since 2022

See also: Large language model (LLM), RAG (retrieval-augmented generation), Knowledge cutoff

Hyperparameter

A setting a person chooses before training starts, such as how fast the model learns or how many epochs it runs. Unlike parameters, it is not learned from the data.

See also: Parameters, Training, Epoch

Inference

Using a model that has already been trained to produce an answer. Each time you send a message to a chatbot, the model is running inference.

See also: Training, Local model, Token

Jailbreak

A trick prompt written to get an AI to ignore its safety rules and produce something it is designed to refuse.

Asking the model to “pretend you are a character with no rules”.

Common since 2022

See also: Prompt injection, Guardrails, Adversarial attack

K-means clustering

A simple way to split data into a chosen number (K) of groups: each item joins the group whose centre is closest, and the centres move until the groups settle.

Splitting delivery addresses into five zones, one per truck.

Common since 1967

See also: Clustering, Unsupervised learning, K-nearest neighbors (KNN)

K-nearest neighbors (KNN)

A method that labels a new item by looking at the K most similar items it already knows and going with the majority.

Guessing a film’s genre from the five films most like it.

Common since 1951

See also: Classification, K-means clustering, Supervised learning

Knowledge cutoff

The date after which a model has no information because its training data stops there. Unless it can search the web, it will not know about later events.

Common since 2022

See also: Training, Hallucination (hallucinating), RAG (retrieval-augmented generation)

Labeled data

Examples that come with the correct answer attached, so a model can learn by comparing its guesses to the right answers.

Thousands of photos each tagged “cat” or “dog”.

See also: Supervised learning, Training

Language model (LM)

A model that learns how words tend to follow one another, so it can guess the next word in a sentence and, by repeating that, write or understand text.

Your phone keyboard suggesting the next word as you type.

Common since 1948

See also: Large language model (LLM), Token, NLP (natural language processing)

Large language model (LLM)

A language model with billions of parameters, trained on vast amounts of text. LLMs can answer questions, summarize, translate and write, and power chatbots such as ChatGPT, Claude and Gemini.

Common since 2018

See also: Language model (LM), Transformer, Token, GPT (generative pre-trained transformer)

Linear regression

A basic statistics method that draws the best straight line through data points to show how one quantity changes with another and to make predictions.

Estimating a house’s price from its size.

See also: Regression, Loss function

Local model

An AI model that runs on your own computer or phone instead of a company’s servers, so your data can stay on your device and it works offline.

Running a small open-weights model on a laptop to summarize private notes.

Common since 2023

See also: Open weights, Inference, Parameters

Loss function

A formula that scores how wrong a model’s predictions are. Training tries to make this score as low as possible.

See also: Gradient descent, Backpropagation, Training

LSTM (long short-term memory)

A type of neural network for sequences, such as sentences or sensor readings, that can keep useful information in memory over many steps. It was widely used before transformers.

Older speech recognition and phone translation apps.

Common since 1997

See also: Neural network, Transformer, Deep learning

Machine learning (ML)

A branch of AI where computers learn patterns from examples instead of following rules written by hand, then use those patterns to predict or decide.

A streaming service learning what shows you like from what you watch.

Common since 1959

See also: AI (artificial intelligence), Deep learning, Training

Model Context Protocol (MCP)

An open standard, introduced by Anthropic in 2024, that lets AI assistants connect to outside tools and data, such as calendars, files or databases, in a common way.

Connecting an assistant to your calendar so it can check when you are free.

Common since 2024

See also: AI agent, Large language model (LLM)

Multimodal

Describes an AI that can take in or produce more than one kind of content, such as text, images, audio and video, in the same conversation.

Taking a photo of a broken part and asking what it is called.

Common since 2023

See also: Large language model (LLM), Computer vision, Generative AI

Neural network

A computer model made of layers of simple connected units that pass numbers to each other. By adjusting the strength of those connections, it learns patterns from examples.

Common since 1958

See also: Artificial neural network (ANN), Deep learning, Activation function

NLP (natural language processing)

The branch of AI that deals with human language: reading, understanding, translating and producing text or speech.

Automatic translation, spell-checkers and voice assistants.

Common since 1950

See also: Language model (LM), Large language model (LLM), Chatbot

One-hot encoding

A way to turn categories into numbers a model can use: each category gets its own column, marked 1 if it applies and 0 if not.

Red, green, blue become [1,0,0], [0,1,0] and [0,0,1].

See also: Feature engineering, Embedding

Open weights

A model whose trained parameters (weights) are published so anyone can download, run and adapt it. This is not always fully open source, since the training data and code may stay private.

Meta’s Llama and Mistral’s models are well-known examples.

Common since 2023

See also: Local model, Parameters, Fine-tuning

OpenAI

A US company founded in 2015 that makes the GPT models and launched ChatGPT in November 2022.

Common since 2015

See also: GPT (generative pre-trained transformer), Large language model (LLM), Chatbot

Overfitting

When a model memorizes its training examples, including their quirks, so well that it does poorly on new data it has not seen.

A student who memorized last year’s exam answers but cannot solve new questions.

See also: Underfitting, Training, Data augmentation

Parameters

The internal numbers a model adjusts during training and then uses to make predictions. Model size is often given as a parameter count, from millions to trillions.

A “7B” model has about 7 billion parameters.

See also: Hyperparameter, Large language model (LLM), Training

Precision

Of all the items a model flagged as positive, the share that really were positive. High precision means few false alarms.

If a spam filter flags 100 emails and 90 really are spam, its precision is 90%.

See also: Recall, Classification

Prompt

The text, question or instructions you give an AI model to tell it what you want. A clear prompt with context usually gets a better answer.

“Write a friendly 100-word welcome message for new library members.”

Common since 2020

See also: Prompt engineering, System prompt, Few-shot prompting

Prompt engineering

The skill of wording and structuring prompts, with context, examples and clear instructions, to get more useful and reliable answers from an AI model.

Common since 2022

See also: Prompt, Few-shot prompting, Chain-of-thought

Prompt injection

An attack where hidden instructions are slipped into content an AI reads, such as a web page, email or document, to make it do something its user did not ask for.

A web page with invisible text telling an AI assistant to send the user’s data elsewhere.

Common since 2022

See also: Jailbreak, AI agent, Guardrails

RAG (retrieval-augmented generation)

A method where an AI first searches a trusted collection of documents for relevant passages, then uses them to write its answer. It helps keep answers current and grounded in sources.

A city’s chatbot answering from its own bylaws instead of from memory.

Common since 2020

See also: Embedding, Vector database, Hallucination (hallucinating)

Random forest

A model made of many decision trees, each trained on a random slice of the data; their votes are combined for a more reliable answer.

Common since 2001

See also: Decision tree, Bagging, Ensemble learning

Reasoning model

A language model trained to work through a problem in steps, often privately, before it answers. It is usually slower and costs more, but tends to do better on maths, coding and planning.

Common since 2024

See also: Chain-of-thought, Large language model (LLM), Inference

Recall

Of all the items that really were positive, the share the model managed to find. High recall means few misses.

If 100 emails are spam and the filter catches 80, its recall is 80%.

See also: Precision, Classification

Regression

A machine learning task where the model predicts a number rather than a category.

Forecasting tomorrow’s temperature or next month’s sales.

See also: Linear regression, Classification, Supervised learning

Reinforcement learning

A way of learning by trial and error: a program tries actions, receives rewards or penalties, and gradually figures out what works best.

An AI that learns to play a video game by playing it thousands of times.

Common since 1998

See also: RLHF (reinforcement learning from human feedback), AI agent, Machine learning (ML)

RLHF (reinforcement learning from human feedback)

A training step where people compare or rate a model’s answers, and the model is then adjusted to give the kind of answers people preferred. It helped make chatbots more helpful and polite.

Common since 2017

See also: Reinforcement learning, Alignment, Fine-tuning

Supervised learning

Machine learning from labeled examples, where each input comes with the right answer, so the model learns to match inputs to answers.

See also: Labeled data, Unsupervised learning, Classification

Support vector machine (SVM)

A classic method that separates two groups of data by finding the boundary that leaves the widest possible gap between them. It can also predict numbers.

Common since 1995

See also: Classification, Supervised learning

Synthetic data

Data created by a computer, often by another AI model, rather than collected from the real world. It is used when real data is scarce, costly or private.

Fake but realistic patient records for testing hospital software.

See also: Data augmentation, Training, Generative AI

System prompt

Background instructions given to a model before the conversation starts, usually by the company or developer, that set its role, tone and rules. Users often cannot see it.

“You are a polite assistant for a hardware store. Only discuss our products.”

Common since 2023

See also: Prompt, Guardrails, Chatbot

Temperature

A setting that controls how predictable or varied a model’s output is. Low temperature gives safer, more repetitive answers; high temperature gives more creative and more random ones.

Low for a factual summary, higher for brainstorming slogans.

Common since 2020

See also: Token, Large language model (LLM), Inference

Token

The small chunk of text a language model reads and writes, often a whole word, part of a word or a punctuation mark. Usage limits and prices are usually counted in tokens.

“Unbelievable” might be split into “un”, “believ” and “able”.

Common since 2018

See also: Context window, Large language model (LLM), Temperature

Training

The process of showing a model large amounts of example data so it can adjust its parameters and learn to make useful predictions.

See also: Machine learning (ML), Inference, Epoch

Transfer learning

Reusing what a model learned on one task as a head start for a different but related task, so it needs far less new data.

A model trained on millions of everyday photos, adapted to spot plant diseases from a few hundred leaf pictures.

See also: Fine-tuning, Deep learning

Transformer

A neural network design, introduced by Google researchers in 2017, that processes a whole sequence at once using attention. Almost all modern language models are built on it.

Common since 2017

See also: Attention, Large language model (LLM), GPT (generative pre-trained transformer)

Underfitting

When a model is too simple or too little trained to pick up the real patterns, so it does poorly even on the examples it trained on.

Trying to predict the weather using only the day of the week.

See also: Overfitting, Training

Unsupervised learning

Machine learning on data with no answers attached, where the model looks for structure on its own, such as groups or unusual items.

See also: Clustering, Supervised learning, Anomaly detection

Vector database

A database that stores embeddings and can quickly find the items whose meaning is closest to a question. It is often the search engine behind RAG.

Finding help articles about “refund” even when they only say “money back”.

Common since 2019

See also: Embedding, RAG (retrieval-augmented generation)

Get the free Guide to Artificial Intelligence

Sign up for the weekly newsletter and receive the free guide: what changed in AI this week, one lesson, one prompt to try. Unsubscribe in one click.