AI glossary

Fourteen key words about AI, explained plainly. Keep it open whenever you’re not sure what something means.

Artificial Intelligence (AI)

Artificial Intelligence began as an area of study in the 1950s, and focuses on developing systems which can complete tasks people set them up to do. These tasks include things like sorting information, finding patterns, predicting what comes next, making text or pictures, and controlling machines.

The concept of “AI” does not have one fixed meaning. Some systems follow exact rules which are defined by people, step by step. Others are not given exact rules, but instead find patterns from many examples (see machine learning). Many systems mix both.

Because it is ambiguous, “AI” is used as a label for many different kinds of systems, and the term is often used to make something sound more advanced or exciting. Tools once called AI, like chess-playing programs, often lose the label once they become commonplace. Since the term is used loosely, it is worth asking what a system actually does, rather than assuming your understanding of AI matches that of others.

Back to all terms

Machine Learning (ML)

Machine learning is one method of building an AI system, and is the most common in AI today. Instead of a person writing out exact rules, systems “learn” the patterns present in a large number of examples (see data). The system then uses those patterns to make predictions about new information.

For example, to build a system that sorts photos of cats and dogs, a person does not write a rule for what a cat looks like. They give the system many photos already labelled “cat” or “dog,” and it works out patterns that tell the two apart.

Back to all terms

Data

Any information, such as facts, numbers, words, or observations which can be used for reference. Data can come in many forms, such as numbers, text, images, or sensor readings, and becomes valuable when it is processed and analyzed to support understanding or decision-making.

Back to all terms

Model (“AI Model”)

A model is the core part of an AI system that turns input into output. It holds the patterns or rules the system uses to do its task. In most systems discussed during the assembly these patterns are learned from data, but in others the rules are written directly by people.

Back to all terms

Training Data

The examples used to build a machine learning model.

During training, the model’s settings are adjusted so that inputs result in outputs which reflect the patterns in these examples. Models are usually built from very large amounts of training data.

Back to all terms

Generative AI

A type of AI that produces content, such as text, images, audio, or video. It works by finding patterns in its training data and using them to build new outputs of the same kind. The outputs are combinations shaped by the patterns.

Back to all terms

Chatbot

A system built to interact with people in everyday language through back-and-forth messages. A person sends a message and the chatbot outputs text in response. Most chatbots today use “large language models” to generate the response, but some chatbots follow fixed scripts or rules written by people.

Back to all terms

Large Language Model (LLM)

A generative AI system built by finding patterns in very large amounts of text, such as books, articles, and websites. These systems work by predicting the next unit of text (roughly, the next word) based on what came before, and repeating this to build sentences and longer phrases.

On its own, a language model will only continue text in this way. To make it useful for tasks like answering questions, it is adjusted after this initial training (creating what sits behind many chatbots).

Back to all terms

AI Agent (“agentic AI”)

An AI system set up to pursue a goal given by a person by carrying out a series of steps, rather than producing a single response. It can be connected to outside tools, such as software that sends messages, searches the web, or changes files, and it uses generated text to work out each next step. Because that text can be wrong in the same ways other AI output can, and the tools take real actions, how much a person reviews the steps before they happen varies, and some systems are designed to run with little or no human checking.

Back to all terms

Deepfake

A video, image, or audio clip created or altered by AI so that it looks or sounds real, even though it shows something that did not actually happen. Deepfakes are often used to make it appear as if a real person said or did something they did not.

Back to all terms

Algorithmic Bias

When an AI system regularly produces worse outcomes for certain groups of people, because of patterns in the data it was built from, choices made in how it was built, or both. This can happen even when no one intended it.

Back to all terms

Prompt

The text (or other input) a person provides to an AI model in order to generate a response. Because the model builds its output from the input it is given, the wording of the prompt affects what comes back, which is why the same question phrased differently can produce different results.

Back to all terms

Hallucination

A term commonly used to refer to when a model produces confident-sounding output that is either false or inconsistent with information provided, such as invented facts, statistics, or references to sources that do not exist.

This term should be used with caution for three reasons. First, it borrows a word for a human experience and applies it to a machine, which does not experience anything. Second, the word is ambiguous even when used about people; whether a given experience counts as a hallucination often depends on who is interpreting it. Third, it bundles together many different kinds of failure, such as factual inaccuracies, made-up sources, and confident but faulty responses, as if they were one thing with one cause, making the problem more difficult to understand.

Back to all terms

Artificial General Intelligence (AGI)

A term for a hypothetical future AI that could handle almost any intellectual task about as well as, or better than, a person, rather than being limited to particular tasks as current systems are. No such system exists.

There is no agreed definition. As one example, OpenAI, the company responsible for ChatGPT, defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Definitions like this vary widely and some experts question whether AGI is even a clear or measurable idea, since it assumes human intelligence is a single thing a machine could match on one scale. The term is also used heavily in marketing and fundraising, often to suggest such systems are close, a claim the evidence does not settle. Predictions about if or when AGI might arrive have a long history of being wrong, so confident timelines, from any source, are worth treating with care.

Back to all terms

How the glossary was made