AI Jargon, Translated: A Plain-English Glossary of 24 Terms

A system that uses a language model to plan and take actions — searching, running tools, filling in forms — across several steps toward a goal, rather than simply replying to one message.

The work of making AI systems behave in line with what people intend: helpful, honest and avoiding harm, even in situations their designers didn't anticipate.

Software that performs tasks we associate with human thinking, such as understanding language, recognising images or making predictions. It's an umbrella term, not a single technology.

How much text a model can take into account at once — your prompt, any documents you've included and its own reply. Anything beyond it is effectively out of sight.

A list of numbers that represents the meaning of a piece of text, arranged so that similar meanings sit close together. Embeddings are what make search-by-meaning possible.

Training an existing model further on a specific set of examples to shape its behaviour, style or a narrow skill. Better for teaching behaviour than for keeping facts up to date.

AI that creates new content — text, images, audio or code — rather than only sorting or scoring content that already exists.

Rules and filters placed around a model to block unsafe, off-topic or policy-breaking requests and responses.

When a model states something false as though it were fact. The word is a metaphor: the model isn't seeing things, it's producing plausible text without a reliable check on whether it's true.

Using a trained model to produce an output. Training is the learning; inference is the answering.

The point at which a model's training data ends. It won't know about anything later unless that information is supplied — through search, for example, or RAG.

A neural network trained on a vast amount of text to predict the next token. At sufficient scale, that single skill covers writing, summarising, translating and answering questions.

A way of building AI in which a system learns patterns from examples instead of following rules written by hand.

Describes a model that works with more than one kind of input or output — for example text and images, or text and audio.

A model built from layers of simple connected units whose connection strengths are adjusted during training. Loosely inspired by the brain, and far simpler than one.

A model whose trained parameters are published so anyone can download and run it. Often loosely called open source, although the training data and code may not be released.

The adjustable numbers inside a model that are set during training. More parameters can mean more capability and more cost, but the count alone doesn't tell you how good a model is.

The instructions and material you give a model. The wording matters more than most people expect — see how to write better AI prompts.

Finding relevant documents and handing them to a model before it answers, so the answer is grounded in real sources. Our full RAG explainer.

A setting that controls how much randomness goes into a model's word choices. Lower means more predictable; higher means more varied — and more likely to wander.

The chunk of text a language model reads and writes — often a whole word, sometimes part of one. Usage limits and pricing are usually counted in tokens, and tokens are also why AI finds puns hard.

The material a model learns from. Its content, quality and end date shape what the model knows and where its blind spots are.

The neural-network design behind most modern language models, introduced in 2017. Its central idea, attention, lets the model weigh how each part of a passage relates to every other part.

A database built to store embeddings and quickly find the ones closest in meaning to a query. A common component of RAG systems.