AI terms explained,
in plain English.
AI terms explained by the Institute of AI: a plain-English glossary that turns the jargon behind the headlines into language anyone can follow. From tokens to transformers, each entry says what a word actually means, without hype and without assuming you already know. Free to read, and free to keep coming back to.
Start with the terms
behind the headlines.
Fourteen of the words you meet most often, each defined in a sentence or two you can actually use. Jump to a letter, or read straight down.
- Agent
- An AI system that does not just answer, but works towards a goal. It can plan, call tools, search, or act across several steps, checking its own progress as it goes. The word describes software that decides what to do next, not only what to say.
- Context window
- The amount of text a model can hold in view at once, counted in tokens. Everything you send, plus everything the model has already said, has to fit inside it. Once a conversation runs past the window, the earliest parts fall out of view and can no longer be referred to.
- Embedding
- A way of turning words, sentences, or whole documents into lists of numbers that capture their meaning, so related ideas sit close together. Search and recommendation systems compare these numbers to find text that means something similar, even when the wording is different.
- Fine-tuning
- Taking a model that has already been trained and giving it extra training on a narrower set of examples, so it gets better at a particular task, subject, or tone. It adjusts an existing model rather than building a new one from scratch.
- Guardrails
- The rules, filters, and checks placed around a model to keep its output safe, on-topic, and within policy. Guardrails can block certain requests, screen responses before they reach a reader, or hold a system to an agreed boundary.
- Hallucination
- When a model states something false as though it were fact, complete with plausible detail. Because these tools predict likely-sounding text rather than look things up, they can invent names, quotes, and figures. Anything that matters needs checking against a real source.
- Inference
- The act of running a trained model to get an answer. Training is how a model learns; inference is every time you actually use it. Each question you ask is one inference.
- Large language modelLLM
- A model trained on vast amounts of text to predict the next token, given everything that has come before. From that single skill it can draft, summarise, translate, and answer questions. The tools behind most chat assistants are large language models.
- Parameters
- The internal values a model adjusts during training, the settings that hold what it has learned. Modern models have billions of them. The count gives a rough sense of how big a model is, not a direct measure of how good it is.
- Prompt
- The instruction or question you give a model, along with any examples or context you include. Clear, specific prompts tend to get clearer, more useful answers, which is why writing them well has become a skill in its own right.
- Retrieval-augmented generationRAG
- A method that lets a model answer from a chosen set of documents rather than memory alone. The system first retrieves the most relevant passages, then hands them to the model to work from. It grounds answers in sources you control and can point back to.
- Token
- The unit a model reads and writes in: a chunk of text that is often a short word or part of a longer one. Length, limits, and cost are all measured in tokens rather than words. As a rough guide, a hundred tokens is about seventy-five words of English.
- Training
- The process of teaching a model by showing it huge quantities of examples and adjusting its parameters until its predictions improve. It happens once, up front, and is far more costly than running the finished model to answer a question.
- Transformer
- The design most modern AI models are built on. Its central idea, called attention, lets the model weigh how strongly each word relates to every other word in a passage. That is what gives these systems their grasp of context, and the T in GPT.
Clear words
beat clever ones.
Most confusion about AI is not confusion about the technology. It is confusion about the words, which are often borrowed, overloaded, or wrapped in marketing. Define them plainly and the subject stops feeling like a members-only conversation.
That is the standard the Institute of AI holds itself to across everything it publishes. No definition here trades accuracy for a neat line, and none assumes you arrived already fluent. When you are ready to go wider, the full glossary carries hundreds more entries in the same voice.
Questions about AI terms.
What is the difference between a token and a word?+
What does LLM stand for?+
Do I need to understand these terms to use AI?+
Why do AI tools sometimes get facts wrong?+
Is this glossary free to use?+
Understand the words,
follow the whole story.
One clear definition often opens up a whole subject. Open the full glossary from the Institute of AI, or explore the rest of Learn AI when you are ready to go deeper.
The full AI glossary
Hundreds of AI terms defined in plain English, from the everyday to the technical, kept current as the field changes.
Browse the glossaryQuick answers
Short, direct responses to the questions people ask most about AI, written to be understood at a glance.
Get quick answersPrompt engineering
Now you know the words, put them to work. Learn how to write prompts that get clearer, more useful results.
Learn prompting