A plain English glossary of AI terms

Thirty words you will meet in the first week, each explained the way you would explain it to a friend.

Every field builds a vocabulary, and most of this one is younger than a decade, which means a lot of it is still being used loosely. What follows is how these words are actually used in practice, in the order you are likely to meet them, rather than alphabetically.

The words you meet in the first hour

Model: The trained system itself, as opposed to the website or app you reach it through. When people compare model names they are comparing engines, not dashboards.
Large language model, or LLM: The kind of model behind a text box. It reads the text so far and produces a ranked guess at what comes next, then repeats. "Large" refers to the scale of the thing, which is enormous.
Prompt: Whatever you type. The mystique around the word is mostly marketing. In practice a good prompt says who you are, what your situation is, what shape of answer you want, and ideally shows an example.
Token: The unit the system actually reads and writes. Roughly three quarters of a word in English, so a page of prose is about five hundred tokens. It matters because limits and charges are counted in tokens rather than words.
Context window: How much text the model can hold in view at once, counted in tokens. Your question, any pasted document and the whole conversation so far all have to fit. When a long conversation exceeds it, the earliest part silently drops out.
Hallucination: An invented fact, source, quotation or number, delivered in the same confident tone as everything true. A misleading name, since it suggests a malfunction. It is a normal consequence of a system trained to produce plausible text rather than verified text.

Words about how the thing was built

Training: The one off, months long process where the model's behaviour was formed by exposure to a very large quantity of text. It is finished before you ever type anything. Your conversation does not train it.
Training cutoff: The date the training text stops. Events after it are unknown unless the product can look things up for you.
Parameters: The internal numbers adjusted during training, counted in billions. Quoted as a rough proxy for size. Bigger is not reliably better, and the figure is increasingly not published at all.
Fine tuning: Further training of an existing model on a narrower set of material, to specialise it. Much cheaper than training from scratch and usually what is meant by a "custom" model.
Alignment: The work of making a model behave helpfully and decline harmful requests. It is where the assistant persona comes from. A raw model out of training does not behave like a polite helper.
Open weights: A model whose trained numbers have been published, so anyone can download and run it. Often loosely called open source, which is not quite the same claim, since the training data usually is not released.

Words about what these systems are wired into

Multimodal: Handles more than text. Usually means it can look at an image you upload, and sometimes audio or video.
Retrieval, or RAG: Fetching relevant documents and putting them into the context window before the model answers, so it is working from actual sources rather than memory. This is how most systems that answer questions about your own files work.
Grounding: Tying an answer to a specific source that can be shown to you. A grounded answer comes with a link you can open. It is the most reliable defence against the fabrication problem.
Agent: A model given the ability to take actions in a loop rather than only produce text: searching, running code, filling in a form. Powerful and currently the least reliable part of the field.
API: The developer facing way in, used to build a model into other software. Relevant to you only if something you use is built on one.
Inference: The technical term for the model actually running and producing an answer. If you see it in an article it just means use, as distinct from training.

Words about steering the output

System prompt: Standing instructions the product puts in front of your message, setting the persona and the rules. You usually cannot see it, and it is part of why the same underlying model behaves differently in different apps.
Temperature: A setting for how much the system varies its choices instead of always taking its top ranked guess. Low is repetitive and predictable, high is inventive and less coherent. Rarely exposed in consumer apps.
Few shot: Including two or three worked examples in your question so the system copies the pattern. The single most effective technique available to an ordinary user.
Chain of thought: Asking the system to work through its reasoning step by step before answering, which measurably improves results on problems with several stages. Newer reasoning models do this internally by default.
Reasoning model: A model that spends longer working before it answers. Better on multi step problems, slower and more expensive, and unnecessary for drafting an email.
Guardrails: The refusals and filters layered around a model. Why an answer sometimes stops short, and why the same question can be answered in one product and declined in another.

Words that get used loosely

Artificial general intelligence, or AGI: A hypothetical system matching human ability across the full range of tasks. No agreed definition and no agreed test, so claims about how close it is are mostly claims about the speaker's definition.
Machine learning: The older, broader field of programs that improve from data. Language models are one branch. Your bank's fraud detection is another and is not remotely the same kind of thing.
Neural network: The general architecture underneath. The brain analogy in the name is historical and about fifty years old. It does not mean the system works like a brain.
Deepfake: Synthetic audio or video of a real person. A separate technology from language models, frequently discussed in the same articles.
Prompt injection: Hidden instructions planted in a document or web page that the model reads and follows as though you had typed them. The main security problem with systems that browse or read files on your behalf.

If a term you meet is not here, the reliable move is to ask a model to explain it at the level you need, then ask it for the objection: "what does that explanation leave out?" The second answer is usually where the real content is.